# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright The LanceDB Authors from __future__ import annotations import asyncio import inspect import json import deprecation import warnings from abc import ABC, abstractmethod from dataclasses import dataclass from datetime import datetime, timedelta from functools import cached_property from typing import ( TYPE_CHECKING, Any, AsyncIterator, Callable, Dict, Iterable, Iterator, List, Literal, Optional, Sequence, Tuple, TypedDict, Union, overload, ) from urllib.parse import urlparse from lancedb.scannable import _hf_batches, _register_optional_converters, to_scannable from . import __version__ from ._blob import ( BlobFile, _normalize_blob_descriptor, _normalize_blob_row_ids, _wrap_blob_files, strip_auto_row_ids, validate_blob_mode, ) from .types import BlobMode from lancedb.arrow import peek_reader from lancedb.background_loop import LOOP, embedding_executor from lancedb.job import AsyncJob, Job, _typed_job from .dependencies import ( _check_for_hugging_face, _check_for_lance, _check_for_pandas, lance, pandas as pd, polars as pl, ) import pyarrow as pa import pyarrow.dataset import pyarrow.compute as pc import pyarrow.fs as pa_fs import numpy as np from .common import DATA, VEC, VECTOR_COLUMN_NAME from .embeddings import EmbeddingFunctionConfig, EmbeddingFunctionRegistry from .index import ( BTree, IvfFlat, IvfPq, IvfSq, Bitmap, IvfRq, LabelList, Fm, HnswPq, HnswSq, HnswFlat, FTS, ) from .expr import Expr from .functions import ( FunctionApplication, RefreshColumnResult as RefreshColumnJobResult, ) from .merge import LanceMergeInsertBuilder from .pydantic import LanceModel, model_to_dict from .query import ( AnalyzePlanDistributedMetrics, AsyncFTSQuery, AsyncHybridQuery, AsyncQuery, AsyncTakeQuery, AsyncVectorQuery, DocumentGranularity, FullTextQuery, LanceEmptyQueryBuilder, LanceFtsQueryBuilder, LanceHybridQueryBuilder, LanceQueryBuilder, LanceVectorQueryBuilder, LanceTakeQueryBuilder, Query, ) from .util import ( add_note, fs_from_uri, get_uri_scheme, infer_vector_column_name, join_uri, value_to_sql, ) from .index import lang_mapping from .schema import ( blob_v2_column_paths, is_blob_v2_field, row_addressable_blob_v2_paths, schema_has_blob_field, ) def _should_push_down_query_table( namespace_client: Optional[Any], pushdown_operations: set ) -> bool: return namespace_client is not None and "QueryTable" in pushdown_operations _CUDA_UNAVAILABLE_MESSAGE = ( "CUDA acceleration is unavailable in this Python environment. Install a " "CUDA-enabled PyTorch build and verify that the CUDA device is accessible, or " "omit `accelerator` (or set `accelerator=None`) to create the index on CPU." ) def _validate_index_accelerator(accelerator: Any) -> None: if not isinstance(accelerator, str) or accelerator.split(":", 1)[0] != "cuda": return try: import torch cuda_available = torch.cuda.is_available() except (ImportError, RuntimeError, AssertionError) as err: raise RuntimeError(_CUDA_UNAVAILABLE_MESSAGE) from err if not cuda_available: raise RuntimeError(_CUDA_UNAVAILABLE_MESSAGE) try: torch.cuda.init() except (RuntimeError, AssertionError) as err: raise RuntimeError(_CUDA_UNAVAILABLE_MESSAGE) from err def _polars_predicate_pushdown_barrier(frame: Any) -> Any: """Return a Polars frame unchanged while blocking predicate pushdown.""" return frame _MODEL_BACKED_TOKENIZER_PREFIXES = ("jieba", "lindera") _MODEL_BACKED_TOKENIZER_ERRORS = ( "unknown base tokenizer", "Invalid directory path:", "Failed to load Jieba", "Failed to load tokenizer config", "Failed to initialize default tokenizer", ) def _optimize_cleanup_since_ms( cleanup_older_than: Optional[timedelta], retrain: bool ) -> Optional[int]: # Called directly by both the sync and async optimize so stacklevel=3 # names the user's call site rather than the background event loop. cleanup_since_ms: Optional[int] = None if cleanup_older_than is not None: cleanup_since_ms = round(cleanup_older_than.total_seconds() * 1000) if cleanup_since_ms <= 0: warnings.warn( "optimize(cleanup_older_than=0) removes every version except " "the latest. Any concurrent reader or writer still using an " "older version will fail. Use a longer cleanup_older_than " "unless no other process is working on this table.", UserWarning, stacklevel=3, ) if retrain: warnings.warn( "The 'retrain' parameter is deprecated and will be removed in a " "future version.", DeprecationWarning, stacklevel=3, ) return cleanup_since_ms def _add_unique_note(exception: BaseException, note: str) -> None: existing_notes = getattr(exception, "__notes__", ()) or () message = ( exception.args[0] if exception.args and isinstance(exception.args[0], str) else "" ) if note not in existing_notes and note not in message: add_note(exception, note) def _is_model_backed_tokenizer(base_tokenizer: str) -> bool: return any( base_tokenizer == prefix or base_tokenizer.startswith(f"{prefix}/") for prefix in _MODEL_BACKED_TOKENIZER_PREFIXES ) def _maybe_add_fts_error_note( exception: BaseException, *, base_tokenizer: str, language: Optional[str] = None ) -> None: message = str(exception) if language is not None and "not support the requested language" in message: supported_langs = ", ".join(lang_mapping.values()) _add_unique_note(exception, f"Supported languages: {supported_langs}") return if not _is_model_backed_tokenizer(base_tokenizer): return if not any(marker in message for marker in _MODEL_BACKED_TOKENIZER_ERRORS): return _add_unique_note( exception, "Model-backed tokenizers such as 'jieba/default' and 'lindera/ipadic' " "require tokenizer models in Lance's language model home. Set " "LANCE_LANGUAGE_MODEL_HOME to override the default platform data " "directory under 'lance/language_models'. Expected layouts include " "'/jieba/default/...' and " "'/lindera/ipadic/...'.", ) if TYPE_CHECKING: from .db import LanceDBConnection from ._lancedb import ( Table as LanceDBTable, OptimizeStats, CleanupStats, CompactionStats, Tag, AddColumnsResult, FunctionErrors, RefreshColumnResult, AddResult, AlterColumnsResult, UpdateFieldMetadataResult, DeleteResult, DropColumnsResult, FtsToken, RTEWriteSpec, MergeResult, UpdateResult, ) from .index import IndexConfig import pandas import PIL from .types import ( QueryType, OnBadVectorsType, AddMode, CreateMode, ScalarIndexType, BaseTokenizerType, ) # Type alias for index configuration objects IndexConfigType = Union[ IvfFlat, IvfPq, IvfSq, IvfRq, HnswFlat, HnswPq, HnswSq, BTree, Bitmap, LabelList, Fm, FTS, ] class CompactionOptions(TypedDict, total=False): """Options that control file compaction during table optimization. Unspecified options use Lance's defaults. Compaction planning is row based. Lowering ``target_rows_per_fragment`` based on the expected row size can bound later compaction passes once oversized fragments have been rewritten. It does not split an existing fragment, so the first pass over an oversized fragment is not subject to that bound. ``max_bytes_per_file`` limits output file size, not compaction memory. Source budgets keep whole planned tasks; if the first task exceeds a budget, that run performs no compaction work. Examples -------- Derive a steady-state row target from the expected row size: >>> desired_fragment_bytes = 750 * 1024 * 1024 >>> average_row_bytes = 1_500_000 >>> options: CompactionOptions = { ... "target_rows_per_fragment": max( ... 1, desired_fragment_bytes // average_row_bytes ... ), ... } >>> await table.optimize(compaction_options=options) # doctest: +SKIP """ target_rows_per_fragment: int """Target rows per fragment; existing oversized fragments are not split.""" max_rows_per_group: int """Maximum number of rows per row group (default: 1,024).""" max_bytes_per_file: Optional[int] """Maximum output data-file size; this does not bound compaction memory.""" materialize_deletions: bool """Whether to rewrite fragments containing deleted rows (default: True).""" materialize_deletions_threshold: float """Minimum deleted-row fraction that makes a fragment eligible (default: 0.1).""" num_threads: Optional[int] """Number of compaction tasks to run in parallel.""" batch_size: Optional[int] """Number of rows per input scan batch.""" io_buffer_size: Optional[int] """Maximum number of bytes queued in the input scan I/O buffer.""" defer_index_remap: bool """Whether to defer index remapping during compaction (default: False).""" index_remap_mode: Literal["direct", "compact"] """How to construct the old-to-new row-address mapping.""" compaction_mode: Optional[ Literal["reencode", "try_binary_copy", "force_binary_copy"] ] """Whether compaction re-encodes data or uses binary copying.""" binary_copy_read_batch_bytes: Optional[int] """Number of bytes read per batch during binary-copy compaction.""" max_source_fragments: Optional[int] """Maximum number of source fragments compacted in one run.""" max_source_rows: Optional[int] """Maximum live source rows per run, applied to whole planned tasks.""" max_source_bytes: Optional[int] """Maximum source bytes per run, applied to whole planned tasks.""" excluded_fragment_ids: List[int] """Fragment IDs to leave unchanged and use as planning boundaries.""" max_overlays_per_fragment: Optional[int] """Maximum overlays before a fragment is fully compacted.""" data_storage_version: str """Data-file version or release selector for compacted output.""" # Known distance metrics for legacy API detection KNOWN_METRICS = {"l2", "cosine", "dot", "hamming"} def _blob_value_to_storage(value: Any) -> Optional[dict]: """Keep a Python blob's inline data or external URI before Arrow infers it.""" if value is None: return None if isinstance(value, (bytes, bytearray, memoryview)): return {"data": bytes(value)} if isinstance(value, str): if not value: raise ValueError("Blob uri cannot be empty") return {"uri": value} if isinstance(value, dict): unknown = value.keys() - {"data", "uri", "position", "size"} if unknown: raise ValueError(f"Unknown blob fields: {sorted(unknown)}") return value try: from lance.blob import Blob except ModuleNotFoundError as err: if err.name not in ("lance", "lance.blob"): raise else: if isinstance(value, Blob): return { "data": value.data, "uri": value.uri, "position": value.position, "size": value.size, } raise TypeError(f"Unsupported blob value: {type(value).__name__}") def _blob_input_to_arrow(data: Any, schema: Optional[pa.Schema]) -> Optional[pa.Table]: """Convert Python rows with blob fields before Arrow loses their value types.""" if schema is None: return None if isinstance(data, list) and data and isinstance(data[0], dict): names = { field.name for field in schema if is_blob_v2_field(field) and any(isinstance(row, dict) and field.name in row for row in data) } if not names: return None values = {name: [row.get(name) for row in data] for name in names} rows = [{**row, **{name: None for name in names}} for row in data] table = pa.Table.from_pylist(rows) elif _check_for_pandas(data) and isinstance(data, pd.DataFrame): names = { field.name for field in schema if is_blob_v2_field(field) and field.name in data.columns } if not names: return None values = { name: [ None if value is pd.NA or isinstance(value, float) and np.isnan(value) else value for value in data[name] ] for name in names } table = pa.Table.from_pandas( data.assign(**{name: None for name in names}), preserve_index=False ).replace_schema_metadata(None) else: return None for field in schema: if field.name not in names: continue storage_type = ( field.type.storage_type if isinstance(field.type, pa.ExtensionType) else field.type ) storage_values = [_blob_value_to_storage(value) for value in values[field.name]] missing_range_fields = {"position", "size"} - set(storage_type.names) # Arrow silently ignores dictionary keys absent from the storage type. if missing_range_fields and any( value is not None and any(value.get(name) is not None for name in missing_range_fields) for value in storage_values ): raise ValueError( f"Blob column '{field.name}' storage schema cannot store byte ranges. " "Create the column with lancedb.blob() to use position and size." ) storage = pa.array(storage_values, type=storage_type) column = ( pa.ExtensionArray.from_storage(field.type, storage) if isinstance(field.type, pa.ExtensionType) else storage ) table = table.set_column( table.schema.get_field_index(field.name), field, column ) return table def _into_pyarrow_reader( data, schema: Optional[pa.Schema] = None ) -> pa.RecordBatchReader: from lancedb.dependencies import datasets if _check_for_hugging_face(data): if isinstance(data, datasets.Dataset): schema = data.features.arrow_schema return pa.RecordBatchReader.from_batches(schema, _hf_batches(data)) elif isinstance(data, datasets.dataset_dict.DatasetDict): schema = _schema_from_hf(data, None) if "split" not in schema.names: schema = schema.append(pa.field("split", pa.string())) return pa.RecordBatchReader.from_batches( schema, _to_batches_with_split(data) ) if isinstance(data, LanceModel): raise ValueError("Cannot add a single LanceModel to a table. Use a list.") if isinstance(data, dict): raise ValueError( "Cannot create or add rows from a single dictionary. " "Use a list of dictionaries instead." ) if isinstance(data, list): # Handle empty list case if not data: if schema is None: raise ValueError("Cannot create table from empty list without a schema") return pa.Table.from_pylist(data, schema=schema).to_reader() # convert to list of dict if data is a bunch of LanceModels if isinstance(data[0], LanceModel): model_schema = data[0].__class__.to_arrow_schema() data = [model_to_dict(d) for d in data] data = _serialize_json_values(data, schema or model_schema) return pa.Table.from_pylist(data, schema=model_schema).to_reader() elif isinstance(data[0], pa.RecordBatch): return pa.Table.from_batches(data).to_reader() else: data = _serialize_json_values(data, schema) table = _blob_input_to_arrow(data, schema) return ( table if table is not None else pa.Table.from_pylist(data) ).to_reader() elif _check_for_pandas(data) and isinstance(data, pd.DataFrame): table = _blob_input_to_arrow(data, schema) if table is None: table = pa.Table.from_pandas(data, preserve_index=False) # Do not serialize Pandas metadata meta = table.schema.metadata if table.schema.metadata is not None else {} meta = {k: v for k, v in meta.items() if k != b"pandas"} return table.replace_schema_metadata(meta).to_reader() elif isinstance(data, pa.Table): return data.to_reader() elif isinstance(data, pa.RecordBatch): return pa.RecordBatchReader.from_batches(data.schema, [data]) elif _check_for_lance(data) and isinstance(data, lance.LanceDataset): return data.scanner().to_reader() elif isinstance(data, pa.dataset.Dataset): return data.scanner().to_reader() elif isinstance(data, pa.dataset.Scanner): return data.to_reader() elif isinstance(data, pa.RecordBatchReader): return data elif ( type(data).__module__.startswith("polars") and data.__class__.__name__ == "DataFrame" ): return data.to_arrow().to_reader() elif ( type(data).__module__.startswith("polars") and data.__class__.__name__ == "LazyFrame" ): return data.collect().to_arrow().to_reader() elif isinstance(data, Iterable): return _iterator_to_reader(data, schema) else: raise TypeError( f"Unknown data type {type(data)}. " "Supported types: list of dicts, pandas DataFrame, polars DataFrame, " "pyarrow Table/RecordBatch, or Pydantic models. " "See https://docs.lancedb.com/tables/ for examples." ) def _iterator_to_reader( data: Iterable, schema: Optional[pa.Schema] = None ) -> pa.RecordBatchReader: # Each batch is treated as it's own reader, mainly so we can # re-use the _into_pyarrow_reader logic. data = iter(data) try: first_batch = next(data) except StopIteration: if schema is None: raise ValueError("Cannot create table from empty iterable without a schema") return pa.RecordBatchReader.from_batches(schema, []) first = _into_pyarrow_reader(first_batch) schema = first.schema def gen(): yield from first for batch in data: table: pa.Table = _into_pyarrow_reader(batch).read_all() if table.schema != schema: try: table = table.cast(schema) except pa.lib.ArrowInvalid: raise ValueError( f"Input iterator yielded a batch with schema that " f"does not match the schema of other batches.\n" f"Expected:\n{schema}\nGot:\n{batch.schema}" ) yield from table.to_batches() return pa.RecordBatchReader.from_batches(schema, gen()) def _sanitize_data( data: "DATA", target_schema: Optional[pa.Schema] = None, metadata: Optional[dict] = None, # embedding metadata on_bad_vectors: OnBadVectorsType = "error", fill_value: float = 0.0, *, allow_subschema: bool = False, ) -> pa.RecordBatchReader: """ Handle input data, applying all standard transformations. This includes: * Converting the data to a PyArrow Table * Adding vector columns defined in the metadata * Adding embedding metadata into the schema * Casting the table to the target schema * Handling bad vectors Parameters ---------- target_schema : Optional[pa.Schema], default None The schema to cast the table to. This is typically the schema of the table if it already exists. Otherwise it might be a user-requested schema. allow_subschema : bool, default False If True, the input table is allowed to omit columns from the target schema. The target schema will be filtered to only include columns that are present in the input table before casting. metadata : Optional[dict], default None The embedding metadata to add to the schema. on_bad_vectors : Literal["error", "drop", "fill", "null"], default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. fill_value : float, default 0.0 The value to use when filling vectors. Only used if on_bad_vectors="fill". NaN or null elements and missing trailing elements are replaced with this value. Valid elements are preserved. """ # At this point, the table might not match the schema we are targeting: # 1. There might be embedding columns missing that will be added # in the add_embeddings step. # 2. If `allow_subschemas` is True, there might be columns missing. reader = _into_pyarrow_reader(data, target_schema) reader = _append_vector_columns(reader, target_schema, metadata=metadata) # This happens before the cast so we can fix vector columns with # incorrect lengths before they are cast to FSL. reader = _handle_bad_vectors( reader, on_bad_vectors=on_bad_vectors, fill_value=fill_value, target_schema=target_schema, metadata=metadata, ) if target_schema is None: target_schema, reader = _infer_target_schema(reader) if metadata: target_schema = target_schema.with_metadata( _merge_metadata(target_schema.metadata, metadata) ) _validate_schema(target_schema) reader = _cast_to_target_schema(reader, target_schema, allow_subschema) return reader def _cast_to_target_schema( reader: pa.RecordBatchReader, target_schema: pa.Schema, allow_subschema: bool = False, ) -> pa.RecordBatchReader: # pa.Table.cast expects field order not to be changed. # Lance doesn't care about field order, so we don't need to rearrange fields # to match the target schema. We just need to correctly cast the fields. if reader.schema.equals(target_schema, check_metadata=True): # Fast path when the schemas are already the same return reader fields = _align_field_types(list(iter(reader.schema)), list(iter(target_schema))) reordered_schema = pa.schema(fields, metadata=target_schema.metadata) if not allow_subschema and len(reordered_schema) != len(target_schema): raise ValueError( "Input table has different number of columns than target schema" ) if allow_subschema and len(reordered_schema) != len(target_schema): fields = _infer_subschema( list(iter(reader.schema)), list(iter(reordered_schema)) ) reordered_schema = pa.schema(fields, metadata=target_schema.metadata) def gen(): for batch in reader: batch = _coerce_blob_write_columns(batch, reordered_schema) # Table but not RecordBatch has cast. cast_batches = ( pa.Table.from_batches([batch]).cast(reordered_schema).to_batches() ) if cast_batches: yield pa.RecordBatch.from_arrays( cast_batches[0].columns, schema=reordered_schema ) return pa.RecordBatchReader.from_batches(reordered_schema, gen()) def _preserve_blob_v2_fields( inferred_schema: pa.Schema, existing_schema: pa.Schema ) -> pa.Schema: """Keep inferred blob semantics without overriding explicit replacement types.""" existing_fields = {field.name: field for field in existing_schema} fields = [ _preserve_blob_v2_field(field, existing_fields[field.name]) if field.name in existing_fields else field for field in inferred_schema ] return pa.schema(fields, metadata=inferred_schema.metadata) def _preserve_blob_v2_field(inferred: pa.Field, existing: pa.Field) -> pa.Field: if is_blob_v2_field(inferred): # An explicitly declared blob layout is part of the replacement schema. return inferred if is_blob_v2_field(existing): return existing inferred_type, existing_type = inferred.type, existing.type if pa.types.is_struct(inferred_type) and pa.types.is_struct(existing_type): existing_children = {field.name: field for field in existing_type} children = [ _preserve_blob_v2_field(field, existing_children[field.name]) if field.name in existing_children else field for field in inferred_type ] return inferred.with_type(pa.struct(children)) if _is_list_like(inferred_type) and _is_list_like(existing_type): value_field = _preserve_blob_v2_field( inferred_type.value_field, existing_type.value_field ) if value_field == inferred_type.value_field: return inferred value_field = value_field.with_name(existing_type.value_field.name) if pa.types.is_fixed_size_list(inferred_type): value_type = pa.list_(value_field, inferred_type.list_size) elif pa.types.is_large_list(inferred_type): value_type = pa.large_list(value_field) else: value_type = pa.list_(value_field) return inferred.with_type(value_type) if blob_v2_column_paths(pa.schema([existing])): return existing return inferred def _coerce_blob_write_columns( batch: pa.RecordBatch, target_schema: pa.Schema ) -> pa.RecordBatch: """Materialize Python bytes and URIs as blob structs before Arrow casting. Existing blob structs retain their input layout. Rust normalizes them to the stored schema and rejects populated ranges on two-field targets. """ columns = [] fields = [] changed = False for field, column in zip(batch.schema, batch.columns): target_field = target_schema.field(field.name) coerced = _coerce_blob_value(column, target_field) if coerced is not column: column = coerced field = pa.field( field.name, coerced.type, field.nullable, target_field.metadata, ) changed = True columns.append(column) fields.append(field) if not changed: return batch return pa.RecordBatch.from_arrays( columns, schema=pa.schema(fields, metadata=batch.schema.metadata) ) def _coerce_blob_value(column: pa.Array, target_field: pa.Field) -> pa.Array: if is_blob_v2_field(target_field) and _can_coerce_to_blob(column.type): return _coerce_value_to_blob(column, target_field) target_type = target_field.type if pa.types.is_struct(target_type) and pa.types.is_struct(column.type): children = [] fields = [] changed = False for source_field in column.type: source_column = column.field(source_field.name) nested_target = next( (field for field in target_type if field.name == source_field.name), None, ) if nested_target is None: children.append(source_column) fields.append(source_field) continue coerced = _coerce_blob_value(source_column, nested_target) if coerced is not source_column: changed = True child_array, child_type = _physical_array_and_type(coerced) children.append(child_array) fields.append( pa.field( source_field.name, child_type, source_field.nullable, nested_target.metadata, ) ) if not changed: return column return pa.StructArray.from_arrays( children, fields=fields, mask=column.is_null() if column.null_count else None, ) if _is_list_like(target_type) and _is_list_like(column.type): return _coerce_blob_list_values(column, target_type.value_field) return column def _coerce_blob_list_values( column: pa.Array, target_value_field: pa.Field ) -> pa.Array: """Coerce blob values inside a list column, preserving offsets and nulls. Works on the raw child values window instead of ``pc.list_flatten`` because flatten drops values spanned by null slots, which would misalign offsets. """ mask = column.is_null() if column.null_count else None if pa.types.is_fixed_size_list(column.type): list_size = column.type.list_size values = column.values.slice(column.offset * list_size, len(column) * list_size) coerced = _coerce_blob_value(values, target_value_field) if coerced is values: return column physical_values, _ = _physical_array_and_type(coerced) return pa.FixedSizeListArray.from_arrays(physical_values, list_size, mask=mask) offsets = column.offsets first_offset = offsets[0].as_py() values = column.values.slice( first_offset, offsets[-1].as_py() - first_offset, ) coerced = _coerce_blob_value(values, target_value_field) if coerced is values: return column physical_values, _ = _physical_array_and_type(coerced) if first_offset: offsets = pc.subtract(offsets, pa.scalar(first_offset, offsets.type)) if pa.types.is_large_list(column.type): return pa.LargeListArray.from_arrays(offsets, physical_values, mask=mask) return pa.ListArray.from_arrays(offsets, physical_values, mask=mask) def _coerce_value_to_blob(values: pa.Array, target_field: pa.Field) -> pa.Array: value_field = "uri" if _is_string_like(values.type) else "data" length = len(values) storage_type = target_field.type if isinstance(storage_type, pa.ExtensionType): storage_type = storage_type.storage_type storage_fields = list(storage_type) children = [] for storage_field in storage_fields: if storage_field.name == value_field and not pa.types.is_null(values.type): if values.type == storage_field.type: children.append(values) else: children.append(values.cast(storage_field.type)) else: children.append(pa.nulls(length, type=storage_field.type)) storage = pa.StructArray.from_arrays( children, fields=storage_fields, mask=values.is_null() if values.null_count else None, ) if isinstance(target_field.type, pa.ExtensionType): return pa.ExtensionArray.from_storage(target_field.type, storage) return storage def _physical_array_and_type(array: pa.Array) -> tuple[pa.Array, pa.DataType]: if isinstance(array.type, pa.ExtensionType): return array.storage, array.type.storage_type return array, array.type def _can_coerce_to_blob(data_type: pa.DataType) -> bool: return ( _is_binary_like(data_type) or _is_string_like(data_type) or pa.types.is_null(data_type) ) def _is_string_like(data_type: pa.DataType) -> bool: return ( pa.types.is_string(data_type) or pa.types.is_large_string(data_type) or pa.types.is_string_view(data_type) ) def _is_binary_like(data_type: pa.DataType) -> bool: return ( pa.types.is_binary(data_type) or pa.types.is_large_binary(data_type) or pa.types.is_binary_view(data_type) ) def _field_extension_name(field: pa.Field) -> Optional[str]: extension_name = getattr(field.type, "extension_name", None) if extension_name is not None: return extension_name metadata = field.metadata or {} extension_name = metadata.get(b"ARROW:extension:name") or metadata.get( "ARROW:extension:name" ) if isinstance(extension_name, bytes): return extension_name.decode() return extension_name def _is_json_field(field: pa.Field) -> bool: return _field_extension_name(field) in ("arrow.json", "lance.json") @dataclass(frozen=True) class _JsonSerializationPlan: arrow_field: pa.Field children: Optional[Dict[str, "_JsonSerializationPlan"]] = None item: Optional["_JsonSerializationPlan"] = None def _json_serialization_plan(field: pa.Field) -> Optional[_JsonSerializationPlan]: if _is_json_field(field): return _JsonSerializationPlan(field) if pa.types.is_struct(field.type): children: Dict[str, _JsonSerializationPlan] = {} for child_field in field.type: child_plan = _json_serialization_plan(child_field) if child_plan is not None: children[child_field.name] = child_plan if children: return _JsonSerializationPlan(field, children=children) if _is_list_like(field.type): item_plan = _json_serialization_plan(field.type.value_field) if item_plan is not None: return _JsonSerializationPlan(field, item=item_plan) return None def _json_serialization_plans( schema: pa.Schema, ) -> Dict[str, _JsonSerializationPlan]: plans: Dict[str, _JsonSerializationPlan] = {} for field in schema: plan = _json_serialization_plan(field) if plan is not None: plans[field.name] = plan return plans def _serialize_json_value(value: Any, plan: _JsonSerializationPlan) -> Any: if value is None or isinstance(value, str): return value if _is_json_field(plan.arrow_field): if isinstance(value, (dict, list)): return json.dumps(value) return value if plan.children is not None and isinstance(value, dict): serialized = None for child_name, child_plan in plan.children.items(): if child_name not in value: continue child_value = _serialize_json_value(value[child_name], child_plan) if child_value is not value[child_name]: if serialized is None: serialized = dict(value) serialized[child_name] = child_value return serialized if serialized is not None else value if plan.item is not None and isinstance(value, list): serialized = None for index, item in enumerate(value): serialized_item = _serialize_json_value(item, plan.item) if serialized_item is not item: if serialized is None: serialized = list(value) serialized[index] = serialized_item return serialized if serialized is not None else value return value def _serialize_json_values(data: Any, target_schema: Optional[pa.Schema]) -> Any: if target_schema is None or not isinstance(data, list): return data plans = _json_serialization_plans(target_schema) if not plans: return data serialized_rows = [] for row in data: if not isinstance(row, dict): serialized_rows.append(row) continue serialized_row = None for field_name, plan in plans.items(): if field_name not in row: continue value = _serialize_json_value(row[field_name], plan) if value is not row[field_name]: if serialized_row is None: serialized_row = dict(row) serialized_row[field_name] = value serialized_rows.append(serialized_row if serialized_row is not None else row) return serialized_rows def _align_field_types( fields: List[pa.Field], target_fields: List[pa.Field], ) -> List[pa.Field]: """ Apply the data types from the target_fields to the fields. """ new_fields = [] for field in fields: target_field = next((f for f in target_fields if f.name == field.name), None) if target_field is None: raise ValueError(f"Field '{field.name}' not found in target schema") new_fields.append(_align_field(field, target_field)) return new_fields def _align_container_child(child: pa.Field, target_child: pa.Field) -> pa.Field: # A list has one child, a map one key and one item, so an inferred child name # ("item") aligns positionally and adopts the table's; pa.Table.cast renames it. return _align_field(child, target_child).with_name(target_child.name) def _arrow_json_storage_type(input_type: pa.DataType) -> Optional[pa.DataType]: """The storage type arrow.json would use for ``input_type``. Returns None if the type cannot hold JSON text. """ if pa.types.is_string(input_type) or pa.types.is_string_view(input_type): return pa.string() if pa.types.is_large_string(input_type): return pa.large_string() return None def _align_field(field: pa.Field, target_field: pa.Field) -> pa.Field: if is_blob_v2_field(target_field): source_type = field.type if pa.types.is_struct(source_type): if not any(child.name in ("data", "uri") for child in source_type): raise ValueError( f"blob struct input for column '{field.name}' must contain " "a 'data' or 'uri' child" ) elif not _can_coerce_to_blob(source_type) and not isinstance( source_type, pa.ExtensionType ): raise ValueError( f"cannot coerce column '{field.name}' with type {source_type} " "into a blob v2 struct. expected binary bytes " "(Binary, LargeBinary, BinaryView), strings " "(Utf8, LargeUtf8, Utf8View), or a Struct with a " "'data' or 'uri' child" ) if pa.types.is_struct(source_type) or ( isinstance(source_type, pa.ExtensionType) and ( not isinstance(target_field.type, pa.ExtensionType) or type(source_type) is type(target_field.type) ) ): # Keep the source layout until Rust can fill missing range fields or # validate every batch before dropping all-null range fields. Arrow # casting here would reject differing layouts before that validation. return field.with_metadata(target_field.metadata) # LanceDB exposes stored JSON columns as lance.json (JSONB-backed LargeBinary), but # casting the input to that storage type here merely relabels the raw JSON bytes as # JSONB. Lance must see arrow.json so it can perform the JSONB encoding. if _field_extension_name(target_field) == "lance.json": if _field_extension_name(field) == "arrow.json": return field # Plain JSON text, which is what pyarrow infers for a column of `str`, only # needs the arrow.json label. json_storage = _arrow_json_storage_type(field.type) if json_storage is not None: # Labelled through metadata rather than pa.json_(), which only exists on # newer PyArrow; Lance reads the extension name off the field either way. # The other metadata keys mirror the table field's: Lance swaps the name # back to lance.json on write and then requires the field to match the # stored one exactly, including the empty ``ARROW:extension:metadata`` # that pyarrow records for a ``pa.json_()`` column. metadata = dict(target_field.metadata or {}) metadata[b"ARROW:extension:name"] = b"arrow.json" return pa.field(field.name, json_storage, field.nullable, metadata) if pa.types.is_struct(target_field.type): if pa.types.is_struct(field.type): new_type = pa.struct( _align_field_types( field.type.fields, target_field.type.fields, ) ) else: new_type = target_field.type elif pa.types.is_list(target_field.type): if _is_list_like(field.type): new_type = pa.list_( _align_container_child( field.type.value_field, target_field.type.value_field ) ) else: new_type = target_field.type elif pa.types.is_large_list(target_field.type): if _is_list_like(field.type): new_type = pa.large_list( _align_container_child( field.type.value_field, target_field.type.value_field ) ) else: new_type = target_field.type elif pa.types.is_fixed_size_list(target_field.type): if _is_list_like(field.type): new_type = pa.list_( _align_container_child( field.type.value_field, target_field.type.value_field ), target_field.type.list_size, ) else: new_type = target_field.type elif pa.types.is_map(target_field.type): if pa.types.is_map(field.type): # A map has exactly one key and one item field, so like a list's child they # align positionally and adopt the table's names. new_type = pa.map_( _align_container_child( field.type.key_field, target_field.type.key_field ), _align_container_child( field.type.item_field, target_field.type.item_field ), keys_sorted=target_field.type.keys_sorted, ) else: new_type = target_field.type else: new_type = target_field.type return pa.field(field.name, new_type, field.nullable, target_field.metadata) def _infer_subschema( schema: List[pa.Field], reference_fields: List[pa.Field], ) -> List[pa.Field]: """ Transform the list of fields so the types match the reference_fields. The order of the fields is preserved. ``schema`` may have fewer fields than `reference_fields`, but it may not have more fields. """ fields = [] lookup = {f.name: f for f in reference_fields} for field in schema: reference = lookup.get(field.name) if reference is None: raise ValueError("Unexpected field in schema: {}".format(field)) if pa.types.is_struct(reference.type): new_type = pa.struct( _infer_subschema( field.type.fields, reference.type.fields, ) ) new_field = pa.field( field.name, new_type, reference.nullable, ) else: new_field = reference fields.append(new_field) return fields def sanitize_create_table( data, schema: Union[pa.Schema, LanceModel], metadata=None, on_bad_vectors: OnBadVectorsType = "error", fill_value: float = 0.0, ): if inspect.isclass(schema) and issubclass(schema, LanceModel): # convert LanceModel to pyarrow schema # note that it's possible this contains # embedding function metadata already schema: pa.Schema = schema.to_arrow_schema() if data is not None: if metadata is None and schema is not None: metadata = schema.metadata data = _sanitize_data( data, schema, metadata=metadata, on_bad_vectors=on_bad_vectors, fill_value=fill_value, ) schema = data.schema else: if schema is not None: data = pa.Table.from_batches([], schema=schema) if schema is None: if data is None: raise ValueError("Either data or schema must be provided") elif hasattr(data, "schema"): schema = data.schema if metadata: metadata = _merge_metadata(schema.metadata, metadata) schema = schema.with_metadata(metadata) # Need to apply metadata to the data as well if isinstance(data, pa.Table): data = data.replace_schema_metadata(metadata) elif isinstance(data, pa.RecordBatchReader): data = pa.RecordBatchReader.from_batches(schema, data) return data, schema def _schema_from_hf(data, schema) -> pa.Schema: """ Extract pyarrow schema from HuggingFace DatasetDict and validate that they're all the same schema between splits """ for dataset in data.values(): if schema is None: schema = dataset.features.arrow_schema elif schema != dataset.features.arrow_schema: msg = "All datasets in a HuggingFace DatasetDict must have the same schema" raise TypeError(msg) return schema def _to_batches_with_split(data): """ Return a generator of RecordBatches from a HuggingFace DatasetDict with an extra `split` column """ for key, dataset in data.items(): for batch in _hf_batches(dataset): table = pa.Table.from_batches([batch]) if "split" not in table.column_names: table = table.append_column( "split", pa.array([key] * batch.num_rows, pa.string()) ) for b in table.to_batches(): yield b def _append_vector_columns( reader: pa.RecordBatchReader, schema: Optional[pa.Schema] = None, *, metadata: Optional[dict] = None, ) -> pa.RecordBatchReader: """ Use the embedding function to automatically embed the source columns and add the vector columns to the table. """ if schema is None: metadata = _merge_metadata(metadata) else: metadata = _merge_metadata(schema.metadata, metadata) functions = EmbeddingFunctionRegistry.get_instance().parse_functions(metadata) if not functions: return reader fields = list(reader.schema) for vector_column, conf in functions.items(): if vector_column not in reader.schema.names: if schema is not None: field = schema.field(vector_column) else: dtype = pa.list_(pa.float32(), conf.function.ndims()) field = pa.field(vector_column, type=dtype, nullable=True) fields.append(field) schema = pa.schema(fields, metadata=reader.schema.metadata) def gen(): for batch in reader: for vector_column, conf in functions.items(): func = conf.function no_vector_column = vector_column not in batch.column_names if no_vector_column or pc.all(pc.is_null(batch[vector_column])).as_py(): col_data = func.compute_source_embeddings_with_retry( batch[conf.source_column] ) # Replace vectors with wrong length (including empty lists # returned for inputs like empty strings) with None so that # _handle_bad_vectors can process them according to the # on_bad_vectors policy instead of crashing when PyArrow # tries to cast them into a fixed-size list array. expected_ndims = conf.function.ndims() col_data = [ v if v is not None and len(v) == expected_ndims else None for v in col_data ] if no_vector_column: batch = batch.append_column( schema.field(vector_column), pa.array(col_data, type=schema.field(vector_column).type), ) else: batch = batch.set_column( batch.column_names.index(vector_column), schema.field(vector_column), pa.array(col_data, type=schema.field(vector_column).type), ) yield batch return pa.RecordBatchReader.from_batches(schema, gen()) def _table_path(base: str, table_name: str) -> str: """ Get a table path that can be used in PyArrow FS. Removes any weird schemes (such as "s3+ddb") and drops any query params. """ uri = _table_uri(base, table_name) # Parse as URL parsed = urlparse(uri) # If scheme is s3+ddb, convert to s3 if parsed.scheme == "s3+ddb": parsed = parsed._replace(scheme="s3") # Remove query parameters return parsed._replace(query=None).geturl() def _table_uri(base: str, table_name: str) -> str: return join_uri(base, f"{table_name}.lance") def _normalize_progress(progress): """Normalize a ``progress`` parameter for :meth:`Table.add`. Returns ``(progress_obj, owns)`` where *owns* is True when we created a tqdm bar that the caller must close. """ if progress is True: from tqdm.auto import tqdm return tqdm(unit=" rows"), True if progress is False or progress is None: return None, False return progress, False class Table(ABC): """ A Table is a collection of Records in a LanceDB Database. Examples -------- Create using [DBConnection.create_table][lancedb.DBConnection.create_table] (more examples in that method's documentation). >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data=[{"vector": [1.1, 1.2], "b": 2}]) >>> table.head() pyarrow.Table vector: fixed_size_list[2] child 0, item: float b: int64 ---- vector: [[[1.1,1.2]]] b: [[2]] Can append new data with [Table.add()][lancedb.table.Table.add]. >>> table.add([{"vector": [0.5, 1.3], "b": 4}]) AddResult(version=2) Can query the table with [Table.search][lancedb.table.Table.search]. >>> table.search([0.4, 0.4]).select(["b", "vector"]).to_pandas() b vector _distance 0 4 [0.5, 1.3] 0.82 1 2 [1.1, 1.2] 1.13 Search queries are much faster when an index is created. See [Table.create_index][lancedb.table.Table.create_index]. """ @property @abstractmethod def name(self) -> str: """The name of this Table""" raise NotImplementedError("name is not supported for this table type") @property def namespace(self) -> List[str]: """Return the namespace path of the table (empty for the root namespace).""" return [] @property def id(self) -> str: """Return the full identifier of the table (namespace$name).""" return "$".join(self.namespace + [self.name]) @property @abstractmethod def version(self) -> int: """The version of this Table""" raise NotImplementedError("version is not supported for this table type") @property @abstractmethod def schema(self) -> pa.Schema: """The [Arrow Schema](https://arrow.apache.org/docs/python/api/datatypes.html#) of this Table """ raise NotImplementedError("schema is not supported for this table type") @property @abstractmethod def tags(self) -> Tags: """Tag management for the table. Similar to Git, tags are a way to add metadata to a specific version of the table. .. warning:: Tagged versions are exempted from the :py:meth:`cleanup_old_versions()` process. To remove a version that has been tagged, you must first :py:meth:`~Tags.delete` the associated tag. Examples -------- .. code-block:: python table = db.open_table("my_table") table.tags.create("v2-prod-20250203", 10) tags = table.tags.list() """ raise NotImplementedError("tags is not supported for this table type") @property def branches(self) -> "Branches": """Branch management for the table. Branches are isolated, writable lines of history forked from another branch (or version). Writes on a branch do not affect ``main``. """ raise NotImplementedError("branches is not supported for this table type") def current_branch(self) -> Optional[str]: """The branch this table handle is scoped to, or ``None`` for ``main``.""" raise NotImplementedError("current_branch is not supported for this table type") def __len__(self) -> int: """The number of rows in this Table""" return self.count_rows(None) @property @abstractmethod def embedding_functions(self) -> Dict[str, EmbeddingFunctionConfig]: """ Get a mapping from vector column name to it's configured embedding function. """ @abstractmethod def count_rows(self, filter: Optional[str] = None) -> int: """ Count the number of rows in the table. Parameters ---------- filter: str, optional A SQL where clause to filter the rows to count. """ raise NotImplementedError("count_rows is not supported for this table type") def to_pandas(self, blob_mode: BlobMode = "lazy", **kwargs) -> "pandas.DataFrame": """Return the table as a pandas DataFrame. Parameters ---------- blob_mode: str, default "lazy" Controls how blob columns are returned for backends that support Lance blob-aware pandas conversion. **kwargs Forwarded to PyArrow / Lance pandas conversion. Returns ------- pd.DataFrame """ return self.to_arrow().to_pandas(**kwargs) @abstractmethod def to_arrow(self, *, blob_mode: BlobMode = "lazy") -> pa.Table: """Return the table as a pyarrow Table. Parameters ---------- blob_mode: str, default "lazy" ``"lazy"`` returns blob v2 descriptors that carry the row id ``fetch_blobs`` reads. ``"descriptions"`` returns them without it. ``"bytes"`` returns payload bytes and raises for legacy blob columns and blobs inside lists. Returns ------- pa.Table """ raise NotImplementedError("to_arrow is not supported for this table type") def to_lance(self, **kwargs) -> lance.LanceDataset: """Return a local table as a lance.LanceDataset. Remote tables cannot be opened as Lance datasets by the client. Use ``table.search().to_arrow()`` to read them through the server. Returns ------- lance.LanceDataset """ raise NotImplementedError("to_lance is not supported for this table type") def to_polars(self, batch_size=None) -> "pl.LazyFrame": """Return the table as a Polars LazyFrame. For remote tables, use ``table.search().to_polars()`` to convert query results to a Polars DataFrame. Parameters ---------- batch_size: int, optional Passed to Polars as the maximum row count for scanned PyArrow record batches. Note ---- The Polars streaming engine is not supported because it does not currently implement Python PyArrow dataset scans. Use the default engine when collecting this LazyFrame. Returns ------- polars.LazyFrame """ raise NotImplementedError("to_polars is not supported for this table type") def _is_legacy_create_index_call( self, first_arg: str, config: Optional[IndexConfigType], num_partitions: Optional[int], num_sub_vectors: Optional[int], vector_column_name: str, accelerator: Optional[str], index_cache_size: Optional[int], ) -> bool: """Detect if this is a legacy create_index call.""" # If config is provided, it's definitely the new API if config is not None: return False # If old-style parameters were explicitly set, it's legacy if any( x is not None for x in (num_partitions, num_sub_vectors, accelerator, index_cache_size) ): return True # If vector_column_name differs from default, it's legacy if vector_column_name != VECTOR_COLUMN_NAME: return True # If first arg is a known metric, assume legacy if first_arg.lower() in KNOWN_METRICS: return True # Otherwise assume new API return False def _build_vector_config_from_legacy_params( self, metric: str, index_type: str, num_partitions: Optional[int], num_sub_vectors: Optional[int], num_bits: int, max_iterations: int, sample_rate: int, m: int, ef_construction: int, target_partition_size: Optional[int], accelerator: Optional[str], ) -> IndexConfigType: """Build an index config object from legacy parameters.""" if index_type == "IVF_FLAT": return IvfFlat( distance_type=metric, num_partitions=num_partitions, max_iterations=max_iterations, sample_rate=sample_rate, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_SQ": return IvfSq( distance_type=metric, num_partitions=num_partitions, max_iterations=max_iterations, sample_rate=sample_rate, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_PQ": return IvfPq( distance_type=metric, num_partitions=num_partitions, num_sub_vectors=num_sub_vectors, num_bits=num_bits, max_iterations=max_iterations, sample_rate=sample_rate, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_RQ": return IvfRq( distance_type=metric, num_partitions=num_partitions, num_bits=num_bits, max_iterations=max_iterations, sample_rate=sample_rate, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_HNSW_PQ": return HnswPq( distance_type=metric, num_partitions=num_partitions, num_sub_vectors=num_sub_vectors, num_bits=num_bits, max_iterations=max_iterations, sample_rate=sample_rate, m=m, ef_construction=ef_construction, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_HNSW_SQ": return HnswSq( distance_type=metric, num_partitions=num_partitions, max_iterations=max_iterations, sample_rate=sample_rate, m=m, ef_construction=ef_construction, target_partition_size=target_partition_size, accelerator=accelerator, ) elif index_type == "IVF_HNSW_FLAT": return HnswFlat( distance_type=metric, num_partitions=num_partitions, max_iterations=max_iterations, sample_rate=sample_rate, m=m, ef_construction=ef_construction, target_partition_size=target_partition_size, ) else: raise ValueError(f"Unknown index type {index_type}") @staticmethod def infer_tokenizer_configs(tokenizer_name: str) -> dict: if tokenizer_name == "default": return { "base_tokenizer": "simple", "language": "English", "max_token_length": 40, "lower_case": True, "stem": False, "remove_stop_words": False, "ascii_folding": False, "ngram_min_length": 3, "ngram_max_length": 3, "prefix_only": False, } elif tokenizer_name == "raw": return { "base_tokenizer": "raw", "language": "English", "max_token_length": None, "lower_case": False, "stem": False, "remove_stop_words": False, "ascii_folding": False, "ngram_min_length": 3, "ngram_max_length": 3, "prefix_only": False, } elif tokenizer_name == "whitespace": return { "base_tokenizer": "whitespace", "language": "English", "max_token_length": None, "lower_case": False, "stem": False, "remove_stop_words": False, "ascii_folding": False, "ngram_min_length": 3, "ngram_max_length": 3, "prefix_only": False, } # or it's with language stemming with pattern like "en_stem" if len(tokenizer_name) != 7: raise ValueError(f"Invalid tokenizer name {tokenizer_name}") lang = tokenizer_name[:2] if tokenizer_name[-5:] != "_stem": raise ValueError(f"Invalid tokenizer name {tokenizer_name}") if lang not in lang_mapping: raise ValueError(f"Invalid language code {lang}") return { "base_tokenizer": "simple", "language": lang_mapping[lang], "max_token_length": 40, "lower_case": True, "stem": True, "remove_stop_words": False, "ascii_folding": False, "ngram_min_length": 3, "ngram_max_length": 3, "prefix_only": False, } # New unified API overload @overload def create_index( self, column: str, /, *, config: IndexConfigType, replace: bool = ..., wait_timeout: Optional[timedelta] = ..., name: Optional[str] = ..., train: bool = ..., ) -> None: ... # Legacy API overload (deprecated) @overload def create_index( self, metric: Literal["l2", "cosine", "dot", "hamming"] = ..., num_partitions: Optional[int] = ..., num_sub_vectors: Optional[int] = ..., vector_column_name: str = ..., replace: bool = ..., accelerator: Optional[str] = ..., index_cache_size: Optional[int] = ..., num_bits: int = ..., index_type: Literal[ "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_RQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ" ] = ..., max_iterations: int = ..., sample_rate: int = ..., m: int = ..., ef_construction: int = ..., *, wait_timeout: Optional[timedelta] = ..., name: Optional[str] = ..., train: bool = ..., target_partition_size: Optional[int] = ..., ) -> None: ... def create_index( self, metric: str = "l2", num_partitions: Optional[int] = None, num_sub_vectors: Optional[int] = None, vector_column_name: str = VECTOR_COLUMN_NAME, replace: bool = True, accelerator: Optional[str] = None, index_cache_size: Optional[int] = None, num_bits: int = 8, index_type: Literal[ "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_RQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ", "IVF_HNSW_FLAT", ] = "IVF_PQ", max_iterations: int = 50, sample_rate: int = 256, m: int = 20, ef_construction: int = 300, *, config: Optional[IndexConfigType] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, target_partition_size: Optional[int] = None, ): """Create an index on a column. This method supports both the new unified API and the legacy API for backwards compatibility. The new API takes the column name as the first positional argument and an index configuration object via ``config``; the legacy API takes the distance metric as the first argument plus separate ``vector_column_name`` / ``num_partitions`` / etc. parameters, and emits a ``DeprecationWarning``. Parameters ---------- metric : str For new API: the column name to index. For legacy API: the distance metric ("l2", "cosine", "dot", "hamming"). config : IndexConfigType, optional The index configuration object. If provided, uses the new unified API. Can be one of: IvfFlat, IvfPq, IvfSq, IvfRq, HnswPq, HnswSq, BTree, Bitmap, LabelList, Fm, FTS. replace : bool, default True Whether to replace an existing index on this column. wait_timeout : timedelta, optional Timeout to wait for async indexing to complete. name : str, optional Custom name for the index. train : bool, default True Whether to train the index with existing data. Examples -------- New API (recommended): >>> table.create_index( # doctest: +SKIP ... "vector", config=IvfPq(distance_type="l2") ... ) >>> table.create_index("category", config=BTree()) # doctest: +SKIP >>> table.create_index("content", config=FTS()) # doctest: +SKIP Legacy API (deprecated): >>> table.create_index( # doctest: +SKIP ... "l2", vector_column_name="vector" ... ) """ raise NotImplementedError("create_index is not supported for this table type") def create_index_async( self, column: str, *, config: IndexConfigType, replace: Optional[bool] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, ) -> Job: """Create an index, returning a handle to the indexing job. Takes the same arguments as :meth:`create_index`. The job may already be complete when returned; callers must not assume the index exists until :meth:`Job.wait` returns. """ raise NotImplementedError( "create_index_async is not supported for this table type" ) def drop_index(self, name: str) -> None: """ Drop an index from the table. Parameters ---------- name: str The name of the index to drop. Notes ----- This does not delete the index from disk, it just removes it from the table. To delete the index, run [optimize][lancedb.table.Table.optimize] after dropping the index. Use [list_indices][lancedb.table.Table.list_indices] to find the names of the indices. """ raise NotImplementedError("drop_index is not supported for this table type") def wait_for_index( self, index_names: Iterable[str], timeout: timedelta = timedelta(seconds=300) ) -> None: """ Wait for indexing to complete for the given index names. This will poll the table until all the indices are fully indexed, or raise a timeout exception if the timeout is reached. Parameters ---------- index_names: str The name of the indices to poll timeout: timedelta Timeout to wait for asynchronous indexing. The default is 5 minutes. """ raise NotImplementedError("wait_for_index is not supported for this table type") @abstractmethod def stats(self) -> TableStatistics: """ Retrieve table and fragment statistics. """ raise NotImplementedError("stats is not supported for this table type") @abstractmethod def create_scalar_index( self, column: str, *, replace: bool = True, index_type: ScalarIndexType = "BTREE", wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, ): """Create a scalar index on a column. Parameters ---------- column : str The column to be indexed. Must be a boolean, integer, float, or string column. replace : bool, default True Replace the existing index if it exists. index_type: Literal["BTREE", "BITMAP", "LABEL_LIST"], default "BTREE" The type of index to create. wait_timeout: timedelta, optional The timeout to wait if indexing is asynchronous. name: str, optional The name of the index. If not provided, a default name will be generated. Examples -------- Scalar indices, like vector indices, can be used to speed up scans. A scalar index can speed up scans that contain filter expressions on the indexed column. For example, the following scan will be faster if the column ``my_col`` has a scalar index: >>> import lancedb # doctest: +SKIP >>> db = lancedb.connect("/data/lance") # doctest: +SKIP >>> img_table = db.open_table("images") # doctest: +SKIP >>> my_df = img_table.search().where("my_col = 7", # doctest: +SKIP ... prefilter=True).to_pandas() Scalar indices can also speed up scans containing a vector search and a prefilter: >>> import lancedb # doctest: +SKIP >>> db = lancedb.connect("/data/lance") # doctest: +SKIP >>> img_table = db.open_table("images") # doctest: +SKIP >>> img_table.search([1, 2, 3, 4], vector_column_name="vector") # doctest: +SKIP ... .where("my_col != 7", prefilter=True) ... .to_pandas() Scalar indices can only speed up scans for basic filters using equality, comparison, range (e.g. ``my_col BETWEEN 0 AND 100``), and set membership (e.g. `my_col IN (0, 1, 2)`) Scalar indices can be used if the filter contains multiple indexed columns and the filter criteria are AND'd or OR'd together (e.g. ``my_col < 0 AND other_col> 100``) Scalar indices may be used if the filter contains non-indexed columns but, depending on the structure of the filter, they may not be usable. For example, if the column ``not_indexed`` does not have a scalar index then the filter ``my_col = 0 OR not_indexed = 1`` will not be able to use any scalar index on ``my_col``. """ raise NotImplementedError( "create_scalar_index is not supported for this table type" ) def create_fts_index( self, field_names: Union[str, List[str]], *, ordering_field_names: Optional[Union[str, List[str]]] = None, replace: bool = False, writer_heap_size: Optional[int] = 1024 * 1024 * 1024, use_tantivy: bool = False, tokenizer_name: Optional[str] = None, with_position: bool = False, # tokenizer configs: base_tokenizer: BaseTokenizerType = "simple", language: str = "English", max_token_length: Optional[int] = 40, lower_case: bool = True, stem: Optional[bool] = None, remove_stop_words: Optional[bool] = None, custom_stop_words: Optional[List[str]] = None, ascii_folding: bool = True, ngram_min_length: int = 3, ngram_max_length: int = 3, prefix_only: bool = False, block_size: int = 128, document_granularity: DocumentGranularity = DocumentGranularity.ROW, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, ): """Create a full-text search index on the table. Warning - this API is highly experimental and is highly likely to change in the future. Parameters ---------- field_names: str or list of str The name of the field to index. Native FTS indexes can only be created on a single field at a time. To search over multiple text fields, create a separate FTS index for each field. replace: bool, default False If True, replace the existing index if it exists. Note that this is not yet an atomic operation; the index will be temporarily unavailable while the new index is being created. writer_heap_size: int, default 1GB Deprecated legacy Tantivy parameter. Any value other than the default raises an error. ordering_field_names: Deprecated legacy Tantivy parameter. Setting this raises an error. tokenizer_name: str, default "default" A compatibility alias for native tokenizer configs. Can be "raw", "default" or the 2 letter language code followed by "_stem". So for english it would be "en_stem". For new native FTS indexes, use ``base_tokenizer`` directly; ``tokenizer_name`` is a legacy compatibility alias and does not expose model-backed tokenizer names such as ``jieba/default`` or ``lindera/ipadic``. use_tantivy: bool, default False Deprecated legacy Tantivy parameter. Setting this to True raises an error. with_position: bool, default False If False, do not store the positions of the terms in the text. This can reduce the size of the index and improve indexing speed. But it will raise an exception for phrase queries. base_tokenizer : str, default "simple" The base tokenizer to use for tokenization. Options are: - "simple": Splits text by whitespace and punctuation. - "whitespace": Split text by whitespace, but not punctuation. - "raw": No tokenization. The entire text is treated as a single token. - "code": Tokenizes source code and identifiers. - "ngram": N-Gram tokenizer. - "icu": ICU dictionary-based word segmentation. - "icu/split": ICU segmentation with simple-style delimiter splitting. - "jieba/*": Jieba tokenizer loaded from Lance's language model home. - "lindera/*": Lindera tokenizer loaded from Lance's language model home. language : str, default "English" The language to use for stemming and stop-word removal. This is not the primary way to enable CJK tokenization. max_token_length : int, default 40 The maximum token length to index. Tokens longer than this length will be ignored. lower_case : bool, default True Whether to convert the token to lower case. This makes queries case-insensitive. stem : bool, optional Whether to stem the token. Stemming reduces words to their root form. For example, in English "running" and "runs" would both be reduced to "run". ``None`` uses the base tokenizer's default: False for ``code`` and ``ngram``, True otherwise. remove_stop_words : bool, optional Whether to remove stop words. Stop words are common words that are often removed from text before indexing. For example, in English "the" and "and". ``None`` uses the base tokenizer's default: False for ``code`` and ``ngram``, True otherwise. custom_stop_words : list of str, optional Custom words that replace the built-in language stop words. ``None`` uses the built-in list; an empty list explicitly uses no stop words. ascii_folding : bool, default True Whether to fold ASCII characters. This converts accented characters to their ASCII equivalent. For example, "café" would be converted to "cafe". ngram_min_length: int, default 3 The minimum length of an n-gram. ngram_max_length: int, default 3 The maximum length of an n-gram. prefix_only: bool, default False Whether to only index the prefix of the token for ngram tokenizer. block_size: int, default 128 The number of documents per compressed posting block. Must be 128 or 256. A value of 256 uses the experimental FTS V3 format and may introduce breaking changes. document_granularity: DocumentGranularity, default ROW ``ROW`` treats the selected text in one table row as one document. ``LIST_ELEMENT`` treats each element of the deepest list on the field path as one document and returns its physical coordinates in ``_doc_index`` for matching queries. wait_timeout: timedelta, optional The timeout to wait if indexing is asynchronous. name: str, optional The name of the index. If not provided, a default name will be generated. Notes ----- Model-backed tokenizers such as ``jieba/default`` and ``lindera/ipadic`` require tokenizer models in Lance's language model home. Set ``LANCE_LANGUAGE_MODEL_HOME`` to override the default platform data directory under ``lance/language_models``. """ raise NotImplementedError( "create_fts_index is not supported for this table type" ) @abstractmethod def add( self, data: DATA, mode: AddMode = "append", on_bad_vectors: OnBadVectorsType = "error", fill_value: float = 0.0, progress: Optional[Union[bool, Callable, Any]] = None, write_parallelism: Optional[int] = None, allow_external_blob_outside_bases: bool = False, ) -> AddResult: """Add more data to the [Table][lancedb.table.Table]. Parameters ---------- data: DATA The data to insert into the table. Acceptable types are: - list-of-dict - pandas.DataFrame - pyarrow.Table or pyarrow.RecordBatch mode: str The mode to use when writing the data. Valid values are "append" and "overwrite". Overwrite adopts explicitly declared blob layouts, while inferred bytes retain the existing blob v2 fields. on_bad_vectors: str, default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. One of "error", "drop", "fill", "null". fill_value: float, default 0. The value to use when filling vectors. Only used if on_bad_vectors="fill". progress: bool, callable, or tqdm-like, optional Progress reporting during the add operation. Can be: - ``True`` to automatically create and display a tqdm progress bar (requires ``tqdm`` to be installed):: table.add(data, progress=True) - A **callable** that receives a dict with keys ``output_rows``, ``output_bytes``, ``total_rows``, ``elapsed_seconds``, ``active_tasks``, ``total_tasks``, and ``done``:: def on_progress(p): print(f"{p['output_rows']}/{p['total_rows']} rows, " f"{p['active_tasks']}/{p['total_tasks']} workers") table.add(data, progress=on_progress) - A **tqdm-compatible** progress bar whose ``total`` and ``update()`` will be called automatically. The postfix shows write throughput (MB/s) and active worker count:: with tqdm() as pbar: table.add(data, progress=pbar) write_parallelism: int, optional Number of partitions to write in parallel. Higher values increase throughput but also peak memory use, since each partition buffers data in flight. Defaults to an estimate based on the data size, capped at the number of CPU cores. Lower this if bulk ingestion is using too much memory. allow_external_blob_outside_bases: bool, default False Store blob URIs that sit outside registered blob bases. The row keeps a reference, so the object has to stay readable. Local tables only. Returns ------- AddResult An object containing the new version number of the table after adding data. """ raise NotImplementedError("add is not supported for this table type") def merge_insert(self, on: Union[str, Iterable[str]]) -> LanceMergeInsertBuilder: """ Returns a [`LanceMergeInsertBuilder`][lancedb.merge.LanceMergeInsertBuilder] that can be used to create a "merge insert" operation This operation can add rows, update rows, and remove rows all in a single transaction. It is a very generic tool that can be used to create behaviors like "insert if not exists", "update or insert (i.e. upsert)", or even replace a portion of existing data with new data (e.g. replace all data where month="january") The merge insert operation works by combining new data from a **source table** with existing data in a **target table** by using a join. There are three categories of records. "Matched" records are records that exist in both the source table and the target table. "Not matched" records exist only in the source table (e.g. these are new data) "Not matched by source" records exist only in the target table (this is old data) The builder returned by this method can be used to customize what should happen for each category of data. Please note that the data may appear to be reordered as part of this operation. This is because updated rows will be deleted from the dataset and then reinserted at the end with the new values. Parameters ---------- on: Union[str, Iterable[str]] A column (or columns) to join on. This is how records from the source table and target table are matched. Typically this is some kind of key or id column. Passing several columns matches on the composite key: a source row updates a target row only when it agrees on every one of them. Examples -------- >>> import lancedb >>> data = pa.table({"a": [2, 1, 3], "b": ["a", "b", "c"]}) >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> new_data = pa.table({"a": [2, 3, 4], "b": ["x", "y", "z"]}) >>> # Perform a "upsert" operation >>> res = table.merge_insert("a") \\ ... .when_matched_update_all() \\ ... .when_not_matched_insert_all() \\ ... .execute(new_data) >>> res MergeResult(version=2, num_updated_rows=2, num_inserted_rows=1, num_deleted_rows=0, num_attempts=1, num_rows=3) >>> # The order of new rows is non-deterministic since we use >>> # a hash-join as part of this operation and so we sort here >>> table.to_arrow().sort_by("a").to_pandas() a b 0 1 b 1 2 x 2 3 y 3 4 z """ # noqa: E501 on = [on] if isinstance(on, str) else list(iter(on)) return LanceMergeInsertBuilder(self, on) @abstractmethod def search( self, query: Optional[ Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery] ] = None, vector_column_name: Optional[str] = None, query_type: QueryType = "auto", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceQueryBuilder: """Create a search query to find the nearest neighbors of the given query vector. We currently support [vector search](https://lancedb.com/docs/search/vector-search/) and [full-text search](https://lancedb.com/docs/search/full-text-search/). All query options are defined in [LanceQueryBuilder][lancedb.query.LanceQueryBuilder]. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> data = [ ... {"original_width": 100, "caption": "bar", "vector": [0.1, 2.3, 4.5]}, ... {"original_width": 2000, "caption": "foo", "vector": [0.5, 3.4, 1.3]}, ... {"original_width": 3000, "caption": "test", "vector": [0.3, 6.2, 2.6]} ... ] >>> table = db.create_table("my_table", data) >>> query = [0.4, 1.4, 2.4] >>> (table.search(query) ... .where("original_width > 1000", prefilter=True) ... .select(["caption", "original_width", "vector"]) ... .limit(2) ... .to_pandas()) caption original_width vector _distance 0 foo 2000 [0.5, 3.4, 1.3] 5.220000 1 test 3000 [0.3, 6.2, 2.6] 23.089996 Parameters ---------- query: list/np.ndarray/str/PIL.Image.Image, default None The targeted vector to search for. - *default None*. Acceptable types are: list, np.ndarray, PIL.Image.Image - If None then the select/where/limit clauses are applied to filter the table vector_column_name: str, optional The name of the vector column to search. The vector column needs to be a pyarrow fixed size list type - If not specified then the vector column is inferred from the table schema - If the table has multiple vector columns then the *vector_column_name* needs to be specified. Otherwise, an error is raised. query_type: str *default "auto"*. Acceptable types are: "vector", "fts", "hybrid", or "auto" - If "auto" then the query type is inferred from the query; - If `query` is a list/np.ndarray then the query type is "vector"; - If `query` is a PIL.Image.Image then either do vector search, or raise an error if no corresponding embedding function is found. - If `query` is a string, then the query type is "vector" if the table has embedding functions else the query type is "fts" Returns ------- LanceQueryBuilder A query builder object representing the query. Once executed, the query returns - selected columns - the vector - and also the "_distance" column which is the distance between the query vector and the returned vector. """ raise NotImplementedError("search is not supported for this table type") @abstractmethod def take_offsets( self, offsets: list[int], *, with_row_id: bool = False ) -> LanceTakeQueryBuilder: """ Take a list of offsets from the table. Offsets are 0-indexed and relative to the current version of the table. Offsets are not stable. A row with an offset of N may have a different offset in a different version of the table (e.g. if an earlier row is deleted). Offsets are mostly useful for sampling as the set of all valid offsets is easily known in advance to be [0, len(table)). No guarantees are made regarding the order in which results are returned. Repeated offsets produce repeated rows, which makes this method suitable for sampling with replacement. Parameters ---------- offsets: list[int] The offsets to take. with_row_id: bool, default False If True, include the "_rowid" column in the results. Returns ------- LanceTakeQueryBuilder A query builder that can be executed to get the rows at the given offsets. """ def __getitems__(self, offsets: list[int]) -> pa.RecordBatch: """ Take a list of offsets from the table and return as a record batch. This method uses the `take_offsets` method to take the rows. However, it aligns the offsets to the passed in offsets. This means the return type is a record batch (and so users should take care not to pass in too many offsets) Note: this method is primarily intended to fulfill the Dataset contract for pytorch. Parameters ---------- offsets: list[int] The offsets to take. Returns ------- pa.RecordBatch A record batch containing the rows at the given offsets. """ # We don't know the order of the results at all. So we calculate a permutation # for ordering the given offsets. Then we load the data with the _rowoffset # column. Then we sort by _rowoffset and apply the inverse of the permutation # that we calculated. # # Note: this is potentially a lot of memory copy if we're operating on large # batches :( num_offsets = len(offsets) indices = list(range(num_offsets)) permutation = sorted(indices, key=lambda idx: offsets[idx]) permutation_inv = [0] * num_offsets for i in range(num_offsets): permutation_inv[permutation[i]] = i columns = self.schema.names columns.append("_rowoffset") tbl = ( self.take_offsets(offsets) .select(columns) .to_arrow() .sort_by("_rowoffset") .take(permutation_inv) .combine_chunks() .drop_columns(["_rowoffset"]) ) return tbl @abstractmethod def take_row_ids( self, row_ids: list[int], *, with_row_id: bool = False ) -> LanceTakeQueryBuilder: """ Take a list of row ids from the table. Row ids are not stable and are relative to the current version of the table. They can change due to compaction and updates. No guarantees are made regarding the order in which results are returned. If you desire an output order that matches the order of the given ids, you will need to add the row id column to the output and align it yourself. Unlike offsets, row ids are not 0-indexed and no assumptions should be made about the possible range of row ids. In order to use this method you must first obtain the row ids by scanning or searching the table. Even so, row ids are more stable than offsets and can be useful in some situations. There is an ongoing effort to make row ids stable which is tracked at https://github.com/lancedb/lancedb/issues/1120 Parameters ---------- row_ids: list[int] The row ids to take. with_row_id: bool, default False If True, include the "_rowid" column in the results. Returns ------- LanceTakeQueryBuilder A query builder that can be executed to get the rows. """ @abstractmethod def blob_columns(self) -> list[str]: """Names of blob v1 and v2 columns that can be fetched by row id. Struct-nested blobs use dotted paths. Blobs inside lists are omitted because a row id does not identify an individual list element. """ @abstractmethod def fetch_blobs( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> pa.LargeBinaryArray: """Materialize full blob bytes for ``column`` at the given rows. Local tables support both legacy blob v1 and blob v2 columns. The result has the same length and order as ``row_ids``. Null blobs produce null slots; valid empty blobs produce ``b""``. ``row_ids`` may be a list, an integer Arrow array or chunked array, or a query table with row ids. ``_rowid`` values stay valid after compaction when the table has stable row ids. One missing or deleted row ID rejects the entire batch; no partial results are returned. Blobs inside lists cannot be fetched by row id yet. Remote servers limit each request to 1024 row IDs and 64 MiB of blob bytes. The client splits requests automatically and reads an individual larger blob through the Range route. This method still materializes all bytes in memory; for large values use :meth:`fetch_blob_files`. """ @abstractmethod def fetch_blob_ranges( self, column: str, requests: Sequence[Sequence[int]], ) -> pa.LargeBinaryArray: """Materialize row-specific byte ranges from a blob column. Local tables support both legacy blob v1 and blob v2 columns. Each request is a three-item ``(row_id, offset, length)`` sequence. Requests may be repeated or reordered, including multiple ranges for the same blob. The result has the same length and order as ``requests``; null blobs produce null slots and empty ranges on non-null blobs produce ``b""``. ``_rowid`` values stay valid after compaction when the table has stable row ids. One missing or deleted row ID rejects the entire batch; no partial results are returned. Row IDs can be obtained from a query with ``with_row_id(True)``. Remote tables read the requested ranges through version-pinned blob handles using HTTP byte-range requests. """ @abstractmethod def fetch_blob_files( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> "list[Optional[BlobFile]]": """Open lazy, seekable :class:`~lancedb._blob.BlobFile` handles. Local tables support both legacy blob v1 and blob v2 columns. Prefer this over :meth:`fetch_blobs` for large payloads. ``row_ids`` is a ``list[int]``, an integer Arrow array or chunked array, or a query ``pyarrow.Table`` carrying row identity via ``_rowid`` or a ``_lance_row_id`` field on the blob descriptor. Null blob values return ``None``. Remote tables require LanceDB Cloud server 0.5.0 or newer. ``_rowid`` values stay valid after compaction when the table has stable row ids. One missing or deleted row ID rejects the entire batch; no partial handles are returned. """ @abstractmethod def open_blob_file( self, descriptor: Union[dict, pa.StructScalar, None] ) -> Optional[BlobFile]: """Open a seekable file from a SQL ``blob_descriptor`` result. Accepts the descriptor dictionary or an Arrow struct scalar. Its source pins the table version, branch, column and row identity, so opening does not change this table's checkout. Null descriptors return ``None``. Payload bytes load only when the returned handle is read. An external blob with an unspecified size may require a metadata request on opening. ```python rows = db.execute_query( "SELECT blob_descriptor(video) AS video FROM videos" ).read_all() with videos.open_blob_file(rows["video"][0]) as video: header = video.read(4096) ``` """ def open_blob_files(self, descriptors: Iterable) -> Iterator[Optional[BlobFile]]: """Iterate over lazy files from descriptor dictionaries or Arrow scalars. Accepts an Arrow array/chunked array or any iterable. Opens each handle as iteration advances, preserving input order and nulls as ``None``. Close each non-null handle after use, normally with a ``with`` block. ```python for video in videos.open_blob_files(rows["video"]): if video is not None: with video: header = video.read(4096) ``` """ for descriptor in descriptors: yield self.open_blob_file(descriptor) @abstractmethod def _execute_query( self, query: Query, *, batch_size: Optional[int] = None, timeout: Optional[timedelta] = None, ) -> pa.RecordBatchReader: ... @abstractmethod def _explain_plan(self, query: Query, verbose: Optional[bool] = False) -> str: ... @abstractmethod def _analyze_plan( self, query: Query, *, distributed_metrics: AnalyzePlanDistributedMetrics = "aggregate", ) -> str: ... @abstractmethod def _output_schema(self, query: Query) -> pa.Schema: ... @abstractmethod def _do_merge( self, merge: LanceMergeInsertBuilder, new_data: DATA, on_bad_vectors: OnBadVectorsType, fill_value: float, ) -> MergeResult: ... @abstractmethod def delete(self, where: Union[str, Expr]) -> DeleteResult: """Delete rows from the table. This can be used to delete a single row, many rows, all rows, or sometimes no rows (if your predicate matches nothing). Parameters ---------- where: str or :class:`~lancedb.expr.Expr` The filter condition. Can be a SQL string or a type-safe :class:`~lancedb.expr.Expr` built with :func:`~lancedb.expr.col` and :func:`~lancedb.expr.lit`. The filter must not be empty, or it will error. Returns ------- DeleteResult An object containing the new version number of the table after deletion. Examples -------- >>> import lancedb >>> data = [ ... {"x": 1, "vector": [1.0, 2]}, ... {"x": 2, "vector": [3.0, 4]}, ... {"x": 3, "vector": [5.0, 6]} ... ] >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 2 [3.0, 4.0] 2 3 [5.0, 6.0] >>> table.delete("x = 2") DeleteResult(num_deleted_rows=1, version=2) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 3 [5.0, 6.0] If you have a list of values to delete, you can combine them into a stringified list and use the `IN` operator: >>> to_remove = [1, 5] >>> to_remove = ", ".join([str(v) for v in to_remove]) >>> to_remove '1, 5' >>> table.delete(f"x IN ({to_remove})") DeleteResult(num_deleted_rows=1, version=3) >>> table.to_pandas() x vector 0 3 [5.0, 6.0] """ raise NotImplementedError("delete is not supported for this table type") @abstractmethod def update( self, where: Optional[Union[str, Expr]] = None, values: Optional[dict] = None, *, values_sql: Optional[Dict[str, str]] = None, ) -> UpdateResult: """ This can be used to update zero to all rows depending on how many rows match the where clause. If no where clause is provided, then all rows will be updated. Either `values` or `values_sql` must be provided. You cannot provide both. Parameters ---------- where: str or [Expr][lancedb.expr.Expr], optional The filter condition. Can be a SQL string or a type-safe [Expr][lancedb.expr.Expr] built with [col][lancedb.expr.col] and [lit][lancedb.expr.lit]. The filter must not be empty, or it will error. values: dict, optional The values to update. The keys are the column names and the values are the values to set. values_sql: dict, optional The values to update, expressed as SQL expression strings. These can reference existing columns. For example, {"x": "x + 1"} will increment the x column by 1. Returns ------- UpdateResult - rows_updated: The number of rows that were updated - version: The new version number of the table after the update Examples -------- >>> import lancedb >>> from lancedb.expr import col >>> import pandas as pd >>> data = pd.DataFrame({"x": [1, 2, 3], "vector": [[1.0, 2], [3, 4], [5, 6]]}) >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 2 [3.0, 4.0] 2 3 [5.0, 6.0] >>> table.update(where=col("x") == 2, values={"vector": [10.0, 10]}) UpdateResult(rows_updated=1, version=2) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 3 [5.0, 6.0] 2 2 [10.0, 10.0] >>> table.update(values_sql={"x": "x + 1"}) UpdateResult(rows_updated=3, version=3) >>> table.to_pandas() x vector 0 2 [1.0, 2.0] 1 4 [5.0, 6.0] 2 3 [10.0, 10.0] """ raise NotImplementedError("update is not supported for this table type") @abstractmethod def cleanup_old_versions( self, older_than: Optional[timedelta] = None, *, delete_unverified: bool = False, ) -> "CleanupStats": """ Clean up old versions of the table, freeing disk space. Parameters ---------- older_than: timedelta, default None The minimum age of the version to delete. If None, then this defaults to two weeks. delete_unverified: bool, default False Because they may be part of an in-progress transaction, files newer than 7 days old are not deleted by default. If you are sure that there are no in-progress transactions, then you can set this to True to delete all files older than `older_than`. Returns ------- CleanupStats The stats of the cleanup operation, including how many bytes were freed. See Also -------- [Table.optimize][lancedb.table.Table.optimize]: A more comprehensive optimization operation that includes cleanup as well as other operations. Notes ----- This function is not available in LanceDb Cloud (since LanceDB Cloud manages cleanup for you automatically) """ @abstractmethod def compact_files(self, *args, **kwargs): """ Run the compaction process on the table. This can be run after making several small appends to optimize the table for faster reads. Arguments are passed onto Lance's `lance.dataset.DatasetOptimizer.compact_files`. For most cases, the default should be fine. See Also -------- [Table.optimize][lancedb.table.Table.optimize]: A more comprehensive optimization operation that includes cleanup as well as other operations. Notes ----- This function is not available in LanceDB Cloud (since LanceDB Cloud manages compaction for you automatically) """ @abstractmethod def optimize( self, *, cleanup_older_than: Optional[timedelta] = None, delete_unverified: bool = False, retrain: bool = False, compaction_options: Optional[CompactionOptions] = None, ): """ Optimize the on-disk data and indices for better performance. Modeled after ``VACUUM`` in PostgreSQL. Optimization covers three operations: * Compaction: Merges small files into larger ones * Prune: Removes old versions of the dataset * Index: Optimizes the indices, adding new data to existing indices Parameters ---------- cleanup_older_than: timedelta, optional default 7 days All files belonging to versions older than this will be removed. Set to 0 days to remove all versions except the latest. The latest version is never removed. .. warning:: Setting this to 0 deletes the data files of every older version, so any other reader or writer still using an older version of the table will fail. Only set it to 0 if no other process is working on this dataset. delete_unverified: bool, default False Files leftover from a failed transaction may appear to be part of an in-progress operation (e.g. appending new data) and these files will not be deleted unless they are at least 7 days old. If delete_unverified is True then these files will be deleted regardless of their age. .. warning:: This should only be set to True if you can guarantee that no other process is currently working on this dataset. Otherwise the dataset could be put into a corrupted state. retrain: bool, default False This parameter is no longer used and is deprecated. compaction_options: CompactionOptions, optional Options that control file compaction. For large rows, derive a lower ``target_rows_per_fragment`` from the expected row size to bound later passes. This does not cap the first pass over an existing oversized fragment; see [CompactionOptions][lancedb.table.CompactionOptions]. Notes ----- The frequency an application should call optimize is based on the frequency of data modifications. If data is frequently added, deleted, or updated then optimize should be run frequently. A good rule of thumb is to run optimize if you have added or modified 100,000 or more records or run more than 20 data modification operations. """ @abstractmethod def list_indices(self) -> Iterable[IndexConfig]: """ List all indices that have been created with [Table.create_index][lancedb.table.Table.create_index] """ @abstractmethod def tokenize( self, query: str, *, column: Optional[str] = None, index_name: Optional[str] = None, ) -> Iterable[FtsToken]: """ Tokenize a query using the tokenizer configured on an FTS index. Specify exactly one of ``column`` or ``index_name``. Model-backed tokenizers such as ``jieba/*`` and ``lindera/*`` are rebuilt in the client process from index metadata. For remote tables, this means the same tokenizer model files must also exist locally (under ``LANCE_LANGUAGE_MODEL_HOME``) and must match the files the server uses. A mismatch cannot be detected: the returned tokens may differ from what the server's index produces. If the model files are missing, an error naming the tokenizer is raised. """ @abstractmethod def index_stats(self, index_name: str) -> Optional[IndexStatistics]: """ Retrieve statistics about an index Parameters ---------- index_name: str The name of the index to retrieve statistics for Returns ------- IndexStatistics or None The statistics about the index. Returns None if the index does not exist. """ @abstractmethod def add_columns( self, transforms: Dict[str, str | FunctionApplication] | FunctionApplication | pa.Field | List[pa.Field] | pa.Schema | None = None, *, computed: Dict[str, str] | None = None, auto_refresh_enabled: bool | None = None, auto_refresh_interval_seconds: int | None = None, ): """ Add new columns with defined values. Parameters ---------- transforms: Dict[str, str | FunctionApplication], FunctionApplication, pa.Field, List[pa.Field], pa.Schema A map of column name to a SQL expression to use to calculate the value of the new column. These expressions will be evaluated for each row in the table, and can reference existing columns. Alternatively, a pyarrow Field or Schema can be provided to add new columns with the specified data types. The new columns will be initialized with null values. A mapping with one ``FunctionApplication`` value keeps its scalar or named-struct result in the named table column. A bare named-struct application expands its ordered result fields as one atomic binding; aliases come from ``rename(columns=...)``. Function columns are supported only on LanceDB Cloud and Enterprise. computed: Dict[str, str], optional Local declaration and refresh require a source build with the ``experimental-functions`` Cargo feature, which published packages do not enable. Remote Cloud/Enterprise operations remain available. A mapping from output column names to SQL expressions derives each output field from its expression. A direct projection of a Blob v2 field inherits Blob v2 semantics; other expressions derive their ordinary Arrow type. Mapping order is declaration and dependency order. Unlike ``transforms``, the expression is stored rather than evaluated now: the column is committed with no values, and rows get them from [`refresh_column`][lancedb.table.Table.refresh_column]. Declaring one therefore costs the same on a large table as on an empty one. A refresh also recomputes the rows whose inputs changed since they were computed, so a mutated input is reflected by the next refresh. While a declaration reads a column, that column cannot be renamed, retyped or dropped. On LanceDB Cloud and Enterprise the expression is planned by the server, and the refresh runs as a server job -- see [`refresh_column_async`][lancedb.table.Table.refresh_column_async]. Cannot be combined with ``transforms``. auto_refresh_enabled: bool, optional Enable automatic refresh for each declared computed output. Defaults to True. Local tables store this option; scheduling requires a server. auto_refresh_interval_seconds: int, optional Positive minimum interval after a successful refresh, in seconds. When omitted, the server determines when to refresh. Applies to every computed output in this call. Scheduling options require computed or Function columns. Returns ------- AddColumnsResult version: the new version number of the table after adding columns. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("computed_demo", [{"x": 1}, {"x": 2}]) >>> table.add_columns(computed={"doubled": "x * 2"}) # doctest: +SKIP AddColumnsResult(version=2) >>> table.refresh_column("doubled") # doctest: +SKIP RefreshColumnResult(rows_filled=2, version=4) >>> table.to_arrow().sort_by("x").to_pandas() # doctest: +SKIP x doubled 0 1 2 1 2 4 """ @abstractmethod def refresh_column(self, column: str) -> "RefreshColumnResult": """ Fill the rows of a computed column that hold no value yet. Local refresh requires the ``experimental-functions`` Cargo feature in a source build. Published packages leave it disabled. Declared with ``add_columns(computed=...)``, a column starts empty and gets its values here. Rows appended since the last refresh are filled by the next one, and rows whose inputs changed since they were computed are recomputed; everything else is left as it is. Local tables only: a remote refresh runs as a server job, through [`refresh_column_async`][lancedb.table.Table.refresh_column_async]. Parameters ---------- column: str The name of the computed column to fill. Returns ------- RefreshColumnResult rows_filled: the number of rows given a value. version: the new version number of the table. """ @abstractmethod def refresh_column_async(self, column: str) -> Job[RefreshColumnJobResult]: """ Like :meth:`refresh_column`, but returns a handle to the refresh job instead of blocking until it completes. Local refresh requires a source build with ``experimental-functions``. Remote Cloud/Enterprise jobs do not require this feature. The job may already be complete when returned; callers must not assume the column is filled until :meth:`Job.wait` returns. Invalid input -- an unknown column, or one that is not computed -- raises here rather than failing the job. On local tables the job runs in-process; on LanceDB Cloud and Enterprise it is the server's backfill job. Returns ------- Job[RefreshColumnResult] A job whose successful ``wait`` returns row counts plus the source and published table versions. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("computed_job_demo", [{"x": 1}, {"x": 2}]) >>> table.add_columns(computed={"doubled": "x * 2"}) # doctest: +SKIP AddColumnsResult(version=2) >>> job = table.refresh_column_async("doubled") # doctest: +SKIP >>> result = job.wait() # doctest: +SKIP >>> result.rows_assigned # doctest: +SKIP 2 >>> job.status() # doctest: +SKIP 'finished' """ @abstractmethod def function_errors( self, job_id: Optional[str] = None, column: Optional[str] = None, limit: Optional[int] = None, ) -> "FunctionErrors": """ The per-row errors Function refreshes recorded on this table. A refresh running under a skip policy records each row it skipped with the input that failed and the error. This lists those records, newest job first, plus a summary for any fragment whose per-row detail was capped. LanceDB Cloud and Enterprise only; reading errors needs read access to the table, since a message carries the value that failed. Parameters ---------- job_id: str, optional Only errors recorded by this job. column: str, optional Only errors on this column. limit: int, optional At most this many records (server default 10000, cap 100000). Returns ------- FunctionErrors ``records``, ``fragments`` and ``truncated``, the last saying whether the listing stopped at its limit. Examples -------- >>> errors = table.function_errors(column="embedding") # doctest: +SKIP >>> for record in errors.records: # doctest: +SKIP ... print(record.job_id, record.row_offset, record.error_message) >>> if errors.truncated: # doctest: +SKIP ... print("listing stopped at the limit") """ @abstractmethod def alter_columns(self, *alterations: Iterable[Dict[str, str]]): """ Alter column names and nullability. Parameters ---------- alterations : Iterable[Dict[str, Any]] A sequence of dictionaries, each with the following keys: - "path": str The column path to alter. For a top-level column, this is the name. For a nested column, this is the dot-separated path, e.g. "a.b.c". - "rename": str, optional The new name of the column. If not specified, the column name is not changed. - "data_type": pyarrow.DataType, optional The new data type of the column. Existing values will be casted to this type. If not specified, the column data type is not changed. - "nullable": bool, optional Whether the column should be nullable. If not specified, the column nullability is not changed. Only non-nullable columns can be changed to nullable. Currently, you cannot change a nullable column to non-nullable. Each alteration must specify at least one of "rename", "data_type", or "nullable". The legacy key "name" is also accepted for renaming. Unknown keys raise ValueError before any alterations are applied. Returns ------- AlterColumnsResult version: the new version number of the table after the alteration. """ @abstractmethod def update_field_metadata( self, *updates: dict[str, Any] ) -> UpdateFieldMetadataResult: """ Update per-field (column) metadata. Parameters ---------- updates : dict One or more dicts, each with: - "path": str — dot-path to the field (e.g. "embedding" or "a.b.c"). - "metadata": dict[str, str | None] — keys to set; a value of ``None`` deletes that key. - "replace": bool, optional — replace the field's whole metadata map instead of merging (default False). The following keys are treated specially, by convention, and should be used when appropriate: - "lancedb:description": for a human-readable description of a field. - ``"lancedb:tag:"`` for a user-defined key-value tag, where the suffix names the tag category; e.g. "lancedb:tag:model": "clip". - "lancedb:status" for status options ("production", "candidate", "deprecated", "archived") to designate the current life cycle state of this column. Returns ------- UpdateFieldMetadataResult version: the new table version after the update. """ @deprecation.deprecated( deprecated_in="0.33.1", current_version=__version__, details="Use update_field_metadata() instead.", ) def replace_field_metadata(self, field_name: str, new_metadata: Dict[str, str]): """Replace all metadata of a field in the schema. .. deprecated:: 0.33.1 Use [update_field_metadata][lancedb.table.Table.update_field_metadata] with ``replace=True`` instead. Parameters ---------- field_name: str The name of the field whose metadata is replaced. new_metadata: dict The new metadata, replacing all existing keys. """ self.update_field_metadata( {"path": field_name, "metadata": new_metadata, "replace": True} ) @abstractmethod def drop_columns(self, columns: Iterable[str]) -> DropColumnsResult: """ Drop columns from the table. Parameters ---------- columns : Iterable[str] The names of the columns to drop. Returns ------- DropColumnsResult version: the new version number of the table dropping the columns. """ @abstractmethod def checkout(self, version: Union[int, str]): """ Checks out a specific version of the Table Any read operation on the table will now access the data at the checked out version. As a consequence, calling this method will disable any read consistency interval that was previously set. This is a read-only operation that turns the table into a sort of "view" or "detached head". Other table instances will not be affected. To make the change permanent you can use the `[Self::restore]` method. Any operation that modifies the table will fail while the table is in a checked out state. To return the table to a normal state use `[Self::checkout_latest]`. Parameters ---------- version: int | str, The version to check out. A version number (`int`) or a tag (`str`) can be provided. """ @abstractmethod def checkout_latest(self): """ Ensures the table is pointing at the latest version This can be used to manually update a table when the read_consistency_interval is None It can also be used to undo a `[Self::checkout]` operation """ @abstractmethod def restore(self, version: Optional[Union[int, str]] = None): """Restore a version of the table. This is an in-place operation. This creates a new version where the data is equivalent to the specified previous version. Data is not copied (as of python-v0.2.1). Parameters ---------- version : int or str, default None The version number or version tag to restore. If unspecified then restores the currently checked out version. If the currently checked out version is the latest version then this is a no-op. """ @abstractmethod def list_versions(self) -> List[Dict[str, Any]]: """List all versions of the table""" @cached_property def _dataset_uri(self) -> str: return _table_uri(self._conn.uri, self.name) def _get_fts_index_path(self) -> Tuple[str, pa_fs.FileSystem, bool]: from .remote.table import RemoteTable if isinstance(self, RemoteTable) or get_uri_scheme(self._dataset_uri) != "file": return ("", None, False) path = join_uri(self._dataset_uri, "_indices", "fts") fs, path = fs_from_uri(path) index_exists = fs.get_file_info(path).type != pa_fs.FileType.NotFound return (path, fs, index_exists) def _ensure_no_legacy_fts_index(self): path, _, exists = self._get_fts_index_path() if exists: raise ValueError( "Legacy Tantivy FTS index detected at " f"{path}. Tantivy-based FTS has been removed. " "Delete the legacy index and recreate it with " "table.create_fts_index(...)." ) @abstractmethod def uses_v2_manifest_paths(self) -> bool: """ Check if the table is using the new v2 manifest paths. Returns ------- bool True if the table is using the new v2 manifest paths, False otherwise. """ @abstractmethod def migrate_v2_manifest_paths(self): """ Migrate the manifest paths to the new format. This will update the manifest to use the new v2 format for paths. This function is idempotent, and can be run multiple times without changing the state of the object store. !!! danger This should not be run while other concurrent operations are happening. And it should also run until completion before resuming other operations. You can use [Table.uses_v2_manifest_paths][lancedb.table.Table.uses_v2_manifest_paths] to check if the table is already using the new path style. """ # WAL-PK-FUSION: delete both hooks, here and on AsyncTable and RemoteTable. def _hybrid_pk_fusion_learned(self) -> bool: """Whether a hybrid query here has already been refused ``_rowid``. Learned from a refusal, never probed, so asking is free. ``False`` for table types that never refuse. """ return False def _note_hybrid_pk_fusion(self) -> None: """Remember a ``_rowid`` refusal, so later hybrid queries skip it.""" class LanceTable(Table): """ A table in a LanceDB database. This can be opened in two modes: standard and time-travel. Standard mode is the default. In this mode, the table is mutable and tracks the latest version of the table. The level of read consistency is controlled by the `read_consistency_interval` parameter on the connection. Time-travel mode is activated by specifying a version number. In this mode, the table is immutable and fixed to a specific version. This is useful for querying historical versions of the table. """ def __init__( self, connection: "LanceDBConnection", name: str, *, namespace_path: Optional[List[str]] = None, storage_options: Optional[Dict[str, str]] = None, index_cache_size: Optional[int] = None, location: Optional[str] = None, namespace_client: Optional[Any] = None, managed_versioning: Optional[bool] = None, pushdown_operations: Optional[set] = None, route_pushdown_to_rust: bool = False, _async: AsyncTable = None, ): if namespace_path is None: namespace_path = [] self._conn = connection self._namespace_path = namespace_path self._location = location # Store location for use in _dataset_path self._namespace_client = namespace_client self._pushdown_operations = pushdown_operations or set() # When the connection built the namespace client natively (e.g. an # enterprise "rest" connection), the underlying Rust table already # executes QueryTable pushdown itself -- and, unlike this Python urllib3 # path, it routes through the read-freshness context provider that emits # the ``x-lancedb-min-timestamp`` header. So we must defer pushdown to # Rust instead of calling the Python ``namespace_client.query_table`` # directly, or reads silently bypass read-freshness (stale results). self._route_pushdown_to_rust = route_pushdown_to_rust if _async is not None: self._table = _async else: self._table = LOOP.run( connection._conn.open_table( name, namespace_path=namespace_path, storage_options=storage_options, index_cache_size=index_cache_size, location=location, namespace_client=namespace_client, managed_versioning=managed_versioning, ) ) # Take queries read the pushdown set from the inner table. self._table._pushdown_operations = self._pushdown_operations @property def name(self) -> str: return self._table.name @property def namespace(self) -> List[str]: """Return the namespace path of the table.""" return self._namespace_path @classmethod async def from_inner(cls, tbl: LanceDBTable): from .db import AsyncConnection, LanceDBConnection async_tbl = AsyncTable(tbl) inner_conn = tbl.database() read_consistency_interval = await AsyncConnection( inner_conn ).get_read_consistency_interval() conn = LanceDBConnection.from_inner(inner_conn, read_consistency_interval) return cls( conn, async_tbl.name, _async=async_tbl, ) @classmethod def open( cls, db, name, *, namespace_path: Optional[List[str]] = None, storage_options: Optional[Dict[str, str]] = None, index_cache_size: Optional[int] = None, location: Optional[str] = None, namespace_client: Optional[Any] = None, managed_versioning: Optional[bool] = None, pushdown_operations: Optional[set] = None, ): if namespace_path is None: namespace_path = [] tbl = cls( db, name, namespace_path=namespace_path, storage_options=storage_options, index_cache_size=index_cache_size, location=location, namespace_client=namespace_client, managed_versioning=managed_versioning, pushdown_operations=pushdown_operations, ) # check the dataset exists try: tbl.version except ValueError as e: if "Not found:" in str(e): raise FileNotFoundError(f"Table {name} does not exist") raise e return tbl @cached_property def _dataset_path(self) -> str: # Cacheable since it's deterministic # If table was opened with explicit location (e.g., from namespace), # use that location directly instead of constructing from base URI if self._location is not None: return self._location return _table_path(self._conn.uri, self.name) def to_lance(self, **kwargs) -> lance.LanceDataset: """Return the LanceDataset backing this table.""" try: import lance except ImportError: raise ImportError( "The lance library is required to use this function. " "Please install with `pip install pylance`." ) branch = self.current_branch() version = None if branch is not None else self.version namespace_client = self._namespace_client if namespace_client is None: conn_uri = getattr(self._conn, "uri", "") if get_uri_scheme(conn_uri) == "namespace" or self._namespace_path: namespace_client = self._conn.namespace_client() self._namespace_client = namespace_client if namespace_client is not None: table_id = self._namespace_path + [self.name] ds = lance.dataset( version=version, storage_options=self._conn.storage_options, namespace_client=namespace_client, table_id=table_id, **kwargs, ) else: ds = lance.dataset( self._dataset_path, version=version, storage_options=self._conn.storage_options, **kwargs, ) if branch is not None: ds = ds.checkout_version((branch, self.version)) return ds @property def schema(self) -> pa.Schema: """Return the schema of the table. Returns ------- pa.Schema A PyArrow schema object.""" return LOOP.run(self._table.schema()) def list_versions(self) -> List[Dict[str, Any]]: """List all versions of the table""" return LOOP.run(self._table.list_versions()) @property def version(self) -> int: """Get the current version of the table""" return LOOP.run(self._table.version()) def take_offsets( self, offsets: list[int], *, with_row_id: bool = False ) -> LanceTakeQueryBuilder: builder = LanceTakeQueryBuilder(self._table.take_offsets(offsets)) if with_row_id: builder.with_row_id() return builder def take_row_ids( self, row_ids: list[int], *, with_row_id: bool = False ) -> LanceTakeQueryBuilder: builder = LanceTakeQueryBuilder(self._table.take_row_ids(row_ids)) if with_row_id: builder.with_row_id() return builder def blob_columns(self) -> list[str]: return LOOP.run(self._table.blob_columns()) def fetch_blobs( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> pa.LargeBinaryArray: return LOOP.run(self._table.fetch_blobs(column, row_ids)) def fetch_blob_ranges( self, column: str, requests: Sequence[Sequence[int]], ) -> pa.LargeBinaryArray: return LOOP.run(self._table.fetch_blob_ranges(column, list(requests))) def fetch_blob_files( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> "list[Optional[BlobFile]]": return LOOP.run(self._table.fetch_blob_files(column, row_ids)) def open_blob_file( self, descriptor: Union[dict, pa.StructScalar, None] ) -> Optional[BlobFile]: return LOOP.run(self._table.open_blob_file(descriptor)) @property def tags(self) -> Tags: """Tag management for the table. Similar to Git, tags are a way to add metadata to a specific version of the table. .. warning:: Tagged versions are exempted from the :py:meth:`cleanup_old_versions()` process. To remove a version that has been tagged, you must first :py:meth:`~Tags.delete` the associated tag. Returns ------- Tags The tag manager for managing tags for the table. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", ... [{"vector": [1.1, 0.9], "type": "vector"}]) >>> table.tags.create("v1", table.version) >>> table.add([{"vector": [0.5, 0.2], "type": "vector"}]) AddResult(version=2) >>> tags = table.tags.list() >>> print(tags["v1"]["version"]) 1 >>> table.checkout("v1") >>> table.to_pandas() vector type 0 [1.1, 0.9] vector """ return Tags(self._table) @property def branches(self) -> "Branches": """Branch management for the table. ``create``/``checkout`` return a new table handle scoped to the branch; writes on it do not affect ``main``. """ return Branches(self) def current_branch(self) -> Optional[str]: """The branch this table handle is scoped to, or ``None`` for ``main``.""" return self._table.current_branch() def _wrap_branch_handle( self, async_table: "AsyncTable", version: Optional[int] = None ) -> "LanceTable": # version is unused locally: the pin already lives on async_table and a # local handle is not reopened via a serialized connection. return LanceTable( self._conn, async_table.name, namespace_path=self._namespace_path, namespace_client=self._namespace_client, pushdown_operations=self._pushdown_operations, route_pushdown_to_rust=self._route_pushdown_to_rust, location=self._location, _async=async_table, ) def checkout(self, version: Union[int, str]): """Checkout a version of the table. This is an in-place operation. This allows viewing previous versions of the table. If you wish to keep writing to the dataset starting from an old version, then use the `restore` function. Calling this method will set the table into time-travel mode. If you wish to return to standard mode, call `checkout_latest`. Parameters ---------- version: int | str, The version to check out. A version number (`int`) or a tag (`str`) can be provided. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", ... [{"vector": [1.1, 0.9], "type": "vector"}]) >>> table.version 1 >>> table.to_pandas() vector type 0 [1.1, 0.9] vector >>> table.add([{"vector": [0.5, 0.2], "type": "vector"}]) AddResult(version=2) >>> table.version 2 >>> table.checkout(1) >>> table.to_pandas() vector type 0 [1.1, 0.9] vector """ LOOP.run(self._table.checkout(version)) def checkout_latest(self): """Checkout the latest version of the table. This is an in-place operation. The table will be set back into standard mode, and will track the latest version of the table. """ LOOP.run(self._table.checkout_latest()) def restore(self, version: Optional[Union[int, str]] = None): """Restore a version of the table. This is an in-place operation. This creates a new version where the data is equivalent to the specified previous version. Data is not copied (as of python-v0.2.1). Parameters ---------- version : int or str, default None The version number or version tag to restore. If unspecified then restores the currently checked out version. If the currently checked out version is the latest version then this is a no-op. Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", [ ... {"vector": [1.1, 0.9], "type": "vector"}]) >>> table.version 1 >>> table.to_pandas() vector type 0 [1.1, 0.9] vector >>> table.add([{"vector": [0.5, 0.2], "type": "vector"}]) AddResult(version=2) >>> table.version 2 >>> table.tags.create("v2", 2) >>> table.restore(1) >>> table.to_pandas() vector type 0 [1.1, 0.9] vector >>> len(table.list_versions()) 3 >>> table.restore("v2") >>> table.to_pandas() vector type 0 [1.1, 0.9] vector 1 [0.5, 0.2] vector >>> len(table.list_versions()) 4 """ if version is not None: LOOP.run(self._table.checkout(version)) LOOP.run(self._table.restore()) def count_rows(self, filter: Optional[str] = None) -> int: return LOOP.run(self._table.count_rows(filter)) def __repr__(self) -> str: return f"{self.__class__.__name__}(name={self.name!r}, _conn={self._conn!r})" def __str__(self) -> str: return self.__repr__() def head(self, n=5) -> pa.Table: """Return the first n rows of the table.""" return LOOP.run(self._table.head(n)) def to_pandas(self, blob_mode: BlobMode = "lazy", **kwargs) -> "pd.DataFrame": """Return the table as a pandas DataFrame. Parameters ---------- blob_mode: str, default "lazy" Controls how Lance blob columns are returned. **kwargs Forwarded to Lance pandas conversion. Returns ------- pd.DataFrame """ validate_blob_mode(blob_mode) if blob_mode == "descriptions" or not schema_has_blob_field(self.schema): arrow_tbl = self.to_arrow() if blob_mode == "descriptions": arrow_tbl = strip_auto_row_ids( arrow_tbl, row_addressable_blob_v2_paths(self.schema) ) return arrow_tbl.to_pandas(**kwargs) if ( blob_mode == "lazy" and self._namespace_client is None and get_uri_scheme(self._dataset_path) == "memory" ): return self.to_arrow().to_pandas(**kwargs) if blob_mode == "bytes" and blob_v2_column_paths(self.schema): return self.search().to_pandas(blob_mode=blob_mode, **kwargs) return self.to_lance().to_pandas(blob_mode=blob_mode, **kwargs) def to_arrow(self, *, blob_mode: BlobMode = "lazy") -> pa.Table: """Return the table as a pyarrow Table. Parameters ---------- blob_mode: str, default "lazy" ``"lazy"`` returns blob v2 descriptors that carry the row id ``fetch_blobs`` reads. ``"descriptions"`` returns them without it. ``"bytes"`` returns payload bytes and raises for legacy blob columns and blobs inside lists. Returns ------- pa.Table""" validate_blob_mode(blob_mode) if blob_mode != "lazy" and schema_has_blob_field(self.schema): return self.search().to_arrow(blob_mode=blob_mode) if ( _should_push_down_query_table( self._namespace_client, self._pushdown_operations ) and not self._route_pushdown_to_rust ): return self._execute_query(Query()).read_all() return LOOP.run(self._table.to_arrow()) def to_polars(self, batch_size=None) -> "pl.LazyFrame": """Return the table as a polars LazyFrame. Parameters ---------- batch_size: int, optional Passed to polars. This is the maximum row count for scanned pyarrow record batches Note ---- 1. This requires polars to be installed separately 2. Currently we've disabled push-down of the filters from polars because polars pushdown into pyarrow uses pyarrow compute expressions rather than SQl strings (which LanceDB supports) 3. The Polars streaming engine is not supported because it does not currently implement Python PyArrow dataset scans. Use the default engine when collecting this LazyFrame. Returns ------- pl.LazyFrame """ from lancedb.integrations.pyarrow import PyarrowDatasetAdapter dataset = PyarrowDatasetAdapter(self) # Polars 1.32's non-PyArrow callback path passes batch_size twice. Keep # the compatible PyArrow path, but block predicates because this adapter # cannot translate PyArrow expressions into LanceDB filters. return pl.scan_pyarrow_dataset(dataset, batch_size=batch_size).map_batches( _polars_predicate_pushdown_barrier, predicate_pushdown=False, ) # New unified API overload @overload def create_index( self, column: str, /, *, config: IndexConfigType, replace: bool = ..., wait_timeout: Optional[timedelta] = ..., name: Optional[str] = ..., train: bool = ..., ) -> None: ... # Legacy API overload (deprecated) @overload def create_index( self, metric: Literal["l2", "cosine", "dot", "hamming"] = ..., num_partitions: Optional[int] = ..., num_sub_vectors: Optional[int] = ..., vector_column_name: str = ..., replace: bool = ..., accelerator: Optional[str] = ..., index_cache_size: Optional[int] = ..., num_bits: int = ..., index_type: Literal[ "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_RQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ" ] = ..., max_iterations: int = ..., sample_rate: int = ..., m: int = ..., ef_construction: int = ..., *, wait_timeout: Optional[timedelta] = ..., name: Optional[str] = ..., train: bool = ..., target_partition_size: Optional[int] = ..., ) -> None: ... def create_index( self, metric: str = "l2", num_partitions: Optional[int] = None, num_sub_vectors: Optional[int] = None, vector_column_name: str = VECTOR_COLUMN_NAME, replace: bool = True, accelerator: Optional[str] = None, index_cache_size: Optional[int] = None, num_bits: int = 8, index_type: Literal[ "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_RQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ", "IVF_HNSW_FLAT", ] = "IVF_PQ", max_iterations: int = 50, sample_rate: int = 256, m: int = 20, ef_construction: int = 300, *, config: Optional[IndexConfigType] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, target_partition_size: Optional[int] = None, ): """Create an index on a column. This method supports both the new unified API and the legacy API for backwards compatibility. The new API takes the column name as the first positional argument and an index configuration object via ``config``; the legacy API takes the distance metric as the first argument plus separate ``vector_column_name`` / ``num_partitions`` / etc. parameters, and emits a ``DeprecationWarning``. Parameters ---------- metric : str For new API: the column name to index. For legacy API: the distance metric ("l2", "cosine", "dot", "hamming"). config : IndexConfigType, optional The index configuration object. If provided, uses the new unified API. Can be one of: IvfFlat, IvfPq, IvfSq, IvfRq, HnswPq, HnswSq, BTree, Bitmap, LabelList, Fm, FTS. replace : bool, default True Whether to replace an existing index on this column. wait_timeout : timedelta, optional Timeout to wait for async indexing to complete. name : str, optional Custom name for the index. train : bool, default True Whether to train the index with existing data. Examples -------- New API (recommended): >>> table.create_index( # doctest: +SKIP ... "vector", config=IvfPq(distance_type="l2") ... ) >>> table.create_index("category", config=BTree()) # doctest: +SKIP >>> table.create_index("content", config=FTS()) # doctest: +SKIP Legacy API (deprecated): >>> table.create_index( # doctest: +SKIP ... "l2", vector_column_name="vector" ... ) """ # Detect whether this is a legacy API call is_legacy = self._is_legacy_create_index_call( metric, config, num_partitions, num_sub_vectors, vector_column_name, accelerator, index_cache_size, ) if is_legacy: warnings.warn( "The create_index() API with metric/num_partitions parameters is " "deprecated and will be removed in a future version. " "Please migrate to the new unified API:\n" " # Old (deprecated):\n" " table.create_index('l2', vector_column_name='my_vector')\n" " # New (recommended):\n" " table.create_index('my_vector', config=IvfPq(distance_type='l2'))", DeprecationWarning, stacklevel=2, ) # Legacy API: first arg is the distance metric column = vector_column_name # Build config from legacy parameters config = self._build_vector_config_from_legacy_params( metric=metric, index_type=index_type, num_partitions=num_partitions, num_sub_vectors=num_sub_vectors, num_bits=num_bits, max_iterations=max_iterations, sample_rate=sample_rate, m=m, ef_construction=ef_construction, target_partition_size=target_partition_size, accelerator=accelerator, ) # Handle accelerator through pylance if accelerator is not None: dataset = self.to_lance() _validate_index_accelerator(accelerator) dataset.create_index( column=column, index_type=index_type, metric=metric, num_partitions=num_partitions, num_sub_vectors=num_sub_vectors, replace=replace, accelerator=accelerator, index_cache_size=index_cache_size, num_bits=num_bits, m=m, ef_construction=ef_construction, target_partition_size=target_partition_size, ) self.checkout_latest() return else: # New API: metric is the column name column = metric # Check if config has accelerator set and dispatch to pylance if config is not None and hasattr(config, "accelerator"): acc = getattr(config, "accelerator", None) if acc is not None: # Dispatch to pylance for GPU acceleration dataset = self.to_lance() _validate_index_accelerator(acc) index_type_map = { "IvfFlat": "IVF_FLAT", "IvfSq": "IVF_SQ", "IvfPq": "IVF_PQ", "IvfRq": "IVF_RQ", "HnswPq": "IVF_HNSW_PQ", "HnswSq": "IVF_HNSW_SQ", } cfg_type = type(config).__name__ lance_index_type = index_type_map.get(cfg_type, "IVF_PQ") quantization_kwargs = {} num_bits = getattr(config, "num_bits", 8) if num_bits is not None: quantization_kwargs["num_bits"] = num_bits dataset.create_index( column=column, index_type=lance_index_type, metric=getattr(config, "distance_type", "l2"), num_partitions=getattr(config, "num_partitions", None), num_sub_vectors=getattr(config, "num_sub_vectors", None), replace=replace, accelerator=acc, m=getattr(config, "m", 20), ef_construction=getattr(config, "ef_construction", 300), target_partition_size=getattr( config, "target_partition_size", None ), **quantization_kwargs, ) self.checkout_latest() return return LOOP.run( self._table.create_index( column, replace=replace, config=config, wait_timeout=wait_timeout, name=name, train=train, ) ) def create_index_async( self, column: str, *, config: IndexConfigType, replace: Optional[bool] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, ) -> Job: """Create an index, returning a handle to the indexing job. The job may already be complete when returned; callers must not assume the index exists until :meth:`Job.wait` returns. """ return Job( LOOP.run( self._table.create_index_async( column, replace=replace, config=config, wait_timeout=wait_timeout, name=name, train=train, ) ) ) def drop_index(self, name: str) -> None: """ Drops an index from the table Parameters ---------- name: str The name of the index to drop """ return LOOP.run(self._table.drop_index(name)) def prewarm_index(self, name: str) -> None: """ Prewarm an index in the table. This is a hint to the database that the index will be accessed in the future and should be loaded into memory if possible. This can reduce cold-start latency for subsequent queries. This call initiates prewarming and returns once the request is accepted. It is idempotent and safe to call from multiple clients concurrently. It is generally wasteful to call this if the index does not fit into the available cache. Not all index types support prewarming; unsupported indices will silently ignore the request. Parameters ---------- name: str The name of the index to prewarm """ return LOOP.run(self._table.prewarm_index(name)) def prewarm_data(self, columns: Optional[List[str]] = None) -> None: """ Prewarm data for the table. This is a hint to the database that the given columns will be accessed in the future and the database should prefetch the data if possible. Currently only supported on remote tables. This call initiates prewarming and returns once the request is accepted. It is idempotent and safe to call from multiple clients concurrently. This operation has a large upfront cost but can speed up future queries that need to fetch the given columns. Large columns such as embeddings or binary data may not be practical to prewarm. This feature is intended for workloads that issue many queries against the same columns. Parameters ---------- columns: list of str, optional The columns to prewarm. If None, all columns are prewarmed. """ return LOOP.run(self._table.prewarm_data(columns)) def wait_for_index( self, index_names: Iterable[str], timeout: timedelta = timedelta(seconds=300) ) -> None: return LOOP.run(self._table.wait_for_index(index_names, timeout)) def stats(self) -> TableStatistics: return LOOP.run(self._table.stats()) @property def uri(self) -> str: return LOOP.run(self._table.uri()) def initial_storage_options(self) -> Optional[Dict[str, str]]: """Get the initial storage options that were passed in when opening this table. For dynamically refreshed options (e.g., credential vending), use :meth:`latest_storage_options`. Warning: This is an internal API and the return value is subject to change. Returns ------- Optional[Dict[str, str]] The storage options, or None if no storage options were configured. """ return LOOP.run(self._table.initial_storage_options()) def latest_storage_options(self) -> Optional[Dict[str, str]]: """Get the latest storage options, refreshing from provider if configured. This method is useful for credential vending scenarios where storage options may be refreshed dynamically. If no dynamic provider is configured, this returns the initial static options. Warning: This is an internal API and the return value is subject to change. Returns ------- Optional[Dict[str, str]] The storage options, or None if no storage options were configured. """ return LOOP.run(self._table.latest_storage_options()) @deprecation.deprecated( deprecated_in="0.25.0", current_version=__version__, details="Use create_index() with config=BTree()/Bitmap()/LabelList() instead.", ) def create_scalar_index( self, column: str, *, replace: bool = True, index_type: ScalarIndexType = "BTREE", wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, ): """Create a scalar index on a column. .. deprecated:: 0.25.0 Use :meth:`create_index` with a BTree, Bitmap, or LabelList config instead. Example: ``table.create_index("column", config=BTree())`` """ if index_type == "scalar": warnings.warn( 'index_type="scalar" is deprecated; use "BTREE" instead', DeprecationWarning, stacklevel=2, ) index_type = "BTREE" if index_type == "BTREE": config = BTree() elif index_type == "BITMAP": config = Bitmap() elif index_type == "LABEL_LIST": config = LabelList() else: raise ValueError(f"Unknown index type {index_type}") return LOOP.run( self._table.create_index( column, replace=replace, config=config, wait_timeout=wait_timeout, name=name, ) ) @deprecation.deprecated( deprecated_in="0.25.0", current_version=__version__, details="Use create_index() with config=FTS() instead.", ) def create_fts_index( self, field_names: Union[str, List[str]], *, ordering_field_names: Optional[Union[str, List[str]]] = None, replace: bool = False, writer_heap_size: Optional[int] = 1024 * 1024 * 1024, use_tantivy: bool = False, tokenizer_name: Optional[str] = None, with_position: bool = False, # tokenizer configs: base_tokenizer: BaseTokenizerType = "simple", language: str = "English", max_token_length: Optional[int] = 40, lower_case: bool = True, stem: Optional[bool] = None, remove_stop_words: Optional[bool] = None, custom_stop_words: Optional[List[str]] = None, ascii_folding: bool = True, ngram_min_length: int = 3, ngram_max_length: int = 3, prefix_only: bool = False, block_size: int = 128, document_granularity: DocumentGranularity = DocumentGranularity.ROW, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, ): """Create a full-text search index on a column. .. deprecated:: 0.25.0 Use :meth:`create_index` with an FTS config instead. Example: ``table.create_index("text_column", config=FTS())`` """ self._ensure_no_legacy_fts_index() if use_tantivy: raise ValueError( "Tantivy-based FTS has been removed. " "Remove use_tantivy and recreate the index with native FTS." ) if ordering_field_names is not None: raise ValueError( "ordering_field_names was only supported by the removed " "Tantivy-based FTS implementation." ) if writer_heap_size != 1024 * 1024 * 1024: raise ValueError( "writer_heap_size was only supported by the removed " "Tantivy-based FTS implementation." ) if not isinstance(field_names, str): raise ValueError( "Native FTS indexes can only be created on a single field " "at a time. To search over multiple text fields, create a " "separate FTS index for each field." ) if tokenizer_name is None: tokenizer_configs = { "base_tokenizer": base_tokenizer, "language": language, "with_position": with_position, "max_token_length": max_token_length, "lower_case": lower_case, "stem": stem, "remove_stop_words": remove_stop_words, "custom_stop_words": custom_stop_words, "ascii_folding": ascii_folding, "ngram_min_length": ngram_min_length, "ngram_max_length": ngram_max_length, "prefix_only": prefix_only, } else: tokenizer_configs = self.infer_tokenizer_configs(tokenizer_name) tokenizer_configs["custom_stop_words"] = custom_stop_words config = FTS( block_size=block_size, document_granularity=document_granularity, **tokenizer_configs, ) try: LOOP.run( self._table.create_index( field_names, replace=replace, config=config, name=name, wait_timeout=wait_timeout, ) ) except (ValueError, RuntimeError) as e: _maybe_add_fts_error_note( e, base_tokenizer=config.base_tokenizer, language=config.language, ) raise e def add( self, data: DATA, mode: AddMode = "append", on_bad_vectors: OnBadVectorsType = "error", fill_value: float = 0.0, progress: Optional[Union[bool, Callable, Any]] = None, write_parallelism: Optional[int] = None, allow_external_blob_outside_bases: bool = False, ) -> AddResult: """Add data to the table. If vector columns are missing and the table has embedding functions, then the vector columns are automatically computed and added. Parameters ---------- data: list-of-dict, pd.DataFrame The data to insert into the table. mode: str The mode to use when writing the data. Valid values are "append" and "overwrite". Overwrite adopts explicitly declared blob layouts, while inferred bytes retain the existing blob v2 fields. on_bad_vectors: str, default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. One of "error", "drop", "fill", "null". fill_value: float, default 0. The value to use when filling vectors. Only used if on_bad_vectors="fill". progress: bool, callable, or tqdm-like, optional A callback or tqdm-compatible progress bar. See :meth:`Table.add` for details. write_parallelism: int, optional Number of partitions to write in parallel. Higher values increase throughput but also peak memory use, since each partition buffers data in flight. Defaults to an estimate based on the data size, capped at the number of CPU cores. Lower this if bulk ingestion is using too much memory. allow_external_blob_outside_bases: bool, default False Allow blob URIs outside registered bases. See :meth:`Table.add`. Local tables only. Returns ------- int The number of vectors in the table. """ progress, owns = _normalize_progress(progress) try: return LOOP.run( self._table.add( data, mode=mode, on_bad_vectors=on_bad_vectors, fill_value=fill_value, progress=progress, write_parallelism=write_parallelism, allow_external_blob_outside_bases=allow_external_blob_outside_bases, ) ) finally: if owns: progress.close() def merge( self, other_table: Union[LanceTable, DATA], left_on: str, right_on: Optional[str] = None, schema: Optional[Union[pa.Schema, LanceModel]] = None, ): """Merge another table into this table. Performs a left join, where the dataset is the left side and other_table is the right side. Rows existing in the dataset but not on the left will be filled with null values, unless Lance doesn't support null values for some types, in which case an error will be raised. The only overlapping column allowed is the join column. If other overlapping columns exist, an error will be raised. Parameters ---------- other_table: LanceTable or Reader-like The data to be merged. Acceptable types are: - Pandas DataFrame, Pyarrow Table, Dataset, Scanner, Iterator[RecordBatch], or RecordBatchReader - LanceTable left_on: str The name of the column in the dataset to join on. right_on: str or None The name of the column in other_table to join on. If None, defaults to left_on. schema: pa.Schema or LanceModel, optional The schema of the other_table. If not provided, the schema is inferred from the data. Examples -------- >>> import lancedb >>> import pyarrow as pa >>> df = pa.table({'x': [1, 2, 3], 'y': ['a', 'b', 'c']}) >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("dataset", df) >>> table.to_pandas() x y 0 1 a 1 2 b 2 3 c >>> new_df = pa.table({'x': [1, 2, 3], 'z': ['d', 'e', 'f']}) >>> table.merge(new_df, 'x') >>> table.to_pandas() x y z 0 1 a d 1 2 b e 2 3 c f """ if isinstance(other_table, LanceTable): other_table = other_table.to_lance() else: other_table = _sanitize_data( other_table, schema, ) self.to_lance().merge( other_table, left_on=left_on, right_on=right_on, schema=schema ) self.checkout_latest() @cached_property def embedding_functions(self) -> Dict[str, EmbeddingFunctionConfig]: """ Get the embedding functions for the table Returns ------- funcs: Dict[str, EmbeddingFunctionConfig] A mapping of the vector column to the embedding function or empty dict if not configured. """ return EmbeddingFunctionRegistry.get_instance().parse_functions( self.schema.metadata ) @overload def search( # type: ignore self, query: Optional[Union[VEC, str, "PIL.Image.Image", Tuple]] = None, vector_column_name: Optional[str] = None, query_type: Literal["vector"] = "vector", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceVectorQueryBuilder: ... @overload def search( self, query: Optional[Union[VEC, str, "PIL.Image.Image", Tuple]] = None, vector_column_name: Optional[str] = None, query_type: Literal["fts"] = "fts", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceFtsQueryBuilder: ... @overload def search( self, query: Optional[ Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery] ] = None, vector_column_name: Optional[str] = None, query_type: Literal["hybrid"] = "hybrid", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceHybridQueryBuilder: ... @overload def search( self, query: None = None, vector_column_name: Optional[str] = None, query_type: QueryType = "auto", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceEmptyQueryBuilder: ... def search( self, query: Optional[ Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery] ] = None, vector_column_name: Optional[str] = None, query_type: QueryType = "auto", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, fast_search: Optional[bool] = None, ) -> LanceQueryBuilder: """Create a search query to find the nearest neighbors of the given query vector. We currently support [vector search](https://lancedb.com/docs/search/vector-search/) and [full-text search](https://lancedb.com/docs/search/full-text-search/). Examples -------- >>> import lancedb >>> db = lancedb.connect("./.lancedb") >>> data = [ ... {"original_width": 100, "caption": "bar", "vector": [0.1, 2.3, 4.5]}, ... {"original_width": 2000, "caption": "foo", "vector": [0.5, 3.4, 1.3]}, ... {"original_width": 3000, "caption": "test", "vector": [0.3, 6.2, 2.6]} ... ] >>> table = db.create_table("my_table", data) >>> query = [0.4, 1.4, 2.4] >>> (table.search(query) ... .where("original_width > 1000", prefilter=True) ... .select(["caption", "original_width", "vector"]) ... .limit(2) ... .to_pandas()) caption original_width vector _distance 0 foo 2000 [0.5, 3.4, 1.3] 5.220000 1 test 3000 [0.3, 6.2, 2.6] 23.089996 Parameters ---------- query: list/np.ndarray/str/PIL.Image.Image, default None The targeted vector to search for. - *default None*. Acceptable types are: list, np.ndarray, PIL.Image.Image - If None then the select/[where][lancedb.query.LanceQueryBuilder.where]/limit clauses are applied to filter the table vector_column_name: str, optional The name of the vector column to search. The vector column needs to be a pyarrow fixed size list type *default "vector"* - If not specified then the vector column is inferred from the table schema - If the table has multiple vector columns then the *vector_column_name* needs to be specified. Otherwise, an error is raised. query_type: str, default "auto" "vector", "fts", or "auto" If "auto" then the query type is inferred from the query; If `query` is a list/np.ndarray then the query type is "vector"; If `query` is a PIL.Image.Image then either do vector search or raise an error if no corresponding embedding function is found. If the `query` is a string, then the query type is "vector" if the table has embedding functions, else the query type is "fts" fts_columns: str or list of str, default None The column(s) to search in for full-text search. If None then the search is performed on all indexed columns. For now, only one column can be searched at a time. Returns ------- LanceQueryBuilder A query builder object representing the query. Once executed, the query returns selected columns, the vector, and also the "_distance" column which is the distance between the query vector and the returned vector. """ if isinstance(query, FullTextQuery): query_type = "fts" vector_column_name = infer_vector_column_name( schema=self.schema, query_type=query_type, query=query, vector_column_name=vector_column_name, ) return LanceQueryBuilder.create( self, query, query_type, vector_column_name=vector_column_name, ordering_field_name=ordering_field_name, fts_columns=fts_columns or [], fast_search=fast_search, ) @classmethod def create( cls, db: LanceDBConnection, name: str, data: Optional[DATA] = None, schema: Optional[pa.Schema] = None, mode: CreateMode = "create", exist_ok: bool = False, on_bad_vectors: OnBadVectorsType = "error", fill_value: float = 0.0, embedding_functions: Optional[List[EmbeddingFunctionConfig]] = None, *, namespace_path: Optional[List[str]] = None, storage_options: Optional[Dict[str, str | bool]] = None, data_storage_version: Optional[str] = None, enable_v2_manifest_paths: Optional[bool] = None, location: Optional[str] = None, namespace_client: Optional[Any] = None, pushdown_operations: Optional[set] = None, route_pushdown_to_rust: bool = False, ): """ Create a new table. Examples -------- >>> import lancedb >>> data = [ ... {"x": 1, "vector": [1.0, 2]}, ... {"x": 2, "vector": [3.0, 4]}, ... {"x": 3, "vector": [5.0, 6]} ... ] >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 2 [3.0, 4.0] 2 3 [5.0, 6.0] Parameters ---------- db: LanceDB The LanceDB instance to create the table in. name: str The name of the table to create. data: list-of-dict, dict, pd.DataFrame, default None The data to insert into the table. At least one of `data` or `schema` must be provided. schema: pa.Schema or LanceModel, optional The schema of the table. If not provided, the schema is inferred from the data. At least one of `data` or `schema` must be provided. mode: str, default "create" The mode to use when writing the data. Valid values are "create", "overwrite", and "append". exist_ok: bool, default False If the table already exists then raise an error if False, otherwise just open the table, it will not add the provided data but will validate against any schema that's specified. on_bad_vectors: str, default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. One of "error", "drop", "fill", "null". fill_value: float, default 0. The value to use when filling vectors. Only used if on_bad_vectors="fill". embedding_functions: list of EmbeddingFunctionModel, default None The embedding functions to use when creating the table. data_storage_version: optional, str, default "stable" Deprecated. Set `storage_options` when connecting to the database and set `new_table_data_storage_version` in the options. enable_v2_manifest_paths: optional, bool, default False Deprecated. Set `storage_options` when connecting to the database and set `new_table_enable_v2_manifest_paths` in the options. """ if namespace_path is None: namespace_path = [] self = cls.__new__(cls) self._conn = db self._namespace_path = namespace_path self._location = location self._namespace_client = namespace_client self._pushdown_operations = pushdown_operations or set() self._route_pushdown_to_rust = route_pushdown_to_rust if data_storage_version is not None: warnings.warn( "setting data_storage_version directly on create_table is deprecated. " "Use database_options instead.", DeprecationWarning, stacklevel=2, ) if storage_options is None: storage_options = {} storage_options["new_table_data_storage_version"] = data_storage_version if enable_v2_manifest_paths is not None: warnings.warn( "setting enable_v2_manifest_paths directly on create_table is " "deprecated. Use database_options instead.", DeprecationWarning, stacklevel=2, ) if storage_options is None: storage_options = {} storage_options["new_table_enable_v2_manifest_paths"] = str( enable_v2_manifest_paths ).lower() self._table = LOOP.run( self._conn._conn.create_table( name, data, schema=schema, mode=mode, exist_ok=exist_ok, on_bad_vectors=on_bad_vectors, fill_value=fill_value, embedding_functions=embedding_functions, namespace_path=namespace_path, storage_options=storage_options, location=location, namespace_client=namespace_client, ) ) return self def delete(self, where: Union[str, Expr]) -> DeleteResult: predicate = where._inner if isinstance(where, Expr) else where return LOOP.run(self._table.delete(predicate)) def update( self, where: Optional[Union[str, Expr]] = None, values: Optional[dict] = None, *, values_sql: Optional[Dict[str, str]] = None, ) -> UpdateResult: """ This can be used to update zero to all rows depending on how many rows match the where clause. Parameters ---------- where: str or [Expr][lancedb.expr.Expr], optional The filter condition. Can be a SQL string or a type-safe [Expr][lancedb.expr.Expr] built with [col][lancedb.expr.col] and [lit][lancedb.expr.lit]. The filter must not be empty, or it will error. values: dict, optional The values to update. The keys are the column names and the values are the values to set. values_sql: dict, optional The values to update, expressed as SQL expression strings. These can reference existing columns. For example, {"x": "x + 1"} will increment the x column by 1. Returns ------- UpdateResult - rows_updated: The number of rows that were updated - version: The new version number of the table after the update Examples -------- >>> import lancedb >>> from lancedb.expr import col >>> import pandas as pd >>> data = pd.DataFrame({"x": [1, 2, 3], "vector": [[1.0, 2], [3, 4], [5, 6]]}) >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 2 [3.0, 4.0] 2 3 [5.0, 6.0] >>> table.update(where=col("x") == 2, values={"vector": [10.0, 10]}) UpdateResult(rows_updated=1, version=2) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 3 [5.0, 6.0] 2 2 [10.0, 10.0] """ return LOOP.run(self._table.update(values, where=where, updates_sql=values_sql)) def _execute_query( self, query: Query, *, batch_size: Optional[int] = None, timeout: Optional[timedelta] = None, ) -> pa.RecordBatchReader: # Branch queries run locally: the server-side query protocol can't # carry a branch yet. # TODO: push down server-side once it can (with remote table support). if ( _should_push_down_query_table( self._namespace_client, self._pushdown_operations ) and not self._route_pushdown_to_rust and self.current_branch() is None and query.take_offsets is None and query.search_effort is None ): from lancedb.namespace import _execute_server_side_query table_id = self._namespace_path + [self.name] return _execute_server_side_query(self._namespace_client, table_id, query) async_iter = LOOP.run( self._table._execute_query(query, batch_size=batch_size, timeout=timeout) ) def iter_sync(): try: while True: yield LOOP.run(async_iter.__anext__()) except StopAsyncIteration: return return pa.RecordBatchReader.from_batches(async_iter.schema, iter_sync()) def _explain_plan(self, query: Query, verbose: Optional[bool] = False) -> str: return LOOP.run(self._table._explain_plan(query, verbose)) def _analyze_plan( self, query: Query, *, distributed_metrics: AnalyzePlanDistributedMetrics = "aggregate", ) -> str: return LOOP.run( self._table._analyze_plan(query, distributed_metrics=distributed_metrics) ) def _output_schema(self, query: Query) -> pa.Schema: return LOOP.run(self._table._output_schema(query)) def _do_merge( self, merge: LanceMergeInsertBuilder, new_data: DATA, on_bad_vectors: OnBadVectorsType, fill_value: float, ) -> MergeResult: return LOOP.run( self._table._do_merge(merge, new_data, on_bad_vectors, fill_value) ) @property def _inner(self) -> LanceDBTable: return self._table._inner @deprecation.deprecated( deprecated_in="0.21.0", current_version=__version__, details="Use `Table.optimize` instead.", ) def cleanup_old_versions( self, older_than: Optional[timedelta] = None, *, delete_unverified: bool = False, ) -> "CleanupStats": """ Clean up old versions of the table, freeing disk space. Parameters ---------- older_than: timedelta, default None The minimum age of the version to delete. If None, then this defaults to two weeks. delete_unverified: bool, default False Because they may be part of an in-progress transaction, files newer than 7 days old are not deleted by default. If you are sure that there are no in-progress transactions, then you can set this to True to delete all files older than `older_than`. Returns ------- CleanupStats The stats of the cleanup operation, including how many bytes were freed. """ return self.to_lance().cleanup_old_versions( older_than, delete_unverified=delete_unverified ) @deprecation.deprecated( deprecated_in="0.21.0", current_version=__version__, details="Use `Table.optimize` instead.", ) def compact_files(self, *args, **kwargs) -> CompactionStats: """ Run the compaction process on the table. This can be run after making several small appends to optimize the table for faster reads. Arguments are passed onto `lance.dataset.DatasetOptimizer.compact_files`. (see Lance documentation for more details) For most cases, the default should be fine. """ stats = self.to_lance().optimize.compact_files(*args, **kwargs) self.checkout_latest() return stats def optimize( self, *, cleanup_older_than: Optional[timedelta] = None, delete_unverified: bool = False, retrain: bool = False, compaction_options: Optional[CompactionOptions] = None, ): """ Optimize the on-disk data and indices for better performance. Modeled after ``VACUUM`` in PostgreSQL. Optimization covers three operations: * Compaction: Merges small files into larger ones * Prune: Removes old versions of the dataset * Index: Optimizes the indices, adding new data to existing indices Parameters ---------- cleanup_older_than: timedelta, optional default 7 days All files belonging to versions older than this will be removed. Set to 0 days to remove all versions except the latest. The latest version is never removed. .. warning:: Setting this to 0 deletes the data files of every older version, so any other reader or writer still using an older version of the table will fail. Only set it to 0 if no other process is working on this dataset. delete_unverified: bool, default False Files leftover from a failed transaction may appear to be part of an in-progress operation (e.g. appending new data) and these files will not be deleted unless they are at least 7 days old. If delete_unverified is True then these files will be deleted regardless of their age. .. warning:: This should only be set to True if you can guarantee that no other process is currently working on this dataset. Otherwise the dataset could be put into a corrupted state. retrain: bool, default False This parameter is no longer used and is deprecated. compaction_options: CompactionOptions, optional Options that control file compaction. For large rows, derive a lower ``target_rows_per_fragment`` from the expected row size to bound later passes. This does not cap the first pass over an existing oversized fragment; see [CompactionOptions][lancedb.table.CompactionOptions]. Notes ----- The frequency an application should call optimize is based on the frequency of data modifications. If data is frequently added, deleted, or updated then optimize should be run frequently. A good rule of thumb is to run optimize if you have added or modified 100,000 or more records or run more than 20 data modification operations. """ LOOP.run( self._table._do_optimize( _optimize_cleanup_since_ms(cleanup_older_than, retrain), delete_unverified, compaction_options, ) ) def list_indices(self) -> Iterable[IndexConfig]: """ List all indices that have been created with Self::create_index """ return LOOP.run(self._table.list_indices()) def tokenize( self, query: str, *, column: Optional[str] = None, index_name: Optional[str] = None, ) -> Iterable[FtsToken]: """ Tokenize a query using the tokenizer configured on an FTS index. Specify exactly one of ``column`` or ``index_name``. Model-backed tokenizers such as ``jieba/*`` and ``lindera/*`` are rebuilt in the client process from index metadata. For remote tables, this means the same tokenizer model files must also exist locally (under ``LANCE_LANGUAGE_MODEL_HOME``) and must match the files the server uses. A mismatch cannot be detected: the returned tokens may differ from what the server's index produces. If the model files are missing, an error naming the tokenizer is raised. """ return LOOP.run( self._table.tokenize(query, column=column, index_name=index_name) ) def index_stats(self, index_name: str) -> Optional[IndexStatistics]: """ Retrieve statistics about an index Parameters ---------- index_name: str The name of the index to retrieve statistics for Returns ------- IndexStatistics or None The statistics about the index. Returns None if the index does not exist. """ return LOOP.run(self._table.index_stats(index_name)) def add_columns( self, transforms: Dict[str, str | FunctionApplication] | FunctionApplication | pa.Field | List[pa.Field] | pa.Schema | None = None, *, computed: Dict[str, str] | None = None, auto_refresh_enabled: bool | None = None, auto_refresh_interval_seconds: int | None = None, ) -> AddColumnsResult: return LOOP.run( self._table.add_columns( transforms, computed=computed, auto_refresh_enabled=auto_refresh_enabled, auto_refresh_interval_seconds=auto_refresh_interval_seconds, ) ) def refresh_column(self, column: str) -> "RefreshColumnResult": """Fill a computed column's unfilled rows and recompute those whose inputs changed. See [`AsyncTable.refresh_column`][lancedb.AsyncTable.refresh_column].""" return LOOP.run(self._table.refresh_column(column)) def refresh_column_async(self, column: str) -> Job[RefreshColumnJobResult]: """Fill a computed column's unfilled rows and recompute those whose inputs changed, returning a handle to the refresh job. See [`Table.refresh_column_async`][lancedb.table.Table.refresh_column_async]. """ return Job(LOOP.run(self._table.refresh_column_async(column))) def function_errors( self, job_id: Optional[str] = None, column: Optional[str] = None, limit: Optional[int] = None, ) -> "FunctionErrors": """The per-row errors Function refreshes recorded on this table. See [`Table.function_errors`][lancedb.table.Table.function_errors].""" return LOOP.run( self._table.function_errors(job_id=job_id, column=column, limit=limit) ) def alter_columns( self, *alterations: Iterable[Dict[str, str]] ) -> AlterColumnsResult: return LOOP.run(self._table.alter_columns(*alterations)) def update_field_metadata( self, *updates: dict[str, Any] ) -> UpdateFieldMetadataResult: return LOOP.run(self._table.update_field_metadata(*updates)) def drop_columns(self, columns: Iterable[str]) -> DropColumnsResult: return LOOP.run(self._table.drop_columns(columns)) def set_unenforced_primary_key(self, columns: Union[str, Iterable[str]]) -> None: """Set the unenforced primary key. See [`AsyncTable.set_unenforced_primary_key`][lancedb.AsyncTable.set_unenforced_primary_key].""" return LOOP.run(self._table.set_unenforced_primary_key(columns)) def enable_rte(self, spec: "RTEWriteSpec") -> None: """Install an RTEWriteSpec. See [`AsyncTable.enable_rte`][lancedb.AsyncTable.enable_rte].""" return LOOP.run(self._table.enable_rte(spec)) def disable_rte(self) -> None: """Remove the RTEWriteSpec. See [`AsyncTable.disable_rte`][lancedb.AsyncTable.disable_rte].""" return LOOP.run(self._table.disable_rte()) def get_rte_spec(self) -> Optional["RTEWriteSpec"]: """Read the installed RTEWriteSpec, or ``None``. See [`AsyncTable.get_rte_spec`][lancedb.AsyncTable.get_rte_spec].""" return LOOP.run(self._table.get_rte_spec()) def checkpoint_rte(self) -> None: """Synchronous version of [`AsyncTable.checkpoint_rte`][lancedb.AsyncTable.checkpoint_rte].""" return LOOP.run(self._table.checkpoint_rte()) def flush_rte(self) -> None: """Synchronous version of [`AsyncTable.flush_rte`][lancedb.AsyncTable.flush_rte].""" return LOOP.run(self._table.flush_rte()) def compact_rte(self) -> None: """Synchronous version of [`AsyncTable.compact_rte`][lancedb.AsyncTable.compact_rte].""" return LOOP.run(self._table.compact_rte()) def get_rte_stats(self, *, include_sstable_rows: bool = False) -> Optional[dict]: """Synchronous version of [`AsyncTable.get_rte_stats`][lancedb.AsyncTable.get_rte_stats].""" return LOOP.run( self._table.get_rte_stats(include_sstable_rows=include_sstable_rows) ) def close_rte_writers(self) -> None: """Close cached RTE shard writers. See [`AsyncTable.close_rte_writers`][lancedb.AsyncTable.close_rte_writers].""" return LOOP.run(self._table.close_rte_writers()) def uses_v2_manifest_paths(self) -> bool: """ Check if the table is using the new v2 manifest paths. Returns ------- bool True if the table is using the new v2 manifest paths, False otherwise. """ return LOOP.run(self._table.uses_v2_manifest_paths()) def migrate_v2_manifest_paths(self): """ Migrate the manifest paths to the new format. This will update the manifest to use the new v2 format for paths. This function is idempotent, and can be run multiple times without changing the state of the object store. !!! danger This should not be run while other concurrent operations are happening. And it should also run until completion before resuming other operations. You can use [LanceTable.uses_v2_manifest_paths][lancedb.table.LanceTable.uses_v2_manifest_paths] to check if the table is already using the new path style. """ LOOP.run(self._table.migrate_manifest_paths_v2()) @deprecation.deprecated( deprecated_in="0.33.1", current_version=__version__, details="Use update_field_metadata() instead.", ) def replace_field_metadata(self, field_name: str, new_metadata: Dict[str, str]): """ Replace the metadata of a field in the schema .. deprecated:: 0.33.1 Use :func:`update_field_metadata` instead. Parameters ---------- field_name: str The name of the field to replace the metadata for new_metadata: dict The new metadata to set """ LOOP.run(self._table.replace_field_metadata(field_name, new_metadata)) def _handle_bad_vectors( reader: pa.RecordBatchReader, on_bad_vectors: Literal["error", "drop", "fill", "null"] = "error", fill_value: float = 0.0, target_schema: Optional[pa.Schema] = None, metadata: Optional[dict] = None, ) -> pa.RecordBatchReader: vector_columns = _find_vector_columns(reader.schema, target_schema, metadata) if not vector_columns: return reader output_schema = _vector_output_schema(reader.schema, vector_columns) def gen(): for batch in reader: pending_dims = [] for vector_column in vector_columns: dim = vector_column["expected_dim"] if target_schema is not None and dim is None: dim = _infer_vector_dim(batch[vector_column["name"]]) pending_dims.append(vector_column) batch = _handle_bad_vector_column( batch, vector_column_name=vector_column["name"], on_bad_vectors=on_bad_vectors, fill_value=fill_value, expected_dim=dim, expected_value_type=vector_column["expected_value_type"], ) for vector_column in pending_dims: if vector_column["expected_dim"] is None: vector_column["expected_dim"] = _infer_vector_dim( batch[vector_column["name"]] ) if batch.schema.equals(output_schema, check_metadata=True): yield batch continue cast_batches = ( pa.Table.from_batches([batch]).cast(output_schema).to_batches() ) if cast_batches: yield pa.RecordBatch.from_arrays( cast_batches[0].columns, schema=output_schema, ) return pa.RecordBatchReader.from_batches(output_schema, gen()) def _find_vector_columns( reader_schema: pa.Schema, target_schema: Optional[pa.Schema], metadata: Optional[dict], ) -> List[dict]: if target_schema is None: vector_columns = [] for field in reader_schema: named_vector_col = ( _is_list_like(field.type) and pa.types.is_floating(field.type.value_type) and field.name == VECTOR_COLUMN_NAME ) likely_vector_col = ( pa.types.is_fixed_size_list(field.type) and pa.types.is_floating(field.type.value_type) and (field.type.list_size >= 10) ) if named_vector_col or likely_vector_col: vector_columns.append( { "name": field.name, "expected_dim": None, "expected_value_type": None, } ) return vector_columns reader_column_names = set(reader_schema.names) active_metadata = _merge_metadata(target_schema.metadata, metadata) embedding_function_columns = set( EmbeddingFunctionRegistry.get_instance().parse_functions(active_metadata).keys() ) vector_columns = [] for field in target_schema: if field.name not in reader_column_names: continue if not _is_list_like(field.type) or not pa.types.is_floating( field.type.value_type ): continue reader_field = reader_schema.field(field.name) named_vector_col = ( field.name in embedding_function_columns or field.name == VECTOR_COLUMN_NAME or (field.name == "embedding" and pa.types.is_fixed_size_list(field.type)) ) typed_fixed_vector_col = ( pa.types.is_fixed_size_list(reader_field.type) and pa.types.is_floating(reader_field.type.value_type) and reader_field.type.list_size >= 10 ) if named_vector_col or typed_fixed_vector_col: vector_columns.append( { "name": field.name, "expected_dim": ( field.type.list_size if pa.types.is_fixed_size_list(field.type) else None ), "expected_value_type": field.type.value_type, } ) return vector_columns def _vector_output_schema( reader_schema: pa.Schema, vector_columns: List[dict], ) -> pa.Schema: columns_by_name = {column["name"]: column for column in vector_columns} fields = [] for field in reader_schema: column = columns_by_name.get(field.name) if column is None: output_type = field.type else: output_type = _vector_output_type(field, column) fields.append(pa.field(field.name, output_type, field.nullable, field.metadata)) return pa.schema(fields, metadata=reader_schema.metadata) def _vector_output_type(field: pa.Field, vector_column: dict) -> pa.DataType: if not _is_list_like(field.type): return field.type if vector_column["expected_value_type"] is not None and ( pa.types.is_null(field.type.value_type) or pa.types.is_integer(field.type.value_type) or pa.types.is_unsigned_integer(field.type.value_type) ): return pa.list_(vector_column["expected_value_type"]) if ( vector_column["expected_dim"] is not None and pa.types.is_fixed_size_list(field.type) and field.type.list_size != vector_column["expected_dim"] ): return pa.list_(field.type.value_type) return field.type def _handle_bad_vector_column( data: pa.RecordBatch, vector_column_name: str, on_bad_vectors: str = "error", fill_value: float = 0.0, expected_dim: Optional[int] = None, expected_value_type: Optional[pa.DataType] = None, ) -> pa.RecordBatch: """ Ensure that the vector column exists and has type fixed_size_list(float) Parameters ---------- data: pa.Table The table to sanitize. vector_column_name: str The name of the vector column. on_bad_vectors: str, default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. One of "error", "drop", "fill", "null". fill_value: float, default 0.0 The value to use when filling vectors. Only used if on_bad_vectors="fill". """ position = data.column_names.index(vector_column_name) vec_arr = data[vector_column_name] if not _is_list_like(vec_arr.type): return data if ( expected_dim is not None and pa.types.is_fixed_size_list(vec_arr.type) and vec_arr.type.list_size != expected_dim ): vec_arr = pa.array(vec_arr.to_pylist(), type=pa.list_(vec_arr.type.value_type)) data = data.set_column(position, vector_column_name, vec_arr) if expected_value_type is not None and ( pa.types.is_null(vec_arr.type.value_type) or pa.types.is_integer(vec_arr.type.value_type) or pa.types.is_unsigned_integer(vec_arr.type.value_type) ): vec_arr = pa.array(vec_arr.to_pylist(), type=pa.list_(expected_value_type)) data = data.set_column(position, vector_column_name, vec_arr) if pa.types.is_floating(vec_arr.type.value_type): has_nan = has_nan_values(vec_arr) else: has_nan = pa.array([False] * len(vec_arr)) if expected_dim is not None: dim = expected_dim elif pa.types.is_fixed_size_list(vec_arr.type): dim = vec_arr.type.list_size else: dim = _infer_vector_dim(vec_arr) if dim is None: return data is_null = pc.is_null(vec_arr) # pc.list_value_length returns null for null list entries, so # pc.not_equal(null, dim) also returns null. Use or_kleene so that # True OR null = True (Kleene three-valued logic), ensuring null vectors # are counted as wrong-dim. has_wrong_dim = pc.or_kleene( is_null, pc.not_equal(pc.list_value_length(vec_arr), dim), ) has_bad_vectors = pc.any(has_nan).as_py() or pc.any(has_wrong_dim).as_py() if has_bad_vectors: is_bad = pc.or_(has_nan, has_wrong_dim) if on_bad_vectors == "error": if pc.any(has_wrong_dim).as_py(): raise ValueError( f"Vector column '{vector_column_name}' has variable length " "vectors. Set on_bad_vectors='drop' to remove them, " "set on_bad_vectors='fill' and fill_value= to replace them, " "or set on_bad_vectors='null' to replace them with null." ) else: raise ValueError( f"Vector column '{vector_column_name}' has NaNs or null values. " "Set on_bad_vectors='drop' to remove them, " "set on_bad_vectors='fill' and fill_value= to replace them, " "or set on_bad_vectors='null' to replace them with null." ) elif on_bad_vectors == "null": vec_arr = pc.if_else( is_bad, pa.scalar(None), vec_arr, ) elif on_bad_vectors == "drop": data = data.filter(pc.invert(is_bad)) vec_arr = data[vector_column_name] elif on_bad_vectors == "fill": if fill_value is None: raise ValueError( "`fill_value` must not be None if `on_bad_vectors` is 'fill'" ) vec_arr = _fill_bad_vector_values(vec_arr, dim, fill_value) else: raise ValueError(f"Invalid value for on_bad_vectors: {on_bad_vectors}") return data.set_column(position, vector_column_name, vec_arr) def _fill_bad_vector_values( arr: Union[pa.Array, pa.ChunkedArray], dim: int, fill_value: float, ) -> pa.Array: if not isinstance(arr, pa.ChunkedArray): arr = pa.chunked_array([arr]) arr = arr.combine_chunks() # A fixed-size slice truncates long vectors and pads short vectors with nulls. # Both padding and existing null elements need to be filled. sliced = pc.list_slice(arr, 0, dim, return_fixed_size_list=True) values = sliced.values needs_fill = pc.is_null(values) if pa.types.is_floating(values.type): values_for_nan_check = ( values.cast(pa.float32()) if pa.types.is_float16(values.type) else values ) needs_fill = pc.or_kleene(needs_fill, pc.is_nan(values_for_nan_check)) fill_scalar = pa.scalar(fill_value).cast(values.type) filled_values = pc.if_else(needs_fill, fill_scalar, values) filled = pa.FixedSizeListArray.from_arrays(filled_values, dim) return filled.cast(arr.type) def has_nan_values(arr: Union[pa.ListArray, pa.ChunkedArray]) -> pa.BooleanArray: """Find vectors with NaN or null components, including pandas-converted NaNs.""" if isinstance(arr, pa.ChunkedArray): values = pa.chunked_array([chunk.flatten() for chunk in arr.chunks]) else: values = arr.flatten() if pa.types.is_float16(values.type): # is_nan isn't yet implemented for f16, so we cast to f32 # https://github.com/apache/arrow/issues/45083 values_has_nan = pc.is_nan(values.cast(pa.float32())) else: values_has_nan = pc.is_nan(values) # Pandas represents NaN components as Arrow nulls. is_nan returns null # for these values, so count them as invalid before filtering parent indices. values_has_nan = pc.fill_null(values_has_nan, True) values_indices = pc.list_parent_indices(arr) if arr.null_count: # Match flatten(), which excludes child spans under null parent lists. values_indices = pc.filter( values_indices, pc.take(pc.is_valid(arr), values_indices) ) has_nan_indices = pc.unique(pc.filter(values_indices, values_has_nan)) indices = pa.array(range(len(arr)), type=pa.uint32()) return pc.is_in(indices, has_nan_indices) def _is_list_like(data_type: pa.DataType) -> bool: return ( pa.types.is_list(data_type) or pa.types.is_large_list(data_type) or pa.types.is_fixed_size_list(data_type) ) def _merge_metadata(*metadata_dicts: Optional[dict]) -> dict: merged = {} for metadata in metadata_dicts: if metadata is None: continue for key, value in metadata.items(): if isinstance(key, str): key = key.encode("utf-8") if isinstance(value, str): value = value.encode("utf-8") merged[key] = value return merged def _name_suggests_vector_column(field_name: str) -> bool: """Check if a field name indicates a vector column.""" name_lower = field_name.lower() return "vector" in name_lower or "embedding" in name_lower def _infer_target_schema( reader: pa.RecordBatchReader, ) -> Tuple[pa.Schema, pa.RecordBatchReader]: schema = reader.schema peeked = None for i, field in enumerate(schema): is_list_type = pa.types.is_list(field.type) or pa.types.is_large_list( field.type ) if _name_suggests_vector_column(field.name) and is_list_type: if peeked is None: peeked, reader = peek_reader(reader) # Use the most common length of the list as the dimensions dim = _modal_list_size(peeked.column(i)) if dim is None: # Every list entry is null (e.g. embeddings not computed yet), # so the dimension is unknowable. Leave the column as a # variable-length list instead of guessing a fixed size. continue # Determine target type based on value type if pa.types.is_floating(field.type.value_type): target_type = pa.list_(pa.float32(), dim) elif pa.types.is_integer(field.type.value_type): values = peeked.column(i) if isinstance(values, pa.ChunkedArray): values = values.combine_chunks() flattened = values.flatten() valid_count = pc.count(flattened, mode="only_valid").as_py() if valid_count == 0: target_type = pa.list_(pa.uint8(), dim) else: min_max = pc.min_max(flattened) min_value = min_max["min"].as_py() max_value = min_max["max"].as_py() if (min_value is not None and min_value < 0) or ( max_value is not None and max_value > 255 ): target_type = pa.list_(pa.float32(), dim) else: target_type = pa.list_(pa.uint8(), dim) else: continue # Skip non-numeric types new_field = pa.field( field.name, # preserve original field name target_type, nullable=field.nullable, metadata=field.metadata, ) schema = schema.set(i, new_field) return schema, reader def _modal_list_size(arr: Union[pa.ListArray, pa.ChunkedArray]) -> Optional[int]: # Use the most common length of the list as the dimensions. pyarrow's # `mode` skips nulls, so an all-null column yields an empty result; return # None in that case so the caller can keep a variable-length list. modes = pc.mode(pc.list_value_length(arr)) if len(modes) == 0: return None return modes[0].as_py()["mode"] def _infer_vector_dim(arr: Union[pa.Array, pa.ChunkedArray]) -> Optional[int]: if not _is_list_like(arr.type): return None lengths = pc.list_value_length(arr) lengths = pc.filter(lengths, pc.greater(lengths, 0)) if len(lengths) == 0: return None return pc.mode(lengths)[0].as_py()["mode"] def _validate_schema(schema: pa.Schema): """ Make sure the metadata is valid utf8 """ if schema.metadata is not None: _validate_metadata(schema.metadata) def _validate_metadata(metadata: dict): """ Make sure the metadata values are valid utf8 (can be nested) Raises ValueError if not valid utf8 """ for k, v in metadata.items(): if isinstance(v, bytes): try: v.decode("utf8") except UnicodeDecodeError: raise ValueError( f"Metadata key {k} is not valid utf8. " "Consider base64 encode for generic binary metadata." ) elif isinstance(v, dict): _validate_metadata(v) class AsyncTable: """ An AsyncTable is a collection of Records in a LanceDB Database. An AsyncTable can be obtained from the [AsyncConnection.create_table][lancedb.AsyncConnection.create_table] and [AsyncConnection.open_table][lancedb.AsyncConnection.open_table] methods. An AsyncTable object is expected to be long lived and reused for multiple operations. AsyncTable objects will cache a certain amount of index data in memory. This cache will be freed when the Table is garbage collected. To eagerly free the cache you can call the [close][lancedb.AsyncTable.close] method. Once the AsyncTable is closed, it cannot be used for any further operations. An AsyncTable can also be used as a context manager, and will automatically close when the context is exited. Closing a table is optional. If you do not close the table, it will be closed when the AsyncTable object is garbage collected. Examples -------- Create using [AsyncConnection.create_table][lancedb.AsyncConnection.create_table] (more examples in that method's documentation). >>> import lancedb >>> async def create_a_table(): ... db = await lancedb.connect_async("./.lancedb") ... data = [{"vector": [1.1, 1.2], "b": 2}] ... table = await db.create_table("my_table", data=data) ... print(await table.query().limit(5).to_arrow()) >>> import asyncio >>> asyncio.run(create_a_table()) pyarrow.Table vector: fixed_size_list[2] child 0, item: float b: int64 ---- vector: [[[1.1,1.2]]] b: [[2]] Can append new data with [AsyncTable.add()][lancedb.table.AsyncTable.add]. >>> async def add_to_table(): ... db = await lancedb.connect_async("./.lancedb") ... table = await db.open_table("my_table") ... await table.add([{"vector": [0.5, 1.3], "b": 4}]) >>> asyncio.run(add_to_table()) Can query the table with [AsyncTable.vector_search][lancedb.table.AsyncTable.vector_search]. >>> async def search_table_for_vector(): ... db = await lancedb.connect_async("./.lancedb") ... table = await db.open_table("my_table") ... results = ( ... await table.vector_search([0.4, 0.4]).select(["b", "vector"]).to_pandas() ... ) ... print(results) >>> asyncio.run(search_table_for_vector()) b vector _distance 0 4 [0.5, 1.3] 0.82 1 2 [1.1, 1.2] 1.13 Search queries are much faster when an index is created. See [AsyncTable.create_index][lancedb.table.AsyncTable.create_index]. """ def __init__( self, table: LanceDBTable, *, namespace_path: Optional[List[str]] = None, namespace_client: Optional[Any] = None, pushdown_operations: Optional[set] = None, route_pushdown_to_rust: bool = False, ): """Create a new AsyncTable object. You should not create AsyncTable objects directly. Use [AsyncConnection.create_table][lancedb.AsyncConnection.create_table] and [AsyncConnection.open_table][lancedb.AsyncConnection.open_table] to obtain Table objects.""" self._inner = table self._namespace_path = namespace_path or [] self._namespace_client = namespace_client self._pushdown_operations = pushdown_operations or set() # See LanceTable.__init__: defer QueryTable pushdown to Rust (which emits # the read-freshness header) for natively-built namespace clients. self._route_pushdown_to_rust = route_pushdown_to_rust def _set_namespace_context( self, *, namespace_path: Optional[List[str]] = None, namespace_client: Optional[Any] = None, pushdown_operations: Optional[set] = None, route_pushdown_to_rust: bool = False, ) -> "AsyncTable": self._namespace_path = namespace_path or [] self._namespace_client = namespace_client self._pushdown_operations = pushdown_operations or set() self._route_pushdown_to_rust = route_pushdown_to_rust return self def __repr__(self): return self._inner.__repr__() def __enter__(self): return self def __exit__(self, *_): self.close() def is_open(self) -> bool: """Return True if the table is open.""" return self._inner.is_open() def close(self): """Close the table and free any resources associated with it. It is safe to call this method multiple times. Any attempt to use the table after it has been closed will raise an error.""" return self._inner.close() async def set_unenforced_primary_key( self, columns: Union[str, Iterable[str]] ) -> None: """Set the unenforced primary key for this table to the given ordered list of columns. "Unenforced" means LanceDB does not check uniqueness on writes; the columns are recorded in the schema as the primary key so that features such as `merge_insert` can use them. Calling this again replaces any previously-set primary key. Parameters ---------- columns : str or Iterable[str] Either a single column name (single-column key) or an ordered iterable of column names (composite key). Each column dtype must be one of: int32, int64, utf8, large_utf8, binary, large_binary, fixed_size_binary. """ if isinstance(columns, str): columns = [columns] else: columns = list(columns) await self._inner.set_unenforced_primary_key(columns) async def enable_rte(self, spec: "RTEWriteSpec") -> None: """Install an RTEWriteSpec on this table. The spec selects Lance's MemWAL LSM-style write path for future `merge_insert` calls. ``RTEWriteSpec`` chooses one of three sharding strategies: - ``RTEWriteSpec.bucket(column, num_buckets)`` — hash-bucket writes by the single-column unenforced primary key. - ``RTEWriteSpec.identity(column)`` — shard by the raw value of a scalar column. - ``RTEWriteSpec.unsharded()`` — route every write to a single shard. All variants require the table to have an unenforced primary key set via [`set_unenforced_primary_key`]; bucket sharding additionally requires it to be the single column being bucketed. By default the RTE maintains every index on the table, re-read as the table changes, so an index created later is maintained too; one of a kind the RTE cannot mirror is skipped. Name the set with ``with_maintained_indexes`` to pin it instead: each name must already exist and be maintainable (a still-building index is rejected, not omitted); ``[]`` maintains none. Parameters ---------- spec : RTEWriteSpec The sharding spec to install. Examples -------- >>> from lancedb import RTEWriteSpec >>> # table.set_unenforced_primary_key("id") >>> # table.enable_rte(RTEWriteSpec.bucket("id", 16)) """ await self._inner.enable_rte(spec) async def disable_rte(self) -> None: """Remove the RTEWriteSpec from this table. Reverts to the standard `merge_insert` write path. Errors if no spec is currently set. """ await self._inner.disable_rte() async def get_rte_spec(self) -> Optional["RTEWriteSpec"]: """Read the RTEWriteSpec currently installed on this table. Returns ``None`` when the RTE write path is not enabled at all — no spec has been set, or one was removed with `disable_rte`. That is a different answer from a spec whose ``maintained_indexes`` is ``None``, which is an installed spec selecting indexes automatically. The spec read back is the one that was installed, selection included: ``None`` maintains every supported index the table has now or gains later, ``[]`` maintains none, and a non-empty list maintains exactly those. All three round-trip. """ return await self._inner.get_rte_spec() # WAL-PK-FUSION: delete both hooks. def _hybrid_pk_fusion_learned(self) -> bool: """See [`Table._hybrid_pk_fusion_learned`][lancedb.table.Table].""" return self._inner.hybrid_pk_fusion_learned() def _note_hybrid_pk_fusion(self) -> None: """See [`Table._note_hybrid_pk_fusion`][lancedb.table.Table].""" self._inner.note_hybrid_pk_fusion() async def checkpoint_rte(self) -> None: """Converge this table's RTE write path into its base table. One flush, freezing every memtable into an SSTable, then compaction triggers until every SSTable that existed at that moment has reached base. The loop runs client-side, reading progress from ``get_rte_stats``. Best-effort: SSTables created *while* it runs are deliberately not waited on, which is what lets it terminate on a table taking writes. Idempotent and safe on a cadence. There is no deadline, and the caller owns that. It returns when the target SSTables are gone, raises on a terminal server fault, and otherwise waits however long the server takes. A slow table and a stuck one are the same picture from the client: the compactor pool is shared across every table on the node, so a checkpoint queued behind unrelated work looks exactly like one that is merging. Wrap this in ``asyncio.wait_for`` for a wall-clock bound; abandoning it partway costs nothing. """ await self._inner.checkpoint_rte() async def flush_rte(self) -> None: """Freeze every table shard's active memtable into an SSTable. Does not touch the base table — compacting SSTables into base is `compact_rte`. On a node that has not claimed this table, this claims it and replays its WAL log first. """ await self._inner.flush_rte() async def compact_rte(self) -> None: """Trigger a background SSTable compaction pass per table shard. Returns once the passes are dispatched, not once they finish: watch ``get_rte_stats`` for progress, or use ``checkpoint_rte`` to loop until the current SSTables have reached base. """ await self._inner.compact_rte() async def get_rte_stats( self, *, include_sstable_rows: bool = False ) -> Optional[dict]: """Read live per-table-shard RTE state. Answers "how far behind is my fresh tier", "which table shard is hot", and "why is my fresh-tier vector search brute-force". Mutates no table state, though on a node that has not claimed this table it claims it, exactly as a read would. Returns ``None`` only when the RTE write path is not enabled. Parameters ---------- include_sstable_rows Report a row count per SSTable. Off by default: each count opens an uncached Lance dataset, and ``checkpoint_rte`` polls this needing only generation numbers. """ return await self._inner.get_rte_stats(include_sstable_rows) async def close_rte_writers(self) -> None: """Drain and close any cached RTE shard writers for this table. When an RTE write spec is installed, `merge_insert` opens RTE shard writers and caches them for reuse across calls. This closes them, flushing pending data; writers reopen lazily on the next `merge_insert`. It is a no-op when no writers are cached. """ await self._inner.close_rte_writers() @property def name(self) -> str: """The name of the table.""" return self._inner.name() async def schema(self) -> pa.Schema: """The [Arrow Schema](https://arrow.apache.org/docs/python/api/datatypes.html#) of this Table """ return await self._inner.schema() async def embedding_functions(self) -> Dict[str, EmbeddingFunctionConfig]: """ Get the embedding functions for the table Returns ------- funcs: Dict[str, EmbeddingFunctionConfig] A mapping of the vector column to the embedding function or empty dict if not configured. """ schema = await self.schema() return EmbeddingFunctionRegistry.get_instance().parse_functions(schema.metadata) async def count_rows(self, filter: Optional[str] = None) -> int: """ Count the number of rows in the table. Parameters ---------- filter: str, optional A SQL where clause to filter the rows to count. """ return await self._inner.count_rows(filter) async def head(self, n=5) -> pa.Table: """ Return the first `n` rows of the table. Parameters ---------- n: int, default 5 The number of rows to return. """ return await self.query().limit(n).to_arrow() def query(self) -> AsyncQuery: """ Returns an [AsyncQuery][lancedb.query.AsyncQuery] that can be used to search the table. Use methods on the returned query to control query behavior. The query can be executed with methods like [to_arrow][lancedb.query.AsyncQuery.to_arrow], [to_pandas][lancedb.query.AsyncQuery.to_pandas] and more. """ return AsyncQuery(self._inner.query(), self) async def to_lance(self, **kwargs) -> lance.LanceDataset: """Return the Lance dataset backing a local table. Remote tables cannot be opened as Lance datasets by the client. Use [to_arrow][lancedb.table.AsyncTable.to_arrow] or [query][lancedb.table.AsyncTable.query] to read them through the server. Parameters ---------- **kwargs Forwarded to `lance.dataset`. Returns ------- lance.LanceDataset The Lance dataset at this table handle's version and branch. Examples -------- >>> async def get_lance_dataset(table): ... return await table.to_lance() """ if not self._inner._is_native(): raise NotImplementedError( "to_lance() is not supported for remote tables; " "query the server instead" ) try: import lance except ImportError: raise ImportError( "The lance library is required to use this function. " "Please install with `pip install pylance`." ) # lance.dataset() can't open a branch directly, so open the base table # and check out the branch ref (a None branch resolves to main). branch = self.current_branch() table_version = await self.version() version = None if branch is not None else table_version ds = lance.dataset( await self.uri(), version=version, storage_options=await self.latest_storage_options(), **kwargs, ) if branch is not None: ds = ds.checkout_version((branch, table_version)) return ds async def to_pandas(self, blob_mode: BlobMode = "lazy", **kwargs) -> "pd.DataFrame": """Return the table as a pandas DataFrame. Parameters ---------- blob_mode: str, default "lazy" Controls how Lance blob columns are returned. Remote tables support "descriptions"; "bytes" and "lazy" are not yet supported. **kwargs Forwarded to PyArrow / Lance pandas conversion. Returns ------- pd.DataFrame """ validate_blob_mode(blob_mode) schema = await self.schema() if blob_mode == "descriptions" or not schema_has_blob_field(schema): arrow_tbl = await self.to_arrow() if blob_mode == "descriptions": arrow_tbl = strip_auto_row_ids( arrow_tbl, row_addressable_blob_v2_paths(schema) ) return arrow_tbl.to_pandas(**kwargs) if not self._inner._is_native(): return await self.query().to_pandas(blob_mode=blob_mode, **kwargs) if blob_mode == "lazy" and get_uri_scheme(await self.uri()) == "memory": return (await self.to_arrow()).to_pandas(**kwargs) if blob_mode == "bytes" and blob_v2_column_paths(schema): return await self.query().to_pandas(blob_mode=blob_mode, **kwargs) return (await self.to_lance()).to_pandas(blob_mode=blob_mode, **kwargs) async def to_arrow(self, *, blob_mode: BlobMode = "lazy") -> pa.Table: """Return the table as a pyarrow Table. Parameters ---------- blob_mode: str, default "lazy" ``"lazy"`` returns blob v2 descriptors that carry the row id ``fetch_blobs`` reads. ``"descriptions"`` returns them without it. ``"bytes"`` returns payload bytes and raises for legacy blob columns and blobs inside lists. Returns ------- pa.Table """ validate_blob_mode(blob_mode) schema = await self.schema() fetches_bytes = blob_mode == "bytes" and schema_has_blob_field(schema) if ( not fetches_bytes and _should_push_down_query_table( self._namespace_client, self._pushdown_operations ) and not self._route_pushdown_to_rust ): tbl = (await self._execute_query(Query())).read_all() if blob_mode == "descriptions": tbl = strip_auto_row_ids(tbl, row_addressable_blob_v2_paths(schema)) return tbl # The query builder refuses bytes under pushdown before running the query. return await self.query().to_arrow(blob_mode=blob_mode) async def create_index( self, column: str, *, replace: Optional[bool] = None, config: Optional[ Union[ IvfFlat, IvfPq, IvfRq, HnswPq, HnswSq, HnswFlat, BTree, Bitmap, LabelList, Fm, FTS, ] ] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, ): """Create an index to speed up queries Indices can be created on vector columns or scalar columns. Indices on vector columns will speed up vector searches. Indices on scalar columns will speed up filtering (in both vector and non-vector searches) Parameters ---------- column: str The column to index. replace: bool, default True Whether to replace the existing index If this is false, and another index already exists on the same columns and the same name, then an error will be returned. This is true even if that index is out of date. The default is True config: default None For advanced configuration you can specify the type of index you would like to create. You can also specify index-specific parameters when creating an index object. wait_timeout: timedelta, optional The timeout to wait if indexing is asynchronous. name: str, optional The name of the index. If not provided, a default name will be generated. train: bool, default True Whether to train the index with existing data. Vector indices always train with existing data. """ if config is not None: if not isinstance( config, ( IvfFlat, IvfSq, IvfPq, IvfRq, HnswPq, HnswSq, HnswFlat, BTree, Bitmap, LabelList, Fm, FTS, ), ): raise TypeError( "config must be an instance of IvfSq, IvfPq, IvfRq, HnswPq, HnswSq," " BTree, Bitmap, LabelList, Fm, or FTS, but got " + str(type(config)) ) try: await self._inner.create_index( column, index=config, replace=replace, wait_timeout=wait_timeout, name=name, train=train, ) except (ValueError, RuntimeError) as e: if isinstance(config, FTS): _maybe_add_fts_error_note( e, base_tokenizer=config.base_tokenizer, language=config.language, ) raise e async def create_index_async( self, column: str, *, replace: Optional[bool] = None, config: Optional[ Union[ IvfFlat, IvfPq, IvfRq, HnswPq, HnswSq, HnswFlat, BTree, Bitmap, LabelList, Fm, FTS, ] ] = None, wait_timeout: Optional[timedelta] = None, name: Optional[str] = None, train: bool = True, ) -> AsyncJob: """Create an index, returning a handle to the indexing job. Takes the same arguments as :meth:`create_index`. The job may already be complete when returned; callers must not assume the index exists until :meth:`AsyncJob.wait` resolves. """ job = await self._inner.create_index_async( column, index=config, replace=replace, wait_timeout=wait_timeout, name=name, train=train, ) return AsyncJob(job) async def drop_index(self, name: str) -> None: """ Drop an index from the table. Parameters ---------- name: str The name of the index to drop. Notes ----- This does not delete the index from disk, it just removes it from the table. To delete the index, run [optimize][lancedb.table.AsyncTable.optimize] after dropping the index. Use [list_indices][lancedb.table.AsyncTable.list_indices] to find the names of the indices. """ await self._inner.drop_index(name) async def prewarm_index(self, name: str) -> None: """ Prewarm an index in the table. This is a hint to the database that the index will be accessed in the future and should be loaded into memory if possible. This can reduce cold-start latency for subsequent queries. This call initiates prewarming and returns once the request is accepted. It is idempotent and safe to call from multiple clients concurrently. It is generally wasteful to call this if the index does not fit into the available cache. Not all index types support prewarming; unsupported indices will silently ignore the request. Parameters ---------- name: str The name of the index to prewarm """ await self._inner.prewarm_index(name) async def prewarm_data(self, columns: Optional[List[str]] = None) -> None: """ Prewarm data for the table. This is a hint to the database that the given columns will be accessed in the future and the database should prefetch the data if possible. Currently only supported on remote tables. This call initiates prewarming and returns once the request is accepted. It is idempotent and safe to call from multiple clients concurrently. This operation has a large upfront cost but can speed up future queries that need to fetch the given columns. Large columns such as embeddings or binary data may not be practical to prewarm. This feature is intended for workloads that issue many queries against the same columns. Parameters ---------- columns: list of str, optional The columns to prewarm. If None, all columns are prewarmed. """ await self._inner.prewarm_data(columns) async def wait_for_index( self, index_names: Iterable[str], timeout: timedelta = timedelta(seconds=300) ) -> None: """ Wait for indexing to complete for the given index names. This will poll the table until all the indices are fully indexed, or raise a timeout exception if the timeout is reached. Parameters ---------- index_names: str The name of the indices to poll timeout: timedelta Timeout to wait for asynchronous indexing. The default is 5 minutes. """ await self._inner.wait_for_index(index_names, timeout) async def stats(self) -> TableStatistics: """ Retrieve table and fragment statistics. """ return await self._inner.stats() async def uri(self) -> str: """ Get the table URI (storage location). For remote tables, this fetches the location from the server via describe. For local tables, this returns the dataset URI. Returns ------- str The full storage location of the table (e.g., S3/GCS path). """ return await self._inner.uri() async def initial_storage_options(self) -> Optional[Dict[str, str]]: """Get the initial storage options that were passed in when opening this table. For dynamically refreshed options (e.g., credential vending), use :meth:`latest_storage_options`. Warning: This is an internal API and the return value is subject to change. Returns ------- Optional[Dict[str, str]] The storage options, or None if no storage options were configured. """ return await self._inner.initial_storage_options() async def latest_storage_options(self) -> Optional[Dict[str, str]]: """Get the latest storage options, refreshing from provider if configured. This method is useful for credential vending scenarios where storage options may be refreshed dynamically. If no dynamic provider is configured, this returns the initial static options. Warning: This is an internal API and the return value is subject to change. Returns ------- Optional[Dict[str, str]] The storage options, or None if no storage options were configured. """ return await self._inner.latest_storage_options() async def add( self, data: DATA, *, mode: Optional[Literal["append", "overwrite"]] = "append", on_bad_vectors: Optional[OnBadVectorsType] = None, fill_value: Optional[float] = None, progress: Optional[Union[bool, Callable, Any]] = None, write_parallelism: Optional[int] = None, allow_external_blob_outside_bases: bool = False, ) -> AddResult: """Add more data to the [AsyncTable][lancedb.table.AsyncTable]. Parameters ---------- data: DATA The data to insert into the table. Acceptable types are: - list-of-dict - pandas.DataFrame - pyarrow.Table or pyarrow.RecordBatch mode: str The mode to use when writing the data. Valid values are "append" and "overwrite". Overwrite adopts explicitly declared blob layouts, while inferred bytes retain the existing blob v2 fields. on_bad_vectors: str, default "error" How to handle vectors with incorrect sizes, NaNs, or null elements. One of "error", "drop", "fill", "null". fill_value: float, default 0. The value to use when filling vectors. Only used if on_bad_vectors="fill". progress: callable or tqdm-like, optional A callback or tqdm-compatible progress bar. See :meth:`Table.add` for details. write_parallelism: int, optional Number of partitions to write in parallel. Higher values increase throughput but also peak memory use, since each partition buffers data in flight. Defaults to an estimate based on the data size, capped at the number of CPU cores. Lower this if bulk ingestion is using too much memory. allow_external_blob_outside_bases: bool, default False Allow blob URIs outside registered bases. See :meth:`Table.add`. Local tables only. """ schema = await self.schema() data = _serialize_json_values(data, schema) blob_table = _blob_input_to_arrow(data, schema) if blob_table is not None: data = blob_table if on_bad_vectors is None: on_bad_vectors = "error" if fill_value is None: fill_value = 0.0 # _sanitize_data is an old code path, but we will use it until the # new code path is ready. if mode == "overwrite": # For overwrite, apply the same preprocessing as create_table # so vector columns are inferred as FixedSizeList. data, _ = await asyncio.to_thread( sanitize_create_table, data, None, on_bad_vectors=on_bad_vectors, fill_value=fill_value, ) if blob_v2_column_paths(schema): # Plain bytes cannot be inferred as blob v2. Keep matching blob # fields while allowing other columns to use the inferred schema. target_schema = _preserve_blob_v2_fields(data.schema, schema) data = _cast_to_target_schema(data, target_schema) elif on_bad_vectors != "error" or ( schema.metadata is not None and b"embedding_functions" in schema.metadata ): data = await asyncio.to_thread( _sanitize_data, data, schema, metadata=schema.metadata, on_bad_vectors=on_bad_vectors, fill_value=fill_value, allow_subschema=True, ) _register_optional_converters() # Converting an iterator peeks at its first item. A synchronous query in # that iterator schedules work on LOOP, so peeking on LOOP's own thread # would deadlock waiting for itself. data = await asyncio.to_thread(to_scannable, data) progress, owns = _normalize_progress(progress) try: return await self._inner.add( data, mode or "append", progress=progress, write_parallelism=write_parallelism, allow_external_blob_outside_bases=allow_external_blob_outside_bases, ) except RuntimeError as e: if "Cast error" in str(e): raise ValueError(e) elif "Vector column contains NaN" in str(e): raise ValueError(e) else: raise finally: if owns: progress.close() def merge_insert(self, on: Union[str, Iterable[str]]) -> LanceMergeInsertBuilder: """ Returns a [`LanceMergeInsertBuilder`][lancedb.merge.LanceMergeInsertBuilder] that can be used to create a "merge insert" operation This operation can add rows, update rows, and remove rows all in a single transaction. It is a very generic tool that can be used to create behaviors like "insert if not exists", "update or insert (i.e. upsert)", or even replace a portion of existing data with new data (e.g. replace all data where month="january") The merge insert operation works by combining new data from a **source table** with existing data in a **target table** by using a join. There are three categories of records. "Matched" records are records that exist in both the source table and the target table. "Not matched" records exist only in the source table (e.g. these are new data) "Not matched by source" records exist only in the target table (this is old data) The builder returned by this method can be used to customize what should happen for each category of data. Please note that the data may appear to be reordered as part of this operation. This is because updated rows will be deleted from the dataset and then reinserted at the end with the new values. Parameters ---------- on: Union[str, Iterable[str]] A column (or columns) to join on. This is how records from the source table and target table are matched. Typically this is some kind of key or id column. Passing several columns matches on the composite key: a source row updates a target row only when it agrees on every one of them. Examples -------- >>> import lancedb >>> data = pa.table({"a": [2, 1, 3], "b": ["a", "b", "c"]}) >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> new_data = pa.table({"a": [2, 3, 4], "b": ["x", "y", "z"]}) >>> # Perform a "upsert" operation >>> res = table.merge_insert("a") \\ ... .when_matched_update_all() \\ ... .when_not_matched_insert_all() \\ ... .execute(new_data) >>> res MergeResult(version=2, num_updated_rows=2, num_inserted_rows=1, num_deleted_rows=0, num_attempts=1, num_rows=3) >>> # The order of new rows is non-deterministic since we use >>> # a hash-join as part of this operation and so we sort here >>> table.to_arrow().sort_by("a").to_pandas() a b 0 1 b 1 2 x 2 3 y 3 4 z """ # noqa: E501 on = [on] if isinstance(on, str) else list(iter(on)) return LanceMergeInsertBuilder(self, on) @overload async def search( self, query: None = None, vector_column_name: Optional[str] = None, query_type: QueryType = "auto", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> AsyncQuery: ... @overload async def search( self, query: str, vector_column_name: Optional[str] = None, query_type: Literal["auto"] = ..., ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> Union[AsyncHybridQuery, AsyncFTSQuery, AsyncVectorQuery]: ... @overload async def search( self, query: str, vector_column_name: Optional[str] = None, query_type: Literal["hybrid"] = ..., ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> AsyncHybridQuery: ... @overload async def search( self, query: Union[VEC, "PIL.Image.Image", Tuple], vector_column_name: Optional[str] = None, query_type: Literal["auto"] = ..., ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> AsyncVectorQuery: ... @overload async def search( self, query: str, vector_column_name: Optional[str] = None, query_type: Literal["fts"] = ..., ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> AsyncFTSQuery: ... @overload async def search( self, query: Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery], vector_column_name: Optional[str] = None, query_type: Literal["vector"] = ..., ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> AsyncVectorQuery: ... async def search( self, query: Optional[ Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery] ] = None, vector_column_name: Optional[str] = None, query_type: QueryType = "auto", ordering_field_name: Optional[str] = None, fts_columns: Optional[Union[str, List[str]]] = None, ) -> Union[AsyncQuery, AsyncHybridQuery, AsyncFTSQuery, AsyncVectorQuery]: """Create a search query to find the nearest neighbors of the given query vector. We currently support [vector search](https://lancedb.com/docs/search/vector-search/) and [full-text search](https://lancedb.com/docs/search/full-text-search/). All query options are defined in [AsyncQuery][lancedb.query.AsyncQuery]. Parameters ---------- query: list/np.ndarray/str/PIL.Image.Image, default None The targeted vector to search for. - *default None*. Acceptable types are: list, np.ndarray, PIL.Image.Image - If None then a plain [AsyncQuery][lancedb.query.AsyncQuery] is returned, equivalent to calling [query][lancedb.table.AsyncTable.query]. The select/where/limit clauses are applied to filter the table. vector_column_name: str, optional The name of the vector column to search. The vector column needs to be a pyarrow fixed size list type - If not specified then the vector column is inferred from the table schema - If the table has multiple vector columns then the *vector_column_name* needs to be specified. Otherwise, an error is raised. query_type: str *default "auto"*. Acceptable types are: "vector", "fts", "hybrid", or "auto" - If "auto" then the query type is inferred from the query; - If `query` is a list/np.ndarray then the query type is "vector"; - If `query` is a PIL.Image.Image then either do vector search, or raise an error if no corresponding embedding function is found. - If `query` is a string, then the query type is "vector" if the table has embedding functions else the query type is "fts" Returns ------- AsyncQuery, AsyncHybridQuery, AsyncFTSQuery, or AsyncVectorQuery A query builder object representing the query. """ if query is None: return self.query() def is_embedding(query): return isinstance(query, (list, np.ndarray, pa.Array, pa.ChunkedArray)) async def get_embedding_func( vector_column_name: Optional[str], query_type: QueryType, query: Optional[Union[VEC, str, "PIL.Image.Image", Tuple, FullTextQuery]], ) -> Tuple[str, EmbeddingFunctionConfig]: if isinstance(query, FullTextQuery): query_type = "fts" schema = await self.schema() vector_column_name = infer_vector_column_name( schema=schema, query_type=query_type, query=query, vector_column_name=vector_column_name, ) funcs = EmbeddingFunctionRegistry.get_instance().parse_functions( schema.metadata ) func = funcs.get(vector_column_name) if func is None: error = ValueError( f"Column '{vector_column_name}' has no registered " "embedding function." ) if len(funcs) > 0: add_note( error, "Embedding functions are registered for columns: " f"{list(funcs.keys())}", ) else: add_note( error, "No embedding functions are registered for any columns." ) raise error return vector_column_name, func async def make_embedding(embedding, query): if embedding is not None: loop = asyncio.get_running_loop() # This function is likely to block, since it either calls an expensive # function or makes an HTTP request to an embeddings REST API. Run it # on a dedicated executor so it can't starve the default executor that # other blocking I/O shares. See # https://github.com/lancedb/lancedb/issues/3310. return ( await loop.run_in_executor( embedding_executor(), embedding.function.compute_query_embeddings_with_retry, query, ) )[0] else: return None if query_type == "auto": # Infer the query type. if is_embedding(query): vector_query = query query_type = "vector" elif isinstance(query, FullTextQuery): query_type = "fts" elif isinstance(query, str): try: ( indices, (vector_column_name, embedding_conf), ) = await asyncio.gather( self.list_indices(), get_embedding_func(vector_column_name, "auto", query), ) except ValueError as e: if "Column" in str( e ) and "has no registered embedding function" in str(e): # If the column has no registered embedding function, # then it's an FTS query. query_type = "fts" else: raise e else: if embedding_conf is not None: vector_query = await make_embedding(embedding_conf, query) if any( i.columns[0] == embedding_conf.source_column and i.index_type == "FTS" for i in indices ): query_type = "hybrid" else: query_type = "vector" else: query_type = "fts" else: # it's an image or something else embeddable. query_type = "vector" elif query_type == "vector": if is_embedding(query): vector_query = query else: vector_column_name, embedding_conf = await get_embedding_func( vector_column_name, query_type, query ) vector_query = await make_embedding(embedding_conf, query) elif query_type == "hybrid": if is_embedding(query): raise ValueError("Hybrid search requires a text query") else: vector_column_name, embedding_conf = await get_embedding_func( vector_column_name, query_type, query ) vector_query = await make_embedding(embedding_conf, query) if query_type == "vector": builder = self.query().nearest_to(vector_query) if vector_column_name: builder = builder.column(vector_column_name) return builder elif query_type == "fts": return self.query().nearest_to_text(query, columns=fts_columns) elif query_type == "hybrid": builder = self.query().nearest_to(vector_query) if vector_column_name: builder = builder.column(vector_column_name) return builder.nearest_to_text(query, columns=fts_columns) else: raise ValueError(f"Unknown query type: '{query_type}'") def vector_search( self, query_vector: Union[VEC, Tuple], ) -> AsyncVectorQuery: """ Search the table with a given query vector. This is a convenience method for preparing a vector query and is the same thing as calling `nearestTo` on the builder returned by `query`. Seer [nearest_to][lancedb.query.AsyncQuery.nearest_to] for more details. """ return self.query().nearest_to(query_vector) def _sync_query_to_async( self, query: Query ) -> ( AsyncHybridQuery | AsyncFTSQuery | AsyncVectorQuery | AsyncQuery | AsyncTakeQuery ): if query.take_offsets is not None: take_query = self.take_offsets(query.take_offsets) if query.columns: take_query = take_query.select(query.columns) if query.use_rte is not None: take_query = take_query.use_rte(query.use_rte) if query.with_row_id: take_query = take_query.with_row_id() return take_query async_query = self.query() if query.limit is not None: async_query = async_query.limit(query.limit) if query.offset is not None: async_query = async_query.offset(query.offset) if query.columns: async_query = async_query.select(query.columns) if query.filter is not None: async_query = async_query.where(query.filter) if query.fast_search: async_query = async_query.fast_search() if query.use_rte is not None: async_query = async_query.use_rte(query.use_rte) if query.with_row_id: async_query = async_query.with_row_id() if query.order_by: async_query = async_query.order_by(query.order_by) if query.vector: async_query = async_query.nearest_to(query.vector).distance_range( query.lower_bound, query.upper_bound ) if query.distance_type is not None: async_query = async_query.distance_type(query.distance_type) if query.nprobes is not None: async_query = async_query.nprobes(query.nprobes) if query.search_effort is not None: async_query = async_query.search_effort(query.search_effort) if query.minimum_nprobes is not None and query.maximum_nprobes is not None: async_query = async_query.minimum_nprobes( query.minimum_nprobes ).maximum_nprobes(query.maximum_nprobes) elif query.minimum_nprobes is not None: async_query = async_query.minimum_nprobes(query.minimum_nprobes) elif query.maximum_nprobes is not None: async_query = async_query.maximum_nprobes(query.maximum_nprobes) if query.refine_factor is not None: async_query = async_query.refine_factor(query.refine_factor) if query.vector_column: async_query = async_query.column(query.vector_column) if query.ef: async_query = async_query.ef(query.ef) if query.bypass_vector_index: async_query = async_query.bypass_vector_index() if query.postfilter: async_query = async_query.postfilter() if query.full_text_query: async_query = async_query.nearest_to_text( query.full_text_query.query, query.full_text_query.columns ) return async_query async def _execute_query( self, query: Query, *, batch_size: Optional[int] = None, timeout: Optional[timedelta] = None, ) -> pa.RecordBatchReader: if ( _should_push_down_query_table( self._namespace_client, self._pushdown_operations ) and not self._route_pushdown_to_rust and query.take_offsets is None and query.search_effort is None ): from lancedb.namespace import _execute_server_side_query table_id = self._namespace_path + [self.name] return _execute_server_side_query(self._namespace_client, table_id, query) # The sync table calls into this method, so we need to map the # query to the async version of the query and run that here. This is only # used for that code path right now. async_query = self._sync_query_to_async(query) return await async_query.to_batches( max_batch_length=batch_size, timeout=timeout ) async def _explain_plan(self, query: Query, verbose: Optional[bool]) -> str: # This method is used by the sync table async_query = self._sync_query_to_async(query) return await async_query.explain_plan(verbose) async def _analyze_plan( self, query: Query, *, distributed_metrics: AnalyzePlanDistributedMetrics = "aggregate", ) -> str: # This method is used by the sync table async_query = self._sync_query_to_async(query) return await async_query.analyze_plan(distributed_metrics) async def _output_schema(self, query: Query) -> pa.Schema: async_query = self._sync_query_to_async(query) return await async_query.output_schema() async def _do_merge( self, merge: LanceMergeInsertBuilder, new_data: DATA, on_bad_vectors: OnBadVectorsType, fill_value: float, ) -> MergeResult: schema = await self.schema() if on_bad_vectors is None: on_bad_vectors = "error" if fill_value is None: fill_value = 0.0 data = _sanitize_data( new_data, schema, metadata=schema.metadata, on_bad_vectors=on_bad_vectors, fill_value=fill_value, allow_subschema=True, ) if isinstance(data, pa.Table): data = pa.RecordBatchReader.from_batches(data.schema, data.to_batches()) return await self._inner.execute_merge_insert( data, dict( on=merge._on, when_matched_update_all=merge._when_matched_update_all, when_matched_update_all_condition=merge._when_matched_update_all_condition, when_not_matched_insert_all=merge._when_not_matched_insert_all, when_not_matched_by_source_delete=merge._when_not_matched_by_source_delete, when_not_matched_by_source_condition=merge._when_not_matched_by_source_condition, when_not_matched_by_source_condition_expr=merge._when_not_matched_by_source_condition_expr, timeout=merge._timeout, use_index=merge._use_index, use_rte=merge._use_rte, validate_single_shard=merge._validate_single_shard, ), ) async def delete(self, where: Union[str, Expr]) -> DeleteResult: """Delete rows from the table. This can be used to delete a single row, many rows, all rows, or sometimes no rows (if your predicate matches nothing). Parameters ---------- where: str or :class:`~lancedb.expr.Expr` The filter condition. Can be a SQL string or a type-safe :class:`~lancedb.expr.Expr` built with :func:`~lancedb.expr.col` and :func:`~lancedb.expr.lit`. The filter must not be empty, or it will error. Examples -------- >>> import lancedb >>> data = [ ... {"x": 1, "vector": [1.0, 2]}, ... {"x": 2, "vector": [3.0, 4]}, ... {"x": 3, "vector": [5.0, 6]} ... ] >>> db = lancedb.connect("./.lancedb") >>> table = db.create_table("my_table", data) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 2 [3.0, 4.0] 2 3 [5.0, 6.0] >>> table.delete("x = 2") DeleteResult(num_deleted_rows=1, version=2) >>> table.to_pandas() x vector 0 1 [1.0, 2.0] 1 3 [5.0, 6.0] If you have a list of values to delete, you can combine them into a stringified list and use the `IN` operator: >>> to_remove = [1, 5] >>> to_remove = ", ".join([str(v) for v in to_remove]) >>> to_remove '1, 5' >>> table.delete(f"x IN ({to_remove})") DeleteResult(num_deleted_rows=1, version=3) >>> table.to_pandas() x vector 0 3 [5.0, 6.0] """ predicate = where._inner if isinstance(where, Expr) else where return await self._inner.delete(predicate) async def update( self, updates: Optional[Dict[str, Any]] = None, *, where: Optional[Union[str, Expr]] = None, updates_sql: Optional[Dict[str, str]] = None, ) -> UpdateResult: """ This can be used to update zero to all rows in the table. If a filter is provided with `where` then only rows matching the filter will be updated. Otherwise all rows will be updated. Parameters ---------- updates: dict, optional The updates to apply. The keys should be the name of the column to update. The values should be the new values to assign. This is required unless updates_sql is supplied. where: str or [Expr][lancedb.expr.Expr], optional The filter condition. Can be a SQL string or a type-safe [Expr][lancedb.expr.Expr] built with [col][lancedb.expr.col] and [lit][lancedb.expr.lit]. Only rows that satisfy this filter will be updated. updates_sql: dict, optional The updates to apply, expressed as SQL expression strings. The keys should be column names. The values should be SQL expressions. These can be SQL literals (e.g. "7" or "'foo'") or they can be expressions based on the previous value of the row (e.g. "x + 1" to increment the x column by 1) Returns ------- UpdateResult An object containing: - rows_updated: The number of rows that were updated - version: The new version number of the table after the update Examples -------- >>> import asyncio >>> import lancedb >>> from lancedb.expr import col >>> import pandas as pd >>> async def demo_update(): ... data = pd.DataFrame({"x": [1, 2], "vector": [[1, 2], [3, 4]]}) ... db = await lancedb.connect_async("./.lancedb") ... table = await db.create_table("my_table", data) ... # x is [1, 2], vector is [[1, 2], [3, 4]] ... await table.update({"vector": [10, 10]}, where=col("x") == 2) ... # x is [1, 2], vector is [[1, 2], [10, 10]] ... await table.update(updates_sql={"x": "x + 1"}) ... # x is [2, 3], vector is [[1, 2], [10, 10]] >>> asyncio.run(demo_update()) """ if updates is not None and updates_sql is not None: raise ValueError("Only one of updates or updates_sql can be provided") if updates is None and updates_sql is None: raise ValueError("Either updates or updates_sql must be provided") if updates is not None: updates_sql = {k: value_to_sql(v) for k, v in updates.items()} predicate = where.to_sql() if isinstance(where, Expr) else where return await self._inner.update(updates_sql, predicate) async def add_columns( self, transforms: dict[str, str | FunctionApplication] | FunctionApplication | pa.Field | List[pa.Field] | pa.Schema | None = None, *, computed: dict[str, str] | None = None, auto_refresh_enabled: bool | None = None, auto_refresh_interval_seconds: int | None = None, ) -> AddColumnsResult: """ Add new columns with defined values. Parameters ---------- transforms: Dict[str, str | FunctionApplication] or FunctionApplication A map of column name to a SQL expression to use to calculate the value of the new column. These expressions will be evaluated for each row in the table, and can reference existing columns. Alternatively, you can pass a pyarrow field or schema to add new columns with NULLs. A mapping with one ``FunctionApplication`` value keeps its scalar or named-struct result in the named table column. A bare named-struct application expands its ordered result fields as one atomic binding; aliases come from ``rename(columns=...)``. Function columns are supported only on LanceDB Cloud and Enterprise. computed: Dict[str, str], optional Local declaration and refresh require a source build with the ``experimental-functions`` Cargo feature, which published packages do not enable. Remote Cloud/Enterprise operations remain available. A mapping from output column names to SQL expressions derives each output field from its expression. A direct projection of a Blob v2 field inherits Blob v2 semantics; other expressions derive their ordinary Arrow type. Mapping order is declaration and dependency order. Unlike ``transforms``, the expression is stored rather than evaluated now: the column is committed with no values, and rows get them from [`refresh_column`][lancedb.table.AsyncTable.refresh_column]. A refresh also recomputes the rows whose inputs changed since they were computed, so a mutated input is reflected by the next refresh. While a declaration reads a column, that column cannot be renamed, retyped or dropped. On LanceDB Cloud and Enterprise the expression is planned by the server. Cannot be combined with ``transforms``. auto_refresh_enabled: bool, optional Enable automatic refresh for each computed output declared by this call. Defaults to True. Local tables store the setting; automatic scheduling requires Cloud or Enterprise. auto_refresh_interval_seconds: int, optional Positive minimum seconds after a successful refresh before another automatic refresh is eligible. When omitted, the server determines when to refresh. Unchanged inputs do not force recomputation. Applies to every declared computed output. These options cannot be used with one-time transforms or ordinary all-null columns. Examples -------- >>> await table.add_columns( # doctest: +SKIP ... computed={"total": "price * quantity"}, ... auto_refresh_enabled=True, ... auto_refresh_interval_seconds=300, ... ) Returns ------- AddColumnsResult version: the new version number of the table after adding columns. """ if auto_refresh_enabled is not None and type(auto_refresh_enabled) is not bool: raise ValueError("auto_refresh_enabled must be a boolean") if auto_refresh_interval_seconds is not None and ( type(auto_refresh_interval_seconds) is not int or not 0 < auto_refresh_interval_seconds <= (2**64 - 1) // 1000 ): raise ValueError( "auto_refresh_interval_seconds must be a positive integer " "no greater than u64::MAX / 1000" ) options_supplied = ( auto_refresh_enabled is not None or auto_refresh_interval_seconds is not None ) function_application = None function_output_name = None if isinstance(transforms, FunctionApplication): function_application = transforms elif isinstance(transforms, dict) and any( isinstance(value, FunctionApplication) for value in transforms.values() ): if len(transforms) != 1 or not all( isinstance(value, FunctionApplication) for value in transforms.values() ): raise ValueError( "one add_columns call declares exactly one Function binding" ) function_output_name, function_application = next(iter(transforms.items())) if function_application is not None: if computed: raise ValueError( "add_columns cannot mix a Function application with SQL " "computed columns" ) function_application._ensure_declarable() return await self._inner.add_function_columns( function_application.to_canonical_json(), function_output_name, auto_refresh_enabled, auto_refresh_interval_seconds, ) if isinstance(transforms, pa.Field): transforms = [transforms] if isinstance(transforms, list) and all( {isinstance(f, pa.Field) for f in transforms} ): transforms = pa.schema(transforms) if computed: if transforms: raise ValueError( "add_columns cannot take both transforms and computed columns" ) return await self._inner.add_computed_columns( list(computed.items()), auto_refresh_enabled, auto_refresh_interval_seconds, ) if options_supplied: raise ValueError( "automatic refresh options require computed or Function columns" ) if transforms is None: raise ValueError("add_columns requires transforms or computed columns") if isinstance(transforms, pa.Schema): return await self._inner.add_columns_with_schema(transforms) else: return await self._inner.add_columns(list(transforms.items())) async def refresh_column(self, column: str) -> RefreshColumnResult: """ Fill the rows of a computed column that hold no value yet. Local refresh requires the ``experimental-functions`` Cargo feature in a source build. Published packages leave it disabled. Declared with ``add_columns(computed=...)``, a column starts empty and gets its values here. Rows appended since the last refresh are filled by the next one, and rows whose inputs changed since they were computed are recomputed; everything else is left as it is. Local tables only: a remote refresh runs as a server job, through [`refresh_column_async`][lancedb.table.Table.refresh_column_async]. Parameters ---------- column: str The name of the computed column to fill. Returns ------- RefreshColumnResult The number of rows filled and the new version of the table. """ return await self._inner.refresh_column(column) async def function_errors( self, job_id: Optional[str] = None, column: Optional[str] = None, limit: Optional[int] = None, ) -> "FunctionErrors": """ The per-row errors Function refreshes recorded on this table. A refresh running under a skip policy records each row it skipped with the input that failed and the error. This lists those records, newest job first, plus a summary for any fragment whose per-row detail was capped. LanceDB Cloud and Enterprise only; reading errors needs read access to the table, since a message carries the value that failed. Parameters ---------- job_id: str, optional Only errors recorded by this job. column: str, optional Only errors on this column. limit: int, optional At most this many records (server default 10000, cap 100000). Returns ------- FunctionErrors ``records``, ``fragments`` and ``truncated``, the last saying whether the listing stopped at its limit. Examples -------- >>> errors = await table.function_errors(column="embedding") # doctest: +SKIP >>> for record in errors.records: # doctest: +SKIP ... print(record.job_id, record.row_offset, record.error_message) >>> if errors.truncated: # doctest: +SKIP ... print("listing stopped at the limit") """ return await self._inner.function_errors( job_id=job_id, column=column, limit=limit ) async def refresh_column_async( self, column: str ) -> AsyncJob[RefreshColumnJobResult]: """ Like :meth:`refresh_column`, but returns a handle to the refresh job instead of blocking until it completes. Local refresh requires a source build with ``experimental-functions``. Remote Cloud/Enterprise jobs do not require this feature. The job may already be complete when returned; callers must not assume the column is filled until :meth:`AsyncJob.wait` resolves. Invalid input -- an unknown column, or one that is not computed -- raises here rather than failing the job. On local tables the job runs in-process; on LanceDB Cloud and Enterprise it is the server's backfill job. Returns ------- AsyncJob[RefreshColumnResult] A job whose successful ``wait`` returns row counts plus the source and published table versions. Examples -------- >>> import asyncio >>> import lancedb >>> async def refresh_in_background(): ... db = await lancedb.connect_async("./.lancedb") ... table = await db.create_table("computed_job_async_demo", [{"x": 1}]) ... await table.add_columns(computed={"doubled": "x * 2"}) ... job = await table.refresh_column_async("doubled") ... result = await job.wait() ... assert result.rows_assigned == 1 ... return await job.status() >>> asyncio.run(refresh_in_background()) # doctest: +SKIP 'finished' """ return _typed_job( await self._inner.refresh_column_async(column), RefreshColumnJobResult.from_json, ) async def alter_columns( self, *alterations: Iterable[dict[str, Any]] ) -> AlterColumnsResult: """ Alter column names and nullability. alterations : Iterable[Dict[str, Any]] A sequence of dictionaries, each with the following keys: - "path": str The column path to alter. For a top-level column, this is the name. For a nested column, this is the dot-separated path, e.g. "a.b.c". - "rename": str, optional The new name of the column. If not specified, the column name is not changed. - "data_type": pyarrow.DataType, optional The new data type of the column. Existing values will be casted to this type. If not specified, the column data type is not changed. - "nullable": bool, optional Whether the column should be nullable. If not specified, the column nullability is not changed. Only non-nullable columns can be changed to nullable. Currently, you cannot change a nullable column to non-nullable. Each alteration must specify at least one of "rename", "data_type", or "nullable". The legacy key "name" is also accepted for renaming. Unknown keys raise ValueError before any alterations are applied. Returns ------- AlterColumnsResult version: the new version number of the table after the alteration. """ return await self._inner.alter_columns(alterations) async def update_field_metadata( self, *updates: dict[str, Any] ) -> UpdateFieldMetadataResult: """Update per-field metadata. See [`Table.update_field_metadata`][lancedb.table.Table.update_field_metadata].""" return await self._inner.update_field_metadata(updates) async def drop_columns(self, columns: Iterable[str]): """ Drop columns from the table. Parameters ---------- columns : Iterable[str] The names of the columns to drop. """ return await self._inner.drop_columns(columns) async def version(self) -> int: """ Retrieve the version of the table LanceDb supports versioning. Every operation that modifies the table increases version. As long as a version hasn't been deleted you can `[Self::checkout]` that version to view the data at that point. In addition, you can `[Self::restore]` the version to replace the current table with a previous version. """ return await self._inner.version() async def list_versions(self): """ List all versions of the table """ versions = await self._inner.list_versions() for v in versions: # Use integer math: float division on ~1e18 nanosecond # values loses sub-millisecond precision. seconds, nanos = divmod(v["timestamp"], 1_000_000_000) v["timestamp"] = datetime.fromtimestamp(seconds) + timedelta( microseconds=nanos // 1000 ) return versions async def checkout(self, version: int | str): """ Checks out a specific version of the Table Any read operation on the table will now access the data at the checked out version. As a consequence, calling this method will disable any read consistency interval that was previously set. This is a read-only operation that turns the table into a sort of "view" or "detached head". Other table instances will not be affected. To make the change permanent you can use the `[Self::restore]` method. Any operation that modifies the table will fail while the table is in a checked out state. To return the table to a normal state use `[Self::checkout_latest]`. Parameters ---------- version: int | str, The version to check out. A version number (`int`) or a tag (`str`) can be provided. """ try: await self._inner.checkout(version) except RuntimeError as e: if "not found" in str(e): raise ValueError( f"Version {version} no longer exists. Was it cleaned up?" ) else: raise async def checkout_latest(self): """ Ensures the table is pointing at the latest version This can be used to manually update a table when the read_consistency_interval is None It can also be used to undo a `[Self::checkout]` operation """ await self._inner.checkout_latest() async def restore(self, version: Optional[int | str] = None): """ Restore the table to the currently checked out version This operation will fail if checkout has not been called previously This operation will overwrite the latest version of the table with a previous version. Any changes made since the checked out version will no longer be visible. Once the operation concludes the table will no longer be in a checked out state and the read_consistency_interval, if any, will apply. """ await self._inner.restore(version) def take_offsets( self, offsets: list[int], *, with_row_id: bool = False ) -> AsyncTakeQuery: """ Take a list of offsets from the table. Offsets are 0-indexed and relative to the current version of the table. Offsets are not stable. A row with an offset of N may have a different offset in a different version of the table (e.g. if an earlier row is deleted). Offsets are mostly useful for sampling as the set of all valid offsets is easily known in advance to be [0, len(table)). No guarantees are made regarding the order in which results are returned. Repeated offsets produce repeated rows. Parameters ---------- offsets: list[int] The offsets to take. with_row_id: bool, default False If True, include the "_rowid" column in the results. Returns ------- AsyncTakeQuery A query object that can be executed to get the rows at the given offsets. """ query = AsyncTakeQuery(self._inner.take_offsets(offsets), self) if with_row_id: query.with_row_id() return query def take_row_ids( self, row_ids: list[int], *, with_row_id: bool = False ) -> AsyncTakeQuery: """ Take a list of row ids from the table. Row ids are not stable and are relative to the current version of the table. They can change due to compaction and updates. Unlike offsets, row ids are not 0-indexed and no assumptions should be made about the possible range of row ids. In order to use this method you must first obtain the row ids by scanning or searching the table. Even so, row ids are more stable than offsets and can be useful in some situations. There is an ongoing effort to make row ids stable which is tracked at https://github.com/lancedb/lancedb/issues/1120 Parameters ---------- row_ids: list[int] The row ids to take. with_row_id: bool, default False If True, include the "_rowid" column in the results. Returns ------- AsyncTakeQuery A query object that can be executed to get the rows. """ query = AsyncTakeQuery(self._inner.take_row_ids(row_ids), self) if with_row_id: query.with_row_id() return query async def blob_columns(self) -> list[str]: return await self._inner.blob_columns() async def fetch_blobs( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> pa.LargeBinaryArray: return await self._inner.fetch_blobs( column, _normalize_blob_row_ids(row_ids, column) ) async def fetch_blob_ranges( self, column: str, requests: Sequence[Sequence[int]], ) -> pa.LargeBinaryArray: return await self._inner.fetch_blob_ranges(column, list(requests)) async def fetch_blob_files( self, column: str, row_ids: Union[list[int], pa.Array, pa.ChunkedArray, pa.Table], ) -> "list[Optional[BlobFile]]": handles = await self._inner.fetch_blob_files( column, _normalize_blob_row_ids(row_ids, column) ) return _wrap_blob_files(handles) async def open_blob_file( self, descriptor: Union[dict, pa.StructScalar, None] ) -> Optional[BlobFile]: """Open a snapshot-bound descriptor as a seekable file. See [Table.open_blob_file][lancedb.table.Table.open_blob_file]. """ descriptor = _normalize_blob_descriptor(descriptor) if descriptor is None: return None return BlobFile(await self._inner.open_blob_file(json.dumps(descriptor))) async def open_blob_files( self, descriptors: Iterable ) -> AsyncIterator[Optional[BlobFile]]: """Asynchronously iterate over files in descriptor order, preserving nulls. Use ``async for file in table.open_blob_files(descriptors)``. Each non-null file must be closed after use. ```python async for video in videos.open_blob_files(rows["video"]): if video is not None: with video: payload = await video.aread() ``` """ for descriptor in descriptors: yield await self.open_blob_file(descriptor) @property def tags(self) -> AsyncTags: """Tag management for the dataset. Similar to Git, tags are a way to add metadata to a specific version of the dataset. .. warning:: Tagged versions are exempted from the :py:meth:`optimize(cleanup_older_than)` process. To remove a version that has been tagged, you must first :py:meth:`~Tags.delete` the associated tag. """ return AsyncTags(self._inner) @property def branches(self) -> AsyncBranches: """Branch management for the table. Branches are isolated, writable lines of history forked from another branch (or version). Writes on a branch do not affect ``main``. """ return AsyncBranches(self._inner) def current_branch(self) -> Optional[str]: """The branch this table handle is scoped to, or ``None`` for ``main``.""" return self._inner.current_branch() async def optimize( self, *, cleanup_older_than: Optional[timedelta] = None, delete_unverified: bool = False, retrain=False, compaction_options: Optional[CompactionOptions] = None, ) -> OptimizeStats: """ Optimize the on-disk data and indices for better performance. Modeled after ``VACUUM`` in PostgreSQL. Optimization covers three operations: * Compaction: Merges small files into larger ones * Prune: Removes old versions of the dataset * Index: Optimizes the indices, adding new data to existing indices Parameters ---------- cleanup_older_than: timedelta, optional default 7 days All files belonging to versions older than this will be removed. Set to 0 days to remove all versions except the latest. The latest version is never removed. .. warning:: Setting this to 0 deletes the data files of every older version, so any other reader or writer still using an older version of the table will fail. Only set it to 0 if no other process is working on this dataset. delete_unverified: bool, default False Files leftover from a failed transaction may appear to be part of an in-progress operation (e.g. appending new data) and these files will not be deleted unless they are at least 7 days old. If delete_unverified is True then these files will be deleted regardless of their age. .. warning:: This should only be set to True if you can guarantee that no other process is currently working on this dataset. Otherwise the dataset could be put into a corrupted state. retrain: bool, default False This parameter is no longer used and is deprecated. compaction_options: CompactionOptions, optional Options that control file compaction. For large rows, derive a lower ``target_rows_per_fragment`` from the expected row size to bound later passes. This does not cap the first pass over an existing oversized fragment; see [CompactionOptions][lancedb.table.CompactionOptions]. Notes ----- The frequency an application should call optimize is based on the frequency of data modifications. If data is frequently added, deleted, or updated then optimize should be run frequently. A good rule of thumb is to run optimize if you have added or modified 100,000 or more records or run more than 20 data modification operations. """ return await self._do_optimize( _optimize_cleanup_since_ms(cleanup_older_than, retrain), delete_unverified, compaction_options, ) async def _do_optimize( self, cleanup_since_ms: Optional[int], delete_unverified: bool, compaction_options: Optional[CompactionOptions] = None, ) -> OptimizeStats: return await self._inner.optimize( cleanup_since_ms=cleanup_since_ms, delete_unverified=delete_unverified, compaction_options=compaction_options, ) async def list_indices(self) -> Iterable[IndexConfig]: """ List all indices that have been created with Self::create_index """ return await self._inner.list_indices() async def tokenize( self, query: str, *, column: Optional[str] = None, index_name: Optional[str] = None, ) -> Iterable[FtsToken]: """ Tokenize a query using the tokenizer configured on an FTS index. Specify exactly one of ``column`` or ``index_name``. Model-backed tokenizers such as ``jieba/*`` and ``lindera/*`` are rebuilt in the client process from index metadata. For remote tables, this means the same tokenizer model files must also exist locally (under ``LANCE_LANGUAGE_MODEL_HOME``) and must match the files the server uses. A mismatch cannot be detected: the returned tokens may differ from what the server's index produces. If the model files are missing, an error naming the tokenizer is raised. """ return await self._inner.tokenize(query, column=column, index_name=index_name) async def index_stats(self, index_name: str) -> Optional[IndexStatistics]: """ Retrieve statistics about an index Parameters ---------- index_name: str The name of the index to retrieve statistics for Returns ------- IndexStatistics or None The statistics about the index. Returns None if the index does not exist. """ stats = await self._inner.index_stats(index_name) if stats is None: return None else: return IndexStatistics(**stats) async def uses_v2_manifest_paths(self) -> bool: """ Check if the table is using the new v2 manifest paths. Returns ------- bool True if the table is using the new v2 manifest paths, False otherwise. """ return await self._inner.uses_v2_manifest_paths() async def migrate_manifest_paths_v2(self): """ Migrate the manifest paths to the new format. This will update the manifest to use the new v2 format for paths. This function is idempotent, and can be run multiple times without changing the state of the object store. !!! danger This should not be run while other concurrent operations are happening. And it should also run until completion before resuming other operations. You can use [AsyncTable.uses_v2_manifest_paths][lancedb.table.AsyncTable.uses_v2_manifest_paths] to check if the table is already using the new path style. """ await self._inner.migrate_manifest_paths_v2() @deprecation.deprecated( deprecated_in="0.33.1", current_version=__version__, details="Use update_field_metadata() instead.", ) async def replace_field_metadata( self, field_name: str, new_metadata: dict[str, str] ): """ Replace the metadata of a field in the schema .. deprecated:: 0.33.1 Use :func:`update_field_metadata` instead. Parameters ---------- field_name: str The name of the field to replace the metadata for new_metadata: dict The new metadata to set """ await self._inner.replace_field_metadata(field_name, new_metadata) @dataclass class IndexStatistics: """ Statistics about an index. Attributes ---------- num_indexed_rows: int The number of rows that are covered by this index. num_unindexed_rows: int The number of rows that are not covered by this index. index_type: str The type of index that was created. distance_type: Optional[str] The distance type used by the index. num_indices: Optional[int] The number of parts the index is split into. """ num_indexed_rows: int num_unindexed_rows: int index_type: Literal[ "IVF_FLAT", "IVF_SQ", "IVF_PQ", "IVF_RQ", "IVF_HNSW_SQ", "IVF_HNSW_PQ", "IVF_HNSW_FLAT", "FTS", "BTREE", "BITMAP", "LABEL_LIST", ] distance_type: Optional[Literal["l2", "cosine", "dot"]] = None num_indices: Optional[int] = None # This exists for backwards compatibility with an older API, which returned # a dictionary instead of a class. def __getitem__(self, key): return getattr(self, key) @dataclass class TableStatistics: """ Statistics about a table and fragments. Attributes ---------- total_bytes: int The total size, in bytes, of the table's data files, index files, and overlay files. Read from the manifest, so this excludes deletion files and manifests. num_rows: int The total number of rows in the table. num_deleted_rows: Optional[int] The total number of rows marked as deleted across all fragments of the table. These rows are not counted in ``num_rows``, but still occupy space on disk until the table is compacted, so a large value here indicates that the table should be optimized. Fragments in which every row was deleted are dropped outright, so their rows are not counted here. ``None`` when the backend does not report deletion counts. num_indices: int The total number of indices in the table. fragment_stats: FragmentStatistics Statistics about fragments in the table. """ total_bytes: int num_rows: int num_deleted_rows: Optional[int] num_indices: int fragment_stats: FragmentStatistics @dataclass class FragmentStatistics: """ Statistics about fragments. Attributes ---------- num_fragments: int The total number of fragments in the table. num_small_fragments: int The total number of small fragments in the table. Small fragments have low row counts and may need to be compacted. lengths: FragmentSummaryStats Statistics about the number of rows in the table fragments. """ num_fragments: int num_small_fragments: int lengths: FragmentSummaryStats @dataclass class FragmentSummaryStats: """ Statistics about fragments sizes Attributes ---------- min: int The number of rows in the fragment with the fewest rows. max: int The number of rows in the fragment with the most rows. mean: int The mean number of rows in the fragments. p25: int The 25th percentile of number of rows in the fragments. p50: int The 50th percentile of number of rows in the fragments. p75: int The 75th percentile of number of rows in the fragments. p99: int The 99th percentile of number of rows in the fragments. """ min: int max: int mean: int p25: int p50: int p75: int p99: int class Tags: """ Table tag manager. """ def __init__(self, table): self._table = table def list(self) -> Dict[str, Tag]: """ List all table tags. Returns ------- dict[str, Tag] A dictionary mapping tag names to version numbers. """ return LOOP.run(self._table.tags.list()) def get_version(self, tag: str) -> int: """ Get the version of a tag. Parameters ---------- tag: str, The name of the tag to get the version for. """ return LOOP.run(self._table.tags.get_version(tag)) def create(self, tag: str, version: int) -> None: """ Create a tag for a given table version. Parameters ---------- tag: str, The name of the tag to create. This name must be unique among all tag names for the table. version: int, The table version to tag. """ LOOP.run(self._table.tags.create(tag, version)) def delete(self, tag: str) -> None: """ Delete tag from the table. Parameters ---------- tag: str, The name of the tag to delete. """ LOOP.run(self._table.tags.delete(tag)) def update(self, tag: str, version: int) -> None: """ Update tag to a new version. Parameters ---------- tag: str, The name of the tag to update. version: int, The new table version to tag. """ LOOP.run(self._table.tags.update(tag, version)) class Branches: """ Table branch manager. """ def __init__(self, parent: "LanceTable"): self._parent = parent self._table = parent._table def list(self) -> Dict[str, Any]: """List all branches, mapping name to branch metadata.""" return LOOP.run(self._table.branches.list()) def create( self, name: str, from_ref: Optional[str] = None, from_version: Optional[int] = None, ) -> "Table": """Create a branch and return a handle scoped to it. Parameters ---------- name: str Name of the new branch. from_ref: str, optional Source branch to fork from. Defaults to ``main``. from_version: int, optional A specific version on ``from_ref`` to fork from. Defaults to latest. """ async_table = LOOP.run( self._table.branches.create(name, from_ref, from_version) ) return self._wrap(async_table) def checkout(self, name: str, version: Optional[int] = None) -> "Table": """Check out an existing branch and return a handle scoped to it. Parameters ---------- name: str Name of the branch to check out. version: int, optional A specific version on the branch to pin. When set, the returned handle is a read-only view of that version; when omitted it tracks the branch's latest and stays writable. """ async_table = LOOP.run(self._table.branches.checkout(name, version)) return self._wrap(async_table, version) def delete(self, name: str) -> None: """Delete a branch.""" LOOP.run(self._table.branches.delete(name)) def diff(self, from_branch: str) -> Dict[str, Any]: """Diff a branch against main.""" return LOOP.run(self._table.branches.diff(from_branch)) def cherry_pick(self, from_branch: str, dry_run: bool = False) -> Dict[str, Any]: """Cherry-pick a branch onto main, or dry-run. Parameters ---------- from_branch: str Branch to cherry-pick from. dry_run: bool, default False When True, only preview. When False, attempt the cherry-pick. Notes ----- A failed cherry-pick returns ``status="failed"`` instead of raising. """ return LOOP.run(self._table.branches.cherry_pick(from_branch, dry_run)) def _wrap( self, async_table: "AsyncTable", version: Optional[int] = None ) -> "Table": # Delegate to the parent so the branch handle keeps its concrete type # (LanceTable / RemoteTable) and connection context; `version` is the # explicit pin so a remote handle can restore branch+version on reopen. return self._parent._wrap_branch_handle(async_table, version) class AsyncTags: """ Async table tag manager. """ def __init__(self, table): self._table = table async def list(self) -> Dict[str, Tag]: """ List all table tags. Returns ------- dict[str, Tag] A dictionary mapping tag names to version numbers. """ return await self._table.tags.list() async def get_version(self, tag: str) -> int: """ Get the version of a tag. Parameters ---------- tag: str, The name of the tag to get the version for. """ return await self._table.tags.get_version(tag) async def create(self, tag: str, version: int) -> None: """ Create a tag for a given table version. Parameters ---------- tag: str, The name of the tag to create. This name must be unique among all tag names for the table. version: int, The table version to tag. """ await self._table.tags.create(tag, version) async def delete(self, tag: str) -> None: """ Delete tag from the table. Parameters ---------- tag: str, The name of the tag to delete. """ await self._table.tags.delete(tag) async def update(self, tag: str, version: int) -> None: """ Update tag to a new version. Parameters ---------- tag: str, The name of the tag to update. version: int, The new table version to tag. """ await self._table.tags.update(tag, version) class AsyncBranches: """Async table branch manager.""" def __init__(self, table): self._table = table async def list(self) -> Dict[str, Any]: """List all branches, mapping name to branch metadata.""" return await self._table.branches.list() async def create( self, name: str, from_ref: Optional[str] = None, from_version: Optional[int] = None, ) -> "AsyncTable": """Create a branch and return a handle scoped to it. Parameters ---------- name: str Name of the new branch. from_ref: str, optional Source branch to fork from. Defaults to ``main``. from_version: int, optional A specific version on ``from_ref`` to fork from. Defaults to latest. """ # "main" and None are two spellings of the root branch in lance; normalize # so from_ref="main" behaves identically to the default. if from_ref == "main": from_ref = None inner = await self._table.branches.create(name, from_ref, from_version) return AsyncTable(inner) async def checkout(self, name: str, version: Optional[int] = None) -> "AsyncTable": """Check out an existing branch and return a handle scoped to it. Parameters ---------- name: str Name of the branch to check out. version: int, optional A specific version on the branch to pin. When set, the returned handle is a read-only view of that version; when omitted it tracks the branch's latest and stays writable. """ return AsyncTable(await self._table.branches.checkout(name, version)) async def delete(self, name: str) -> None: """Delete a branch.""" await self._table.branches.delete(name) async def diff(self, from_branch: str) -> Dict[str, Any]: """Diff a branch against main.""" return await self._table.branches.diff(from_branch) async def cherry_pick( self, from_branch: str, dry_run: bool = False ) -> Dict[str, Any]: """Cherry-pick a branch onto main, or dry-run. A failed cherry-pick returns ``status="failed"`` instead of raising. """ return await self._table.branches.cherry_pick(from_branch, dry_run)