Extract Schema And Citations

Integrator calls /extract with a JSON schema describing the fields to pull from an invoice (invoice_number, total, line_items[]). They make every field a loosely typed string and add no field descriptions.

Define a precise schema: correct types (numbers as numbers, dates as dates), arrays for repeated structures (line_items[]), and per-field descriptions that disambiguate (e.g., 'total' = grand total incl. tax). A well-specified schema is the primary lever on extraction accuracy. Validate returned va…

Extract returns each field value alongside a citation pointing to the source location in the document (page + region/block). The integrator stores only the values and discards the citations.

Persist the citation alongside each extracted value so the field is auditable and a human can verify it against the source. Source-grounded extraction is the core trust feature — discarding citations turns a verifiable extraction into an unverifiable guess. Surface citations in any human-review UI.…

Integrator already parses every document, then sends the full parsed text to a general LLM with an ad-hoc 'pull these fields' prompt instead of using /extract.

Prefer the purpose-built /extract path (schema-constrained, source-cited) over an ad-hoc LLM prompt for structured field extraction: it returns typed, validated, grounded output without bespoke prompt engineering or a separate hallucination guard. Reserve the parse-then-LLM path for open-ended summ…