tPj UdZddlmZddlmZdZeed<dedefdZ d edefd Z d edefd Z d edefdZ dee efdefdZy)z Single chokepoint for reading prompt-compression token savings out of a parsed SpendLog ``metadata`` JSON dict. Imported by the daily-spend DB writer and by cost-savings read endpoints. )Mapping)FinalheadroomHEADROOM_GUARDRAIL_PROVIDERvaluereturncrt|tst|ttfsy|dkryt|S)Nr) isinstanceboolintfloat)rs u/home/oem/CyberBrain26-unified/.venv/lib/python3.12/site-packages/litellm/proxy/spend_tracking/compression_savings.py_saved_tokens_or_zeror s0%je &E qy u:statscXt|tsyt|jdS)Nr tokens_saved)r rrget)rs r_tokens_saved_from_statsrs# eW % >!: ;;rentryct|tsy|jdtk7ryt |jdS)Nrguardrail_providerguardrail_response)r rrrr)rs r_headroom_entry_saved_tokensrs: eW % yy%&*EE #EII.B$C DDrguardrail_informationcrt|tr|gn|}t|tsytd|DS)Nrc32K|]}t|yw)N)r).0rs r z)_headroom_saved_tokens..'sHu+E2Hs)r rlistsum)rentriess r_headroom_saved_tokensr##s70:;PRY0Z+,`uG gt $ HH HHrmetadataclt|jdt|jdzS)a; Return the total prompt tokens saved by compression for one request. Sums two disjoint sources: - the native ``compression_savings`` key, written only by ``CompressionInterceptionLogger`` in its pre-call deployment hook - ``guardrail_information`` entries with ``guardrail_provider == "headroom"``, written only by the Headroom guardrail Each writer records only its own transform pass and the two run at different stages (guardrail pre-call vs deployment pre-call), so when both fire on one request their measured savings are independent and additive; summing them never double-counts. Malformed or missing values contribute 0. A bare dict ``guardrail_information`` is treated as a single entry, matching the spend-log redactor's normalization of that legacy shape. compression_savingsr)rrr#)r$s r extract_compression_saved_tokensr'*s6$ $HLL1F$G HKa ,-L rN)__doc__collections.abcrtypingrr__annotations__objectr rrrr#strr'rrr/s $%/U/C