jy%jSrSSKJr SSKrSSKJr SSKJrJrJ r J r J r J r J r Jr SSKJr SSKJrJr SSKJr SS KJrJr SS KJrJr "S S \5r"S S\5rS\R<"\R?5SS9S3r \"SS\ 55r!"SS5r""SS5r#\ Sr$SSSjjr%g)u{ TableDetector Protocol — the page image is the arbiter of what is a table and where. The detector looks at a rendered page image and returns the real tables on that page, in top-to-bottom reading order, each marked with a rough bounding box and a short description. Its list is the reference point for the whole Camelot-correspondence flow: it sets *how many* tables a page has and *roughly where* each one sits. Camelot chunks are matched to these detected tables downstream, in `quber.core.extractors.camelot.correspondence` (`matching.assign_chunks`, called from `orchestrator.py`). The boxes should tightly bound each table: correspondence matches Camelot chunks to detected tables by 2D bbox overlap, so box accuracy is load-bearing. The description plays no part in matching. It becomes the table's title when structure correction supplies none, and it is recorded as `detected_description`. Kept separate from `LLMClient`/`TableClassifier`/`TableUnifier` for the same single-responsibility reasons: its own system prompt, its own output schema, its own retry handling, and a model that can be tuned independently of the other agents. ) annotationsN)Path)AnyDictListLiteralOptionalProtocolTupleruntime_checkable)logger) BaseModelField)supports_sampling_temperature)LangSmithTracerusage_metadata_from)DEFAULT_LLM_MODEL get_settingsc^\rSrSr%\"SSS9rS\S'\"SSS 9rS \S '\"S S S 9rS\S'Sr g) DetectedTable(zI1-based position of this table in top-to-bottom reading order on the page)ge descriptionintordinalNzBox tightly enclosing the table, normalized to 0..1 with the page's top-left as origin: (x1, y1, x2, y2). Correspondence matches Camelot chunks to detected tables by 2D bbox overlap, so accuracy matters.)defaultrz+Optional[Tuple[float, float, float, float]]bboxuShort description of what this table is — its visible title/caption or, if none, a one-line summary of its columns. The content hook the correspondence step matches Camelot chunks against.strr) __name__ __module__ __qualname____firstlineno__rr__annotations__rr__static_attributes__r!3/home/mande/repo/quber/src/quber/agents/detector.pyrr(sR _GS9> Q9D 5 BKr(rc.\rSrSr%\"\SS9rS\S'Srg)DetectorResult?zKReal tables on the page, ordered top to bottom. Empty if the page has none.)default_factoryrzList[DetectedTable]tablesr!N) r"r#r$r%rlistr.r&r'r!r(r)r+r+?s"'a#F r(r+u"You are looking at a rendered image of a single PDF page. Identify the *real data tables* on the page. A real data table is a rectangular grid: a header naming columns AND at least one row of aligned data cells. Report a table ONLY if you can see both. NEVER report these as tables: - Charts or graphs (bar, line, pie) — even when they carry numeric data labels. A chart is not a table. - Page furniture: footers, headers, decorative banners, logos, page numbers, section dividers. - Footnote lists, narrative paragraphs, bullet lists, multi-column prose, Table-of-Contents entries. Return every real table on the page, in top-to-bottom reading order: - Number them with `ordinal` starting at 1 for the topmost table. - Give each a `bbox` normalized to 0..1 (page top-left is the origin): (x1, y1, x2, y2). The box must TIGHTLY enclose the table's header and data cells — not the surrounding whitespace, caption, or footer. If you are not sure a table is present, do not report it. - Give each a short `description`: its visible title/caption if it has one, otherwise a one-line summary of its columns. This is what later steps use to recognize the table, so make it specific. If the page has no real tables, return an empty list. Return a JSON object conforming exactly to this schema: )indentuE Return only the JSON object — no prose, no markdown code fences. c\rSrSrSSjrSrg) TableDetectornc # g7fNr!)self page_images r)detectTableDetector.detectpssr!Nr8rreturnr+)r"r#r$r%r9r'r!r(r)r3r3nsCr(r3cX\rSrSrSrSrSSS\4S SjjrS SjrS SjrSr g) PydanticAITableDetectorsuTableDetector backed by pydantic-ai with an Anthropic model. Sends the page image as `BinaryContent`; the image is the arbiter of what tables exist. Auth resolution mirrors the other agents: `ANTHROPIC_AUTH_TOKEN` (OAuth subscription) preferred over `ANTHROPIC_API_KEY`. Model defaults to Haiku 4.5 — the codebase default — but the detector is the agent most likely to move to a stronger model, since the table count it returns is load-bearing for everything downstream. gNcSSKJn SSKJn SSKJn SSKJn [5Rn U=(d U R=(d [nUb[U5(dSnU=(d U Rn U=(d U Rn U (aSSKJn U "X5n OU (a U"U S9nU"XS9n O [#S 5eXl X@lUcSOU"US 9nU"U [&[(US 9Ul[-S S 9Ulg)Nr)Agent)AnthropicModel)AnthropicProvider) ModelSettings)make_oauth_anthropic_model)api_key)providerzwPydanticAITableDetector: neither ANTHROPIC_AUTH_TOKEN nor ANTHROPIC_API_KEY is set. Provide one via env or constructor.) temperature) output_type system_promptmodel_settingszquber-detector)run_name) pydantic_airApydantic_ai.models.anthropicrBpydantic_ai.providers.anthropicrCpydantic_ai.settingsrDrllmmodelrranthropic_auth_tokenanthropic_api_keyquber.agents._oauth_gaterE RuntimeErrorrHr+DETECT_TABLES_PROMPTagentrtracer)r7rR auth_tokenrFrHrArBrCrD llm_settings resolved_auth resolved_keyrE anth_modelrGrKs r)__init__ PydanticAITableDetector.__init__s &?E6#~)) @++@/@  "+H+O+OK"Gl&G&G @,"@"@  K3EIJ (>H'AJP   &!,!4-T_:` &.)   &/?@ r(c2SS[S.SSSS.SS S./S./0$) NmessagessystemrolecontentusertextDetect the tables on this page.)typerhimagez[page image attached])rW)r7s r) trace_inputs$PydanticAITableDetector.trace_inputss< !.BC"!'1RS!(2IJ    r(c# SSKJn U"[U5R5SS9nUR 5nUR R SX@RS9IShvN nURRSU/5IShvN nURnSUR5S ./[UR5[UR5S .UlSSS5IShvN U$NNgN !,IShvN (df  W$=f!["a>n[$R&"S [U5R(U5 [+/S 9sSnA$SnAff=f7f) Nr) BinaryContentz image/png)data media_type detect_tables)rRri assistantrd)rb table_countusage_metadatazGdetect: LLM call failed; returning empty detection page_image={} exc={}r.)rMror read_bytesrlrYllm_runrRrXrunoutputmodel_dump_jsonlenr.rusageoutputs Exceptionr errornamer+) r7r8rorkinputsryresultrzexcs r)r9PydanticAITableDetector.detects-4 #3#>#>#@[Y""$ -{{**?F***UUY\#zz~~/PRW.XYY*5&BXBXBZ![ \#&v}}#5&9&,,&G VUMVYVUUM - LLYZ %%  ", ,  -s5E(D C+!D $!C1C-AC1 D $C/%D *E+D -C1/D 1D 7C: 8 D D E D E3E E EEE)rXrRrHrY) rR Optional[str]rZrrFrrHzOptional[float]r<None)r<zDict[str, Any]r;) r"r#r$r%__doc__DEFAULT_TEMPERATUREr_rlr9r'r!r(r)r>r>ssc $ $$(!%': +A+A"+A +A % +A  +AZ  -r(r>c0\rSrSrSrSSSjjrS SjrSrg) MockTableDetectorz+Returns a canned DetectorResult. For tests.Nc2U=(d [/S9Ulg)Nrv)r+r)r7rs r)r_MockTableDetector.__init__s9b 9 r(c&# UnUR$7fr6r)r7r8_s r)r9MockTableDetector.detects {{srr6)rzOptional[DetectorResult]r<rr;)r"r#r$r%rr_r9r'r!r(r)rrs5:r(r)apimockcU=(d [5RRnUS:Xa [5$US:Xa [ 5$[ SU<S35e)Nrrz Unknown QUBER_DETECTOR_BACKEND: z. Expected api|mock.)rrQdetector_backendr>r ValueError)backendselecteds r) get_detectorrsR=,.,,==H5&((6 "" 7|CWX YYr(r6)rzOptional[DetectorBackend]r<r3)&r __future__rjsonpathlibrtypingrrrrr r r r logurur pydanticrrquber.agents.factoryrquber.agents.langsmith_tracerrrquber.settingsrrrr+dumpsmodel_json_schemarWr3r>rDetectorBackendrr!r(r)rs2# YYY%>N:I.Y>N , , .q9:;?"JDHDDe-e-P-(Zr(