66&j]6SrSSKJr SSKrSSKrSSKJr SSKJrJ r J r J r SSK J r SSKJrJr SS KJr "S S \5rS \R("\R+5S S9S3r\ "SS\ 55rSSjr"SS5r"SS5r\SrSSSjjrg)u TableCorrespondence Protocol — does this Camelot chunk belong to this detected table? (QUE-230 rebuild, Step 2.) The detector (`agents.detector`) returns the real tables on a page in top-to-bottom order; Camelot returns chunks, also in page order. Order does most of the alignment work — the extractor walks both lists in sequence. This agent is the *content confirmation* the walk leans on: it judges a single (chunk, detected table) pairing and answers whether the chunk is the table shown there. It is deliberately per-pairing and narrow. The order-driven orchestration (which pairings to try, lattice-before-stream, split/combined handling) lives in the extractor, not here. Asking a content question — "is this chunk that table?" — is robust where comparing pixel coordinates is fragile. Kept separate from the other agents for the usual reasons: its own system prompt, output schema, retry handling, and an independently tunable model. ) annotationsN)Path)LiteralOptionalProtocolruntime_checkable)logger) BaseModelField) DetectedTablecD\rSrSr%\"SS9rS\S'\"SSS9rS \S 'S rg ) CorrespondenceVerdict$zTrue if the Camelot chunk is the content of the detected table shown on the page (the same physical table, even if the chunk is only part of it or has minor extraction noise). False if the chunk is a different table, non-table text, or unrelated content.) descriptionboolmatchesz`One short sentence justifying the decision; what content tied the chunk to the table, or didn't.)defaultrstrreasonN) __name__ __module__ __qualname____firstlineno__r r__annotations__r__static_attributes__r9/home/mande/repo/quber/src/quber/agents/correspondence.pyrr$s2 EGTvFCrruGYou are deciding whether one extracted table chunk corresponds to one specific table shown on a PDF page. You are given: 1. A rendered image of the PDF page (ground truth). 2. The target table: its position on the page (ordinal, top to bottom), a rough bounding box, and a description of what it is. 3. The chunk: markdown a table-extraction tool pulled from the page. Decide whether the chunk is the content of the target table. - *Matches* — the chunk is that physical table. It is a match even if the chunk holds only part of the table (a truncated top or bottom section), or has minor extraction noise, as long as its rows/columns clearly belong to the target table. - *Does not match* — the chunk is a different table on the page, non-table text (a footnote list, prose, a heading), or unrelated content. Judge by *content* — do the chunk's headers, row labels, and values match what the target table contains? Do not rely on coordinates; the bounding box only tells you which table on the page is the target. Return a JSON object conforming exactly to this schema: indentuE Return only the JSON object — no prose, no markdown code fences. c.\rSrSrSSjrSrg)TableCorrespondenceVc # g7fNr)self page_imagetargetchunk_markdowns r rTableCorrespondence.matchesXs !$srNr*rr+r r,rreturnr)rrrrrrrrr r%r%Vs-$$$ $  $rr%cH[R"UR5SS9$)Nr!r")jsondumps model_dump)r+s r _target_payloadr4`s ::f'')! 44rcX\rSrSrSrSrSS SjjrS SjrSrg) PydanticAICorrespondenceduTableCorrespondence backed by pydantic-ai with an Anthropic model. Sends the page image as `BinaryContent` alongside the target-table metadata and the chunk markdown. Auth resolution mirrors the other agents. Model defaults to Haiku 4.5. Correspondence is the safest fit for Haiku of the three QUE-230 agents — document order does most of the work and the agent only confirms and disambiguates. zclaude-haiku-4-5-20251001NcSSKJn SSKJn SSKJn U=(d2 [ RRS5=(d URnU=(d [ RRS5nU=(d [ RRS5nU(aSS K J n U "X5n OU(a U"US 9n U"XS 9n O [S 5eXl U"U [[S 9Ulg)Nr)Agent)AnthropicModel)AnthropicProviderQUBER_LLM_MODELANTHROPIC_AUTH_TOKENANTHROPIC_API_KEYr )make_oauth_anthropic_model)api_key)providerzxPydanticAICorrespondence: neither ANTHROPIC_AUTH_TOKEN nor ANTHROPIC_API_KEY is set. Provide one via env or constructor.) output_type system_prompt) pydantic_air9pydantic_ai.models.anthropicr:pydantic_ai.providers.anthropicr;osenvironget DEFAULT_MODEL _oauth_gater? RuntimeErrormodelrCORRESPOND_PROMPTagent) r)rM auth_tokenr@r9r:r; resolved_auth resolved_keyr? anth_modelrAs r __init__!PydanticAICorrespondence.__init__rs &?EP(9:Pd>P>P"Lbjjnn5K&L E"**..1D"E  ?3EIJ (>H'AJP    -+  rc# SSKJn S[U5SU3nU"[U5R 5SS9nUR R XV/5IShvN nUR$N![aJn[R"S[U5RURU5 [SS S 9sSnA$SnAff=f7f) Nr) BinaryContentzTarget table on this page: z Chunk to test: z image/png)data media_typezScorrespond: LLM call failed; defaulting to no-match page_image={} ordinal={} exc={}Fz*default no-match on correspondence failurerr)rDrWr4r read_bytesrOrunoutput Exceptionr errornameordinalr) r)r*r+r,rW user_textimageresultexcs r r PydanticAICorrespondence.matchess .+?6+B*CC[\j[k l 4 #3#>#>#@[Y m::>>9*<==F== > m LLeZ %%   )?kl l msF6CA+A)A+(C)A++ B?5?B:4B?5C:B??C)rOrM)NNN)rM Optional[str]rPrgr@rgr/Noner.) rrrr__doc__rJrTrrrrr r6r6dsu 0M $$(!%    "      Dmmm m  mrr6c@\rSrSrSrSSSjjrS SjrSrg) MockCorrespondencez$Returns a canned verdict. For tests.Nc4U=(d [SSS9Ulg)NTmockrZ)rverdict)r)ros r rTMockCorrespondence.__init__sT"7V"T rc*# XU4nUR$7fr(ro)r)r*r+r,_s r rMockCorrespondence.matchess  0||srrr()rozOptional[CorrespondenceVerdict]r/rhr.)rrrrrirTrrrrr rkrks6.U   rrk)apirncU=(d [RRSS5nUS:Xa [5$US:Xa [ 5$[ SU<S35e)NQUBER_CORRESPONDENCE_BACKENDrurnz&Unknown QUBER_CORRESPONDENCE_BACKEND: z. Expected api|mock.)rGrHrIr6rk ValueError)backendselecteds r get_correspondencer{sTO"**..)GOH5'))6!## =h\I]^ __r)r+r r/rr()ryzOptional[CorrespondenceBackend]r/r%)ri __future__rr1rGpathlibrtypingrrrrlogurur pydanticr r detectorr rr2model_json_schemarNr%r4r6rkCorrespondenceBackendr{rrr rs,# AA%# I $4 ! 3 3 5a@AB5@$($$5FmFmR    .`r