Document Claims

Read deterministic provenance claims for a document with GET /v1/documents/:id/claims: atomic subject-predicate-object triples with evidence and confidence.

GET /v1/documents/{id}/claims returns the document's provenance claims: atomic (subject, predicate, object) triples synthesized deterministically from the document's captured fields and provenance spans. No LLM is involved in the synthesis. Each claim carries the evidence quote it was derived from, a grounded flag marking whether a source span backs it, and a confidence score, so it answers "where does this value come from" at the document level in an auditable form.

Claims are ordered grounded-first, then by descending confidence, so the most defensible statements come first. Pass grounded=true to restrict the response to span-backed claims only, which is the right default when feeding claims into a downstream system that must never surface unverifiable statements. The response is a flat { data, total } envelope with no pagination.

Response

{
  "data": [
    {
      "subject": "Invoice INV-2024-0042",
      "predicate": "has_total_amount",
      "object": "4950.00",
      "qualifier": "EUR",
      "evidence": "Total due: EUR 4,950.00",
      "segment_id": "seg_12",
      "grounded": true,
      "confidence": 0.97,
      "origin": "field_capture"
    }
  ],
  "total": 1
}

Errors

A common pattern is to run compliance or audit tooling over the grounded claims of every processed document: list documents, read each document's claims with grounded=true, and store the triples beside the structured output. Because synthesis is deterministic, re-reading a document's claims returns stable results unless the document is re-extracted.