j:j$~SrSSKJr SSKrSSKJr SSKJrJrJ r J r J r J r J r Jr SSKJr SSKJrJr SSKJrJr SS KJrJr "S S \5r"S S \5r"SS\5rS\R:"\R=5SS9S3r\"SS\ 55r SSjr!"SS5r""SS5r#\ Sr$SSSjjr%g) uK TableUnifier Protocol — unifies the Camelot flavor outputs for a page into one canonical table per physical table. Camelot is run in two flavors (lattice + stream) on purpose: each captures structure the other misses. When both produce an extraction of the same physical table, the unifier combines them into one canonical table (most complete header from one, full row set from the other), using the page image as ground truth. Candidates that are genuinely distinct tables pass through untouched. This is not deduplication — the second flavor's output is a deliberate, complementary source, not an unwanted duplicate. For each page with multiple surviving classifier-positive candidates, the unifier receives the page image and all candidates and returns the canonical set. Pure-LLM (no geometric heuristics): the LLM judges which candidates describe the same physical table and emits the unified best version. Kept separate from `LLMClient` for the same reasons as `TableClassifier`: single-responsibility, potentially different model than structure-correction, isolated retry semantics. ) annotationsN)Path)AnyDictListLiteralOptionalProtocolTupleruntime_checkable)logger) BaseModelField)LangSmithTracerusage_metadata_from)DEFAULT_LLM_MODEL get_settingscZ\rSrSr%\"SS9rS\S'\"SS9rS\S'\"SS S 9rS \S 'S r g)CandidateInput&z@Stable id used to reference this candidate in the unifier output descriptionstr candidate_idz-Camelot-extracted markdown for this candidatemarkdownNzfCamelot's geometric box (x1, y1, x2, y2); informational, not load-bearing for the LLM's unify decision)defaultrz+Optional[Tuple[float, float, float, float]]bbox) __name__ __module__ __qualname____firstlineno__rr__annotations__rr__static_attributes__r2/home/mande/repo/quber/src/quber/agents/unifier.pyrr&s<*lmL#m&UVHcV8=89D 5r%rcB\rSrSr%\"SS9rS\S'\"SS9rS\S'S rg ) UnifiedTable0z)Canonical markdown for this logical tablerrrzcandidate_ids the unifier combined into this output. Single-id list means the candidate stood alone; multi-id means an actual unify.z List[str]source_candidate_idsrN) rr r!r"rrr#r*r$rr%r&r(r(0s*&QRHcR&+Z')r%r(c.\rSrSr%\"\SS9rS\S'Srg) UnifierResult8z?Canonical tables on this page after unifying the flavor outputs)default_factoryrzList[UnifiedTable]tablesrN) rr r!r"rlistr/r#r$rr%r&r,r,8s!&U"F r%r,u+You are reconciling candidate tables extracted from a single PDF page by different table-extraction strategies. Multiple candidates may represent the same logical table (extracted twice with slight differences) or distinct tables that happen to be on the same page. You are given: 1. A rendered image of the PDF page (ground truth). 2. A list of candidate tables, each with an id, markdown, and optional bbox. Your job: return the canonical set of tables present on this page. Rules: - *Same logical table* — two candidates describe the same physical table on the page (even if their bboxes, row counts, or cell content differ slightly). Return ONE unified entry combining the best parts of each (most complete header, all data rows, correct column count). List ALL contributing candidate_ids in `source_candidate_ids`. - *Distinct tables* — two candidates are different physical tables that happen to be on the same page. Return each as its own entry with `source_candidate_ids` containing only its own id. - *Use the page image* as the source of truth for which physical tables exist and what their structure should be. - *Preserve numeric values exactly* from whichever candidate(s) you draw them from. Do not invent or alter numbers. Return a JSON object conforming exactly to this schema: indentuE Return only the JSON object — no prose, no markdown code fences. c*\rSrSrSSjrSrg) TableUnifiergc # g7fNr)self page_image candidatess r&unifyTableUnifier.unifyis srNr:rr;List[CandidateInput]returnr,)rr r!r"r<r$rr%r&r5r5gs#)  r%r5cr[R"UVs/sHoR5PM snSS9$s snf)Nr1r2)jsondumps model_dump)r;cs r&candidates_payloadrFps/ ::!+,A, ,s4cZ\rSrSrSrSS SjjrS SjrS SjrSrg) PydanticAIUnifierwuTableUnifier backed by pydantic-ai with an Anthropic model. Sends the page image as `BinaryContent` alongside the candidate payload — the image is the ground-truth reference for unify decisions. NcSSKJn SSKJn SSKJn [ 5RnU=(d UR=(d [nU=(d URnU=(d URn U(aSSK J n U "X5n OU (a U"U S9n U"XS9n O [S5eXlU"U [[ S 9Ul[%S S 9Ulg) Nr)Agent)AnthropicModel)AnthropicProvider)make_oauth_anthropic_model)api_key)providerzqPydanticAIUnifier: neither ANTHROPIC_AUTH_TOKEN nor ANTHROPIC_API_KEY is set. Provide one via env or constructor.) output_type system_promptz quber-unifier)run_name) pydantic_airKpydantic_ai.models.anthropicrLpydantic_ai.providers.anthropicrMrllmmodelranthropic_auth_tokenanthropic_api_keyquber.agents._oauth_gaterN RuntimeErrorr,UNIFY_FLAVORS_PROMPTagentrtracer) r9rX auth_tokenrOrKrLrM llm_settings resolved_auth resolved_keyrN anth_modelrPs r&__init__PydanticAIUnifier.__init__s &?E#~)) @++@/@"Gl&G&G @,"@"@  K3EIJ (>H'AJP    %.  &? r%c2SS[S.SSUS.SSS./S./0$) Nmessagessystemrolecontentusertext)typernimagez[page image attached])r])r9 user_texts r& trace_inputsPydanticAIUnifier.trace_inputss; !.BC"!';!(2IJ    r%c X# SSKJn U(d [/S9$S[U53nU"[ U5R 5SS9nUR U5nURRSX`RS9IShvN nURRXE/5IShvN nURn S U R5S ./[U R5[!UR"5S .UlSSS5IShvN U $NNgN !,IShvN (df  W $=f![&asn [(R*"S UR,[U5U 5 [UV s/sH"n [/U R0U R2/S 9PM$ Os sn fsn S9sSn A $Sn A ff=f7f)Nr) BinaryContentr/zCandidates on this page: z image/png)data media_type unify_flavors)rX assistantrj)rh table_countusage_metadatazPunify: LLM call failed; returning candidates as-is page_image={} count={} exc={}rr*)rTrur,rFr read_bytesrrr_llm_runrXr^runoutputmodel_dump_jsonlenr/rusageoutputs Exceptionr errornamer(rr) r9r:r;rurqrpinputsrresultrexcrEs r&r<PydanticAIUnifier.unifysj . + +01CJ1O0PQ 4 #3#>#>#@[Y""9- {{**?F***UUY\#zz~~y.@AA*5&BXBXBZ![ \#&v}}#5&9&,,&G VUMVAVUUM  LLbJ   !('!!**ANNK[\'  sAF*(D*?D D* D#D $AD7 D*D D*F* D* D D*D'D  D'"D*&F*'D** F'45F"))F  F"F'F*"F''F*)r^rXr_)NNN)rX Optional[str]r`rrOrr@None)rqrr@zDict[str, Any]r>) rr r!r"__doc__rerrr<r$rr%r&rHrHwsi $$(!% $@$@"$@ $@  $@L  ##)#  #r%rHc<\rSrSrSrSSSjjrS SjrSrg) MockUnifierzReturns a canned UnifierResult, or pass-through if none injected. Pass-through behavior: emit one UnifiedTable per candidate, each referencing only its own candidate_id (no unifying). For tests. NcXlgr8r)r9rs r&reMockUnifier.__init__s r%c # UnURb UR$[UVs/sH"n[URUR/S9PM$ snS9$s snf7f)Nr}rv)rr,r(rr)r9r:r;_rEs r&r<MockUnifier.unifys]  ;; ";; blbl]^ ajjGWXbl  s%A)A Arr8)rzOptional[UnifierResult]r@rr>)rr r!r"rrer<r$rr%r&rrs-     )     r%r)apimockcU=(d [5RRnUS:Xa [5$US:Xa [ 5$[ SU<S35e)NrrzUnknown QUBER_UNIFIER_BACKEND: z. Expected api|mock.)rrWunifier_backendrHr ValueError)backendselecteds r& get_unifierrsQ<,.,,<rs0# YYY%N:Y9I>M + + -a89:?"J8__D  2'Yr%