uPjEUdZddlmZddlmZmZddlmZddlm Z m Z m Z ddl Z ddl mZmZmZddlmZdd lmZmZmZdd lmZe rdd lmZd Zd ed<dZd ed<dZd ed<dZ d ed<dZ!d ed<dZ"d ed<dZ#d ed<dZ$d ed<dZ%d ed<dZ&d ed <Gd!d"eZ'Gd#d$eZ(Gd%d&eZ)Gd'd(eZ*Gd)d*eZ+Gd+d,eZ,Gd-d.eZ-dEd/Z.dFd0Z/dGd1Z0dHd2Z1dId3Z2dJd4Z3Gd5d6eZ4Gd7d8eZ5Gd9d:eZ6Gd;deZ8Gd?d@eZ9dKdAZ:dLdBZ;GdCdDeZ.services.ai.azure.com/api/projects/ 2. Set BING_GROUNDING_MODEL to a model deployment in that project (e.g. gpt-4.1); it runs the grounded search and its tokens are billed on that deployment 3. Optional: set BING_GROUNDING_CONNECTION_ID to a Grounding with Bing Search project connection id to use the `bing_grounding` tool; without it the project's built-in `web_search` tool is used 4. Auth: pass api_key (an Azure API key, sent in the api-key header), or set BING_GROUNDING_TOKEN to an Entra bearer token for scope https://ai.azure.com/.default, or configure azure-identity (AZURE_CLIENT_ID / AZURE_CLIENT_SECRET / AZURE_TENANT_ID, managed identity, or any DefaultAzureCredential source) and the token is minted automatically Usage: response = litellm.search( query="latest AI developments", search_provider="bing_grounding", max_results=5, ) ) annotations)CallableMapping)MappingProxyType) TYPE_CHECKINGFinalLiteralN) BaseModel ConfigDictValidationError)BaseLLMException)BaseSearchConfigSearchResponse SearchResult)get_secret_str)LoggingzUhttps://learn.microsoft.com/en-us/azure/ai-foundry/agents/how-to/tools/bing-groundingr _DOCS_URLBING_GROUNDING_PROJECT_ENDPOINTPROJECT_ENDPOINT_ENVBING_GROUNDING_MODEL MODEL_ENVBING_GROUNDING_CONNECTION_IDCONNECTION_ID_ENVBING_GROUNDING_TOKEN TOKEN_ENVzhttps://ai.azure.com/.default ENTRA_SCOPEz/openai/v1/responses_RESPONSES_PATHi,_SNIPPET_FALLBACK_LENGTHi_UPSTREAM_ERROR_STATUSz$llm_provider-x-litellm-response-cost_RESPONSE_COST_HEADERcheZdZUeddZdZded<dZded <dZded <dZ d ed <dZ d ed <y) _AnnotationignoreTextrafrozenstrtypeN str | Noneurltitle int | None start_index end_index) __name__ __module__ __qualname__r model_configr)__annotations__r+r,r.r/m/home/oem/CyberBrain26-unified/.venv/lib/python3.12/site-packages/litellm/llms/azure/search/transformation.pyr"r"@s?HT:LD#NCE:"K" Iz r6r"cLeZdZUeddZdZded<dZded<dZd ed <y ) _ContentPartr#Tr$r'r(r)textr5ztuple[_Annotation, ...]rN) r0r1r2r r3r)r4r:rr5r6r7r9r9Js+HT:LD#ND#N+-K(-r6r9c>eZdZUeddZdZded<dZded <y ) _OutputItemr#Tr$r'r(r)r5ztuple[_ContentPart, ...]contentN)r0r1r2r r3r)r4r=r5r6r7r<r<Rs"HT:LD#N(*G %*r6r<c0eZdZUeddZdZded<y) _ErrorBodyr#Tr$Nr*message)r0r1r2r r3r@r4r5r6r7r?r?YsHT:LGZr6r?c0eZdZUeddZdZded<y)_IncompleteDetailsr#Tr$Nr*reason)r0r1r2r r3rCr4r5r6r7rBrB_sHT:LFJr6rBcZeZdZUdZeddZded<dZded <dZd ed <dZ d ed <y)_ResponsesEnvelopea7A Foundry Responses API body. `output` is required: a body without it is not a Responses API response and must not be reported as a successful empty search. A 200 body can still carry `status` `failed` or `incomplete`; those are surfaced as errors rather than reported as a successful empty search.r#Tr$ztuple[_OutputItem, ...]outputNr*status_ErrorBody | Noneerrorz_IncompleteDetails | Noneincomplete_details) r0r1r2__doc__r r3r4rGrIrJr5r6r7rErEes=A HT:L ##FJ#E #4818r6rEc0eZdZUeddZdZded<y)_ErrorEnveloper#Tr$NrHrI)r0r1r2r r3rIr4r5r6r7rMrMtsHT:L#E #r6rMc tj|}|jr|jjnd}||S t j|}|jxs|S#t$r|cYSwxYw#t$r|cYSwxYw)a' Surface the human-readable message inside Foundry's error envelope. Tool failures nest a second JSON document as a string inside `error.message` (observed live for `bing_grounding` connection errors), so the unwrap runs twice. Falls back to the raw body for anything else. N)rMmodel_validate_jsonr rIr@r?) error_messageenveloper@nesteds r7_unwrap_error_detailrSzs(<<]K08~~X^^++4G"66w? >> $W$  s#A%A6% A32A36 BBc|j}|d|cxkrt|krnn|n t|}|d|jdddj}|r |t dS|dtS)a  The text a citation supports, not the citation marker itself. A url_citation's start/end indices span the inline marker ("([host](url))"), which follows the claim it backs, so the snippet is the marker's own line up to where the marker starts. Nr )r.lenrsplitstripr)r: annotationstart marker_startclaims r7_snippetr_s}))E#(#4e9Ps4y9P%VYZ^V_L &--dA6r:@@BE ../00 )) **r6ctd|jD}tt|Dcic]}|j|c}tfdt j d|DDScc}w)zMOne result per cited URL: first occurrence wins, order preserved as answered.c 3RK|]}|jdk(r|jD]}|jdk(rn|jD]_}|jdk(rN|jrBt |j xsd|jxsdt |j|ddayw)r@ output_text url_citationr'N)r,r+snippetdate last_updated)r)r=rr+rr,r_r:).0itempartr[s r7 z$_citation_results..s  99 !LL  99 %**  ??n , ""(b$"TYY 3     sB%B'c3(K|] }| ywNr5)rgr+ first_by_urls r7rjz$_citation_results..s\sc"\sc34K|]}|jywrl)r+)rgresults r7rjz$_citation_results..s;[6FJJ;[s)tuplerFrreversedr+dictfromkeys)rQcitedrorms @r7_citation_resultsruslOOE+XV[_+]6FJJ,>+]^L \dmm;[UZ;[.[\ \\,^sA9cVt|tst|tsy|dkDr|SdS)zA positive-int `max_results`, else None. Rejects bools, an `int` subclass, and non-positive values so neither the request-side `count` nor the response-side cap forwards a value the other would silently ignore. Nr) isinstanceboolint) max_resultss r7_valid_max_resultsr{s+ +t$J{C,H%/;3t3r6cz|jd}t|tsyt|jdS)aThe unified `max_results` cap the caller asked for, if any. The built-in web_search tool has no server-side result-count knob, so the cap is enforced here after the fact; connection mode also honors it as a hard ceiling on top of the tool's `count` hint. optional_paramsNrz)getrwrr{)response_kwargsr}s r7_requested_max_resultsrs8-001BCO ow / o11-@ AAr6c||d|S|Srlr5)resultsrzs r7_cappedrs$/$;7eZdZUedZded<ded<ded<y) _ResponsesRequestTrr(modelinputz/tuple[_BingGroundingTool | _WebSearchTool, ...]toolsNrr5r6r7rrsT*L J J ::r6rc2tt}|jd}|jd}|r-t|t |}t t |fSt|trt|jnd}t|S)Nrzr)rr)r)r)r)r) rrr~rr{rrrwr(rupperr)r} connection_idrzr configurationlocations r7 _search_toolrs)*;.services.ai.azure.com/api/projects/./)rrrrstripendswithr)rrr}datar resolved_basetrimmeds r7get_complete_urlz*BingGroundingSearchConfig.get_complete_urlJsk (O>:N+O '()ab '--c2   O ,N?+,,r6c tt}|sttdt|t |t rdj |n|t|f}|jddS)az Transform Search request to the Foundry Responses API format. The unified params map as far as the API allows: - max_results -> the bing_grounding search configuration's `count`; the built-in web_search tool has no result-count knob, so that mode instead caps the returned results after the fact (see transform_search_response) - country -> web_search's approximate `user_location` (bing_grounding's `market` wants a full locale like en-US, which a bare country code cannot fill) - search_domain_filter, max_tokens_per_page -> no API equivalent, dropped zm is not set. Set it to a model deployment in the Foundry project that runs the grounded search, e.g. gpt-4.1. )rrrjsonT)mode exclude_none) rrrrrwlistjoinr model_dump)rqueryr}rrrequests r7transform_search_requestz2BingGroundingSearchConfig.transform_search_request\sv"&i0+?@ +%/t%<#((5/%02  !!vD!AAr6c  tj|j}|jdk(rL|jr,|jjr|jjnd}|j||tt|t|}|s^|jdk(rO|jr,|jj r|jj nd}|jd|||j#|S#t$r9}|j d||j t |jd}~wwxYw)Nz:response does not match the Foundry Responses API schema: rP status_coderfailedzthe grounded search failed incompletezunknown reasonz$the grounded search was incomplete: )rErOr=r get_error_classrrrrrGrIr@_upstream_errorrrurrJrC_priced) r raw_response logging_objrparsedrdetailrrCs r7transform_search_responsez3BingGroundingSearchConfig.transform_search_responsezs:  .BB -];BB4H#26(& =Y  r6rl)rzCallable[[], str] | NonereturnNonerr()NN) rdict[str, str]rr*rr*rrrr)rr*rr*rzMapping[str, str])rr*rr() rr*r}dict[str, object]rz2dict[str, object] | list[dict[str, object]] | Nonerrrr()rzstr | list[str]r}rrrrr)rhttpx.ResponserLiteLLMLoggingObjrrrr)rr(rrrr)rtuple[SearchResult, ...]rr)rPr(rryrrrr)r0r1r2r staticmethodrrrrrrrrrr __classcell__)rs@r7rrs16,, ##          &F" &DH --+-A -  - -$BB+B B  B<%$%'% %  %: "          r6r)rPr(rr()r:r(r[r"rr()rQrErr)rzrrr-)rMapping[str, object]rr-)rrrzr-rr)r}rrz#_BingGroundingTool | _WebSearchToolr)=rK __future__rcollections.abcrrtypesrtypingrrr httpxpydanticr r r )litellm.llms.base_llm.chat.transformationr +litellm.llms.base_llm.search.transformationrrrlitellm.secret_managers.mainr*litellm.litellm_core_utils.litellm_loggingrrrr4rrrrrrrrr r"r9r<r?rBrErMrSr_rur{rrrrrrrrrrrr5r6r7r s|8#-"00 ;;F 8Wj 5j?e?) 5)959) 5)4 U4//"%%% ##EuE!)!.9.+)+  9 9$Y$ %, + ]*4 BI9<9< ))I/Y/; ; 2B l 0l r6