Invite-only pilot · API and browser upload

Know where every answer came from.

Source-preserving PDF parsing for search and RAG. Native text and visual OCR remain separate, with page coordinates and explicit conflict records attached.

Sign in to the pilot Read the API guide

Example page evidence

Native reading17conflict
OCR reading23preserved
{
  "page": 14,
  "sources": ["native", "ocr"],
  "state": "conflict",
  "action": "block"
}
How it works

Two witnesses per page. Evidence, not guesses.

01

Submit a PDF

Upload in the dashboard or use plain HTTPS. The service accepts PDFs up to 90 MiB and 200 pages during the pilot.

02

Keep both readings

Native PDF extraction and CPU OCR read the document. Raw observations stay separate even when they match.

03

Return a defensible record

Derived associations reference their source observations. Material disagreement becomes a blocking conflict, not a silent choice.

The output

JSON your retrieval layer can defend.

Page geometry, native observations, OCR observations, associations, and conflicts remain explicit. Markdown and other convenient views stay derived from that evidence.

{
  "page": 14,
  "nativeObservations": ["..."],
  "ocrObservations": ["..."],
  "criticalConflicts": ["..."],
  "requiresEscalation": true
}
Where it fits

RAG grounding

Carry page-level evidence into retrieval instead of indexing bare strings.

Enterprise search

Preserve provenance and trust metadata beside searchable content.

Audit workflows

Trace extracted values to their source page and exact observation geometry.

Pilot

Metered by completed page.

$0.001 estimated usage / completed page

This is an invite-only value proof. The dashboard records a placeholder usage estimate; it is not an invoice or a measured per-job cloud bill.

FAQ

How is this different from a plain OCR API?

PageSpatial keeps native and OCR evidence separate, associates it without overwriting either source, and exposes material conflicts.

How long are files retained?

Input PDFs and result objects are retained for up to two days. Download results into your own store.

What is supported?

PDF only, up to 90 MiB and 200 pages. Enrichment is off in the external pilot.

Does it claim to know which witness is right?

No. A material disagreement is preserved and blocks trusted output instead of being resolved from confidence alone.

Is my document used for model training?

No. The deployed parse path uses deterministic native extraction, local OCR, and rule-based association.

What happens after an idle period?

The parser scales to zero. The first job after idle can include roughly 70–102 seconds of cold-start time.