j,#SrSSKJr SSKJr SSKJrJrJrJ r SSK J r SSK J r JrJr SSKJr SSKJrJr SS KJr \"S S 15r\"S S 15rSr"SS\5rSSSjjrSSjrSSjrSSjrSSjr g)uWhich pages get scanned, read off the parse. ADE is the right reader for three things a financial document prints on a page, and nomination is how we aim it at them. It chooses pages, not regions. ADE locates and bounds whatever it finds on the page it is handed, so there is nothing here to bound and no second judgement to make. The only decision is where to send it. The parse already reports the two picture signals, and the table signal is a count of words in the source's text layer, so nomination makes no model call: - A picture the parse classified as a chart. Plotted values are the whole point of a chart, and no component in the pipeline reads them. - A picture the parse classified as something else with content in it — a photograph, a map, a diagram, a calendar. Also read by nothing today. - A table with no text layer under it. Its cells hold text, so something read them, and the only thing that could have is the parse reading the page image. Such a table never reached the Set-of-Mark and Camelot engine either, because that engine extracts from a text layer and there was none, so what stands in the document is the parse's own reading with nothing having checked it since. `table_verdicts` measures this against the source document rather than against a record written earlier, so it holds however the tables have been renumbered since and needs no artifact beyond the two the workflow already takes. Logos and icons are left where they are. A wordmark is printed on every page of a deck as identity rather than content, and there is nothing inside it to read. Every nominated page is scanned; nothing filters the list after it is built. Nomination is a cost optimization, not a guarantee: a page whose content the parse never classified — an infographic drawn entirely in the text layer — has no signal here and is invisible to it. The orchestrator's `pages="all"` mode exists for exactly those documents, scanning every page so nothing depends on the parse having seen what matters. ) annotations)Path)DictListMappingOptional) CoordOrigin)DoclingDocument PictureItem TableItem)logger) BaseModelField) page_wordslogoiconocremptyg?cx\rSrSr%Sr\"SS9rS\S'\"\SS9r S \S '\"\S S9r S \S '\"\S S9r S \S'Sr g) NominatedPageEz9One page the parse says holds something nothing has read.z1-based source page) descriptionintpagez=The pictures on the page that nominated it, in document order)default_factoryr List[str] picture_refsz)The class the parse assigned each of thempicture_classeszGTables on the page the parse read off the page image, in document order table_refsN) __name__ __module__ __qualname____firstlineno____doc__rr__annotations__listrrr__static_attributes__r 9/home/mande/repo/quber/src/quber/core/figures/nominate.pyrrEs^C"78D#8#*iL)"'*U"OY"]J r)rNc^ U=(d 0n0m SU 4SjjnURHnUR(dM[U5nU(a[SU55(aMAU"URSR5nUR R UR5 URRU5 M URH}nUR(dMURUR5[;aM;U"URSR5RR UR5 M [T 5Vs/sHnT UPM sn$s snf)aPages of `document` holding a picture or a table nothing has read. `table_verdicts` maps a table's `self_ref` to a verdict on whether its text stands in the page's text layer. The orchestrator passes the map this module's `table_verdicts` measures against the source document. A table nominates its page when its verdict is in `UNREAD_VERDICTS`. Omitted, no table nominates a page and only the pictures do. c6>TRU[US95$)N)r) setdefaultr)rby_pages r* page_entry"nominate_pages..page_entryds!!$ 4(@AAr)c32# UH o[;v M g7fN)FURNITURE_CLASSES.0cs r* !nominate_pages..ksC7a 117sr)rrreturnr)picturesprovrallpage_norappendself_refextendtablesgetUNREAD_VERDICTSrsorted) documenttable_verdictsverdictsr/pictureclasses nominatedtablepr.s @r*nominate_pagesrMUs#H(*GB$$|| !'* sC7CCC w||A667 %%g&6&67!!((1%zz  << ' > 5::a=(()44;;ENNK !!'w 01GAJ 00 0s!E3c ^ ^ 0nURHNm T R(dMURT RSR/5R T 5 MP 0n[ UR 55HxupE[X5unm nUH]m [ST RR 55n [U U 4SjU55n U ["U -:aSOSUT R$'M_ Mz U$![aAn[R"SU[U5RU[U55 SnAMSnAff=f)uWhether each table's text stands in the page's text layer, keyed by reference. A table is `native` when the page prints text under its box, and `ocr` when it holds cells the page has no text for — the parse read those off the page image. The comparison is a count of words under the box against the cells the table holds, and the two cases are far apart: a table typeset as text carries at least a word per filled cell, one printed as an image carries none. A page whose text cannot be read at all gives its tables no entry in the map. `nominate_pages` treats a missing entry as read, so a failure here never sends a page to be scanned on a signal nobody measured. rzsNomination: page {} text layer unreadable ({}: {}); its {} table(s) are taken as typeset text and none is nominatedNc3x# UH0oR=(d SR5(dM,Sv M2 g7f)N)textstripr4s r*r7!table_verdicts..s%U$:qvv|>R>R>T$:s+: :c3b># UH$n[U[TT55(dM Sv M& g7f)rQN) _center_in _top_left_box)r5wheightrKs r*r7rTs#X5aJq-v:V,W5s/ /nativer)rAr;r-r=r>rDitemsr Exceptionr warningtyper!lensumdata table_cellsTEXT_LAYER_RATIOr?) rEsourcer.rGrrA_widthwordsexcfilledunderrYrKs @@r*rFrF{s+-G :::   uzz!}44b 9 @ @ G! "Hw}}/  $.v$< !FFEEUEJJ$:$:UUFX5XXE38)rHrrrrIann predicteds r*rrs 7FD )D>B>NWT#3T:TXNBPB\'.->bfKA))**Gw r2C!4d; 9 NN9Q<22 33 Nr)r2)rEr rFzOptional[Mapping[str, str]]r9zList[NominatedPage])rEr rdrr9zDict[str, str])rKr rtfloatr9!tuple[float, float, float, float])rxz&tuple[float, float, float, float, str]ryrr9bool)rHr r9r)!r% __future__rpathlibrtypingrrrrdocling_core.types.doc.baser docling_core.types.doc.documentr r r logurur pydanticrrquber.agents.completenessr frozensetr3rCrcrrMrFrWrVrr r)r*rs"H#003SS%0 vv./UG,- I $37#1#1/#1#1L$Nba r)