Checkpoint: Table + OCR + Chunked JSON Pipeline

No new API, just the pieces put together

Lessons 7-12 covered pipeline options, table structure, OCR, picture extraction, chunking, and JSON export separately, one concept at a time. This checkpoint runs the actual pipeline you'd build from those pieces: convert two structurally different documents with one default DocumentConverter, chunk each with HybridChunker, and write everything out as one JSON file, source file, headings, and both the raw and contextualized text for every chunk.

Why these two specific files

quarterly_report.pdf has a real text layer and a table, it exercises the table-structure path from Lesson 8. scanned_invoice.pdf has no text layer at all, it exercises the OCR path from Lesson 9. Both convert correctly with the exact same default DocumentConverter(), no per-file configuration needed, do_ocr and do_table_structure are on by default and each only does work where there's something for it to do.

The output shape

{
    "source": filename,
    "headings": chunk.meta.headings,
    "text": chunk.text,
    "contextualized_text": chunker.contextualize(chunk=chunk),
}

This is close to the actual row shape you'd insert into a vector store: contextualized_text is what gets embedded, text and headings stay around as metadata for display and provenance. Lesson 18's capstone extends this exact shape into something queryable.

scanned_invoice.pdf's single chunk ends up with headings: None, that document has no section headings at all (it's a flat invoice, not a sectioned report), HybridChunker still chunks it correctly, there's just nothing to attach to .meta.headings.