Combined per-page extraction
One vision call per page extracts heading candidates, tables (LaTeX tabular), formulas (LaTeX math), and figures in a single prompt.
Semantic structure is a separate layer.
get_semantic_document_structure() writes semantic-structure.json alongside the page index.
It adds heading and section structure plus cross-page merged tables, formulas, and figures without changing the shape of pages[].
Cross-page aggregation
Headings are assembled into a section tree. Tables, formulas, and figures are merged across page boundaries using truncation flags.
Output shape
root contains the heading tree. Optional top-level tables[], formulas[], and figures[] carry cross-page merged elements with startPage/endPage.
LLM requirement
Semantic extraction requires an explicitly configured local provider and model.
detector = agent-structured-v1
Downstream rule
Read detector and strategy metadata before assuming semantic richness or cache reuse.
Not domain logic
The semantic layer is general document structure. Domain-specific interpretation belongs downstream.
Metadata that matters
| Field | Why it exists |
|---|---|
detector | identifies which semantic extraction path produced the artifact |
strategyKey | changes when provider, model, or extraction budget changes enough to invalidate reuse |
pageIndexArtifactPath | links the semantic layer back to the stable page index |
pageArtifactPath | lets section nodes point back to the originating page artifact |
Still not domain logic.
The semantic layer is general document structure. It should not encode datasheet-specific, EDA-specific, or other downstream product semantics.