jlSrSSKJr SSKJr SSKJrJrJrJ r SSK J r J r SSK Jr SSKJr SSKJr SS KJr SS KJr SS KJr SS KJrJr SS KJr SSKJrJ r SSK!J"r" Sr#SSjr$Sr%Sr&"SS\ 5r'\"SS9S Sj5r(S!Sjr)S"S#Sjjr*S$Sjr+S"S%Sjjr,S"S&Sjjr-g)'acThe playground's Claude model factory, its answer-context format, and the prose baseline. `anthropic_model` decides which credential a playground agent runs on. Every Pydantic AI agent defined in this package builds its model through it: the answering agents, the selection agent in `selection.py`, and the planner in `answers/planned.py`. Two agents the playground also runs, the cover-page attributor at upload and the footnote lookup at ingest, are built in `quber.agents` and choose their credential there. `answer_model` fixes which Claude model answers, for the live answering agents in `answers/fixed.py` and `answers/declared.py` and for the prose agent here. `build_context` is the chunk-to-prompt format the answering agents share. The prose agent (`answer`, `GroundedAnswer`) answers in free text, strictly from the retrieved chunks, and returns the ids it grounded on. A table chunk's HTML carries an id on every cell (``), so it can cite individual cells. It is not on the playground's request path. `benchmark/run_shapes.py` uses `answer` as its baseline, and `answer_stream` has no caller. ) annotations) lru_cache)AnyDictListOptional) BaseModelField)Agent)AnthropicModel)AnthropicProvider) ModelSettings)make_oauth_anthropic_model)usage_metadata_from)DocumentIdentity with_document)RetrievedChunk) run_metadatatracer) get_settingsz claude-opus-5c[5RnUR(a[XR5$UR(a[ U[ URS9S9$[S5e)aAn Anthropic model over whichever credential the host presents. The subscription OAuth token wins when it is set, routed through the OAuth gate in `quber.agents._oauth_gate`. Without one the API key is used through the stock model, which is how the hosted playground runs: the token never leaves the developer host, and the task is given only the key. Neither set is a configuration error, reported as such. )api_key)providerz\Neither ANTHROPIC_AUTH_TOKEN nor ANTHROPIC_API_KEY is set; the playground needs one of them.)rllmanthropic_auth_tokenranthropic_api_keyr r RuntimeError)model_idrs 4/home/mande/repo/quber/src/quber/playground/agent.pyanthropic_modelr +s] .  C )(4L4LMM h1B3K`K`1abb f c [[5$)zThe playground's answering model. One place decides which model answers. The fixed and declared answering agents and the prose baseline below all build from here rather than each pinning its own copy of the name. )r DEFAULT_MODELr!r answer_modelr%>s = ))r!uYou are a financial-document analyst answering questions about ONE parsed document. You are given a set of context chunks retrieved from that document. Rules: - Answer ONLY from the provided context chunks. Never use outside knowledge and never invent figures. If the context does not contain the answer, say so plainly. - Be precise and quantitative. Report exact values, units, and periods as they appear in the source. - Populate `cited_ids` with the ids that support your answer. Each context chunk begins with a header line: `[chunk id= | page | type ]`. * By default, cite the `` from the header of each chunk you used. * If a chunk's content exposes finer-grained element ids — e.g. table cells tagged like `` — you MAY cite those specific cell ids instead, for tighter grounding. * Only cite ids that literally appear in the provided context (a header `` or an element id inside a chunk's content). NEVER invent, guess, or construct an id that is not present verbatim. cD\rSrSr%\"SS9rS\S'\"\SS9rS\S 'S r g ) GroundedAnswer^z/The answer, drawn only from the context chunks.) descriptionstranswerz8Chunk ids and/or table cell ids that support the answer.)default_factoryr)z List[str] cited_idsr$N) __name__ __module__ __qualname____firstlineno__r r+__annotations__listr-__static_attributes__r$r!rr'r'^s)$UVFCV NIyr!r')maxsizec H[[5[[[ SS9S9$)Ng) temperature) output_type system_promptmodel_settings)r r%r' SYSTEM_PROMPTrr$r!r_agentr=fs" "#$5  r!c/nUHPnSURSURS-SURS3nURUSUR35 MR SR U5$)Nz [chunk id=z | page r5z | type ] z )chunk_idpage chunk_typeappendcontentjoin)chunksblockscheaders r build_contextrKpsh F ajj\!&&1*Xall^STU 199+./ ;;v r!Nc@[U5nSUSUS3n[XB5$)NzContext chunks: z Question: z: Answer from the context above and cite the ids you used.)rKr)questionrGdocumentcontextprompts r_promptrQys;F#G gY'JC C  **r!cbS[S.SUS./UVs/sHo"RPM snS.$s snf)NsystemrolerEuser)messages chunk_ids)r<rA)rPrGrIs r _trace_inputsrYs>- 8 / +11&Qjj&1   2s,c# [XU5n[5RS[X15[[ 5S9IShvN n[ 5RU5IShvN nURnSUR5S./[UR5S.Ul SSS5IShvN U$NrNSN !,IShvN (df  W$=f7f)Nanswer_groundedmodelextra_metadata assistantrTrWusage_metadata) rQrllm_runrYr#rr=runoutputmodel_dump_jsonrusageoutputs)rMrGrNrPrcresultrds rr+r+sXx 0Fx=8 ^j^l  x||F++!'"-&:P:P:RST1&,,?   M, MsZACB5CB;#B7$?B;# C.B9/C7B;9C;CC  CCc# [XU5n[5RS[X15[[ 5S9IShvN n[ 5RU5IShvN nSnURSS9ShvN n[USS5=(d SnU(dM(X:wdM/UnSU47v M:NqNQN9 UR5IShvN n S U R5S ./[UR5S .UlSSS5IShvN O!,IShvN (df  O=fS W 47v SSS5IShvN g!,IShvN (df  g=f7f) zwYield ("delta", ) as the model generates, then ("final", GroundedAnswer) once the full output validates.r[r\Ng?) debounce_byr+deltar_rTr`final)rQrrbrYr#rr= run_stream stream_outputgetattr get_outputrerrfrg) rMrGrNrPrcrhlastpartialtextrms r answer_streamrus/ Xx 0Fx=8 ^j^l  8&&v..&D!'!5!5$!5!G (gw"5;4DLD!4-'/ (!G +1*;*;*=$=$=E&1e>S>S>UVW"5fll"CCK/.....unsAE"B4E"E$B6%E(D:B:>B8 ?B:D"D) D4E"6E8B::DC4D EDED, D D, (E6 E"EE"EE  EE")rr*)returnzAgent[None, GroundedAnswer])rGList[RetrievedChunk]rvr*)N)rMr*rGrwrNOptional[DocumentIdentity]rvr*)rPr*rGrwrvzDict[str, Any])rMr*rGrwrNrxrvr')rMr*rGrwrNrx).__doc__ __future__r functoolsrtypingrrrrpydanticr r pydantic_air pydantic_ai.models.anthropicr pydantic_ai.providers.anthropicr pydantic_ai.settingsrquber.agents._oauth_gaterquber.agents.langsmith_tracerr!quber.playground.answers.documentrrquber.playground.retrievalrquber.playground.tracingrrquber.settingsrr#r r%r<r'r=rKrQrYr+rur$r!rrs(#,,%7=.?=M59' &* ,Y 1+Y]  / ;U  "Y]/;Ur!