Use parse results

A ParseResult holds everything the parser saw: markdown, structure, figures, and page renders. This page shows how to get each out.

Whole-document markdown

result = parse("report.pdf")
print(result.markdown)          # every page, assembled

Tables come out as HTML (merged cells preserved), formulas as LaTeX, charts as markdown data tables with a caption. This single string is often all you need if you're feeding another system.

Per-page structure

for page in result.pages:
    print(page.page_num, len(page.regions), "regions")
    print(page.markdown)        # this page only

Every block on a page is a Region with a type, its recognized content, and a bounding box:

for region in result.pages[0].regions:
    print(region.region_id, region.region_type, region.bbox.as_tuple())
    # 0 text     (72.0, 90.5, 523.0, 140.2)
    # 1 table    (72.0, 160.0, 523.0, 380.9)
    # 2 chart    (90.0, 400.0, 500.0, 640.0)

Region ids are what chunks cite later — the provenance chain hangs off them.

Figures and charts

Charts and figures are cropped out as PNGs with captions and AI descriptions:

for page in result.pages:
    for fig in page.figures:
        print(fig.caption, "—", fig.description[:60])

result.save_images("out/")      # writes every crop as a PNG file

Page renders

Each page keeps its rendered image (page.image_bytes, PNG) — the canvas for drawing citation highlights.

Reuse a parse

Parsing is the expensive stage, so reuse the ParseResult object — feed it straight to the other operations (below), no re-parse. Once a document is ingested, its structure lives in the registry and reads back with get_document:

from ingestlib.services import get_document

doc = get_document(doc_id)          # a stored (ingested) document, from the registry
doc.pages, doc.regions, doc.chunks  # parsed structure, as plain dicts
doc.markdown()                      # whole-document markdown (from the artifact store)
doc.page_image(1)                   # a page render on demand (PNG bytes)

get_document reads structure from the registry; bytes (page renders, whole-doc markdown) are fetched lazily from the artifact store. An unknown or tombstoned doc_id returns None.

Feed the other operations

label  = classify(result)                        # no OCR — reuses the parse
chunks = split(result, category=label.category)  # chunks with region provenance

Passing the ParseResult (rather than the file path) is what gives split exact region ids — the standalone-path chunks cite pages only.

---

Next: Extract structured data — your schema, filled and cited from the parse.