Document Parsing — Stack

Technology landscape as of 2026-10-10. Knowledge cutoff May 2026.

Selection criteria

Evaluate on the document types that dominate your corpus, with the structures that your queries depend on, and accept that no single tool wins across formats.

• Table fidelity. Whether row and column association survives, including merged cells, multi-row headers and tables that span pages. Test this explicitly; it is the most common silent defect.
• Reading-order recovery on multi-column, sidebar and footnote-heavy layouts, which is a different capability from character extraction.
• Structural output. Whether the result is a typed element tree with heading levels, or a string. A string cannot be ungarbled later.
• Anchor richness. Page, bounding box, element identity — whatever a citation will need to resolve to a location a human can be shown.
• Determinism. Whether two runs over identical bytes produce identical text, which decides whether content-derived identifiers are usable.
• Throughput and cost per page, measured on your real page mix rather than on clean prose, since layout analysis and model-based parsing differ by orders of magnitude.
• Format coverage weighted by corpus composition — office formats, email with attachments, presentations, spreadsheets, markup and source code each behave differently.
• Failure behaviour. Whether an unparseable file raises or returns empty, and whether partial success is distinguishable from complete success.
• Deployment constraints. Whether documents may leave your boundary at all, which removes every hosted option regardless of quality.

Options