j!vSrSSKJr SSKrSSKJr SSKJrJrJ r J r SSK J r SSK Jr SSKJr SS KJrJr SS KJr SS KJr SS KJr SS KJr SSKJr SSKJr SSK J!r! SSK"J#r#J$r$ SSK%J&r&J'r' \ \(\(\(\(4r)SSjr*"SS5r+SSjr,SSjr-SSjr.g)aFuse the document and table extractions into the two corrected outputs. The fusion step: given a `ParseResult` (the document extraction) and the Set-of-Mark/Camelot tables (the table extraction), match the two by overlap, split any region SoM merged, annotate charts and image tables, and graft the Camelot bodies into the docling spine. An LLM heading review then demotes page decoration the parser labeled as section headings in the unified document, and reports each demotion in `heading_flags`. The result carries both corrected outputs: the SoM/Camelot tables (split, annotated) and the unified `DoclingDocument`. `fuse_artifacts` is the standalone core: it takes already-produced data and the source PDF and knows nothing about how the data was produced, so a cloud job can run it against artifacts loaded from S3. `DocumentFusion` is the local control flow: it runs the document and table extractions in-process, one after the other, then calls `fuse_artifacts`. ) annotationsN)Path)DictListOptionalTuple) CoordOrigin) TableItem)logger) LLMClientget_llm_client)ExtractedTable)camelot_bbox_to_norm)SetOfMarkExtractor) split_table)build_unified_document)review_headings) match_tables) FusionResult RegionMatch)Parser ParseResultc @# URVs0sH%oDRURUR4_M' nnURR Vs0sHofR U_M nn[X5n[XX'XS5IShvN upU (a [X5n[X5 [URXU5upU Hn [R"SU 5 M [X[U5S9IShvN n [U UUU U US9$s snfs snfNN7f)apFuse already-produced document and table extractions into both outputs. Takes the document extraction (`parse`), the table extraction (`som_tables`), and the source PDF (needed to re-extract sub-regions when a merged region is split). Knows nothing about how the inputs were produced, so the same call serves the local flow and a cloud job loading artifacts from S3. Nzfuse: {})source)documenttablesmatcheserrors heading_flagsparse)pagespage_nowidthheightrrself_refr_split_merged_regions _annotaterr errorrstrr)r som_tablesrllmp page_dimst tables_by_refr split_anyunifiedrerrrs 6/home/mande/repo/quber/src/quber/core/fusion/fusion.pyfuse_artifactsr4+s:?EAQWWahh//IE,1NN,A,AB,AqZZ],AMB5-G #8VI#Ju1g"-U^^ZR[\OG Z% *'s6{KKM #  5FB Ls:D,DDD-D D A2D=D>DDcT\rSrSrSrSS SjjrS SjrS SjrSrg) DocumentFusion\uLocal control flow: run both extractions in-process, then fuse them. Holds the parser, the table extractor, and the LLM client. The client serves the split pass, the heading review, and the default `SetOfMarkExtractor` built when no extractor is passed. `fuse` runs the whole flow; `fuse_sync` wraps it for the CLI. The cloud flow does not use this class — it runs the document and table extractions as separate jobs and calls `fuse_artifacts` directly on the loaded artifacts. NcUcSSKJn U"U5nXlU=(d [S5UlU=(d [ URS9Ulg)Nr)parser_for_preset)r+)quber.core.parsersr9parserr r+r extractor)selfr;r<r+presetr9s r3__init__DocumentFusion.__init__gsD > =&v.F ..."F&8TXX&Fc# [R"URRU5nURR U5IShvN nUIShvN n[ XCXR5IShvN $N,N$N7fN)asyncio to_threadr;r r<extract_tablesr4r+)r=r parse_taskr*r s r3fuseDocumentFusion.fusezsb&&t{{'8'8&A >>88@@ #--#EvxxHHHA-Hs6A A? A9 A?A;A?4A=5A?;A?=A?cL[R"URU55$rC)rDrunrH)r=rs r3 fuse_syncDocumentFusion.fuse_syncs{{499V,--rA)r<r+r;)NNNztuned-financial) r;zOptional[Parser]r<zOptional[SetOfMarkExtractor]r+zOptional[LLMClient]r>r)returnNone)rrrNr) __name__ __module__ __qualname____firstlineno____doc__r?rHrL__static_attributes__rAr3r6r6\s[$(26#'' G G0G! G  G  G&I.rAr6c # [U5nSnUGHLnURS:wd[UR5S:waM/URSn [UR5n URV s/sH"n X;dM [ X;XHR 5PM$ n n U V s/sH ocMU PM n n [U 5U :waM[U5Xil[XiXU5IShvN n[U5S:dMUSXi'[USSSS9HunnURX-U5 M Sn[R"SUR [U55 GMO Xg4$s sn fs sn fN7f) aReplace each SoM-merged table with the tables `split_table` recovers. Only the single-SoM-table case is handled (one fused region docling sees as several); a region with more than one SoM table is too ambiguous to split and is left for the graft pass to report. F som_mergedrN)startTz7fuse: split SoM-merged region on page {} into {} tables)listkindlen som_indicesdocling_table_refs_docling_norm_boxpager)rr enumerateinsertr info)rr*rr/r-r+outr0matchidxcountr boundariesbsplitsoffsetextras r3r&r&sf z CI :: %U->->)?1)D "E,,--- -! G m. :: F-  ",=Aa = z?e # f+"38UDD v;?ayCH!*6!":Q!?  3</"@I KKI F  /8 >/ > Es=A1E23 E&E&E2"E+,E+27E2)E0*E2?A2E2cUH_nURS:XaURH nSXlM M1URS:XdMCURH nSXlM Ma g)zGSet `ExtractedTable.kind` for image tables and charts from the matches. image_tablechartN)r\r^)rr*rfis r3r'r'sX :: &&&%2 "' ZZ7 "&&%, "' rAcUR(dgURUS5up4URSRnUR[R :Xa:UR U- URU- URU- URU- 4$[UR URURUR4X45$)zKA docling table's box in the normalized top-left frame split_table expects.N)g @g@r) provgetbbox coord_originr TOPLEFTlr.rirkr)tabler-rar#r$rvs r3r`r`s ::MM$7ME ::a=  D K///%&QQ  @% PPrA) r rr*List[ExtractedTable]rrr+r rNr)rList[RegionMatch]r*r{rrr/zDict[str, TableItem]r-Dict[int, Tuple[float, float]]r+r rNz!Tuple[List[ExtractedTable], bool])rr|r*r{rNrO)rzr r-r}raintrNzOptional[NormBox])/rT __future__rrDpathlibrtypingrrrrdocling_core.types.doc.baser docling_core.types.doc.documentr logurur quber.agents.llm_clientr r quber.core.extractors.baser5quber.core.extractors.camelot.correspondence.geometryr!quber.core.extractors.set_of_markr'quber.core.extractors.set_of_mark.splitrquber.core.fusion.graftr quber.core.fusion.heading_reviewrquber.core.fusion.matchingrquber.core.fusion.modelsrrr:rrfloatNormBoxr4r6r&r'r`rVrAr3rs $#..35=5V@?:<3>2 ueU* +. .$. .  .  .b(.(.V, ,$, ,( , . ,  ,',^- Q  Q!? QGJ Q QrA