Document Intelligence API
Document image → structured fields, answers, and tables — one endpoint family.

How we compare

The hyperscaler IDP services are powerful but priced per feature and per document type — OCR is one rate, forms another, tables another, expense/ID parsers another again, and you stitch the calls together yourself. Brainiall folds recognition, doc-type-aware field extraction, document Q&A and table extraction into one endpoint family at a single per-page price, self-serve from the first call.

Prices are list-price approximations for orientation, not quotes — hyperscaler IDP pricing is tiered and feature-specific. Always check each vendor's current pricing page.

Pricing

One per-page price covers OCR, field extraction and Q&A; table extraction is a small step up. The free tier is generous enough to wire up an end-to-end pipeline.

PAYG: $0.01 / page for /extract and /query, $0.012 / page for /tables (Brainiall Document Intelligence engine). One page = one document image. No per-feature surcharges, no minimum spend, no contract.

Three calls: extract, query, tables

# 1. Extract structured fields — doc_type picks the schema
POST https://api.brainiall.com/v1/document/extract
  {"image": "<base64 png/jpeg>", "doc_type": "receipt"}
  -> {"doc_type": "receipt",
      "fields": {"merchant_name": "Blue Bottle Cafe", "date": "2026-05-12",
                 "items": [{"name": "Cappuccino", "quantity": 2, "unit_price": 4.00, "total": 8.00}, ...],
                 "subtotal": 15.50, "tax": 1.24, "total": 16.74, "payment_method": null},
      "text": "<recognised plain text>",
      "extraction_engine": "Brainiall Document Intelligence engine"}
  # doc_type ∈ receipt | invoice | id | contract | form | generic

# 2. Ask a natural-language question about the document
POST https://api.brainiall.com/v1/document/query
  {"image": "<base64 png/jpeg>", "question": "What was the total amount paid?"}
  -> {"answer": "16.74", "found": true, "supporting_text": "TOTAL  16.74"}

# 3. Pull every table out as headers + rows
POST https://api.brainiall.com/v1/document/tables
  {"image": "<base64 png/jpeg>"}
  -> {"table_count": 1,
      "tables": [{"title": "Line items", "headers": ["Item", "Qty", "Price"],
                  "rows": [["Cappuccino", "2", "8.00"], ["Croissant", "1", "3.50"]],
                  "row_count": 2, "column_count": 3}]}

Pass doc_type: "generic" when you don't know the document kind — you get a short description, a best-guess document type, all labelled key-value pairs, and detected dates, amounts and entities. For multi-page documents, split into page images and call once per page. If a page has no readable text the API returns a 422 rather than guessing.

What it's for

• Accounts-payable & expense automation: drop in a scanned invoice or receipt and get the vendor, dates, line items, tax and total as JSON — straight into your ledger, no template configuration.
• Onboarding & KYC document capture: parse the name, document number, dates and MRZ off an ID document into structured fields your verification flow can check.
• Contract & agreement review: pull parties, effective dates, term, governing law and key obligations out of a contract page, or ask a direct question ("what's the notice period?") and get the answer plus the supporting line.
• Form & questionnaire intake: turn a filled-in form into a list of label/value pairs and checkbox states — useful for digitising paper intake at scale.
• Table-heavy reports: lift every table out of a financial statement, price list or lab report into clean headers and rows for downstream analysis.
• One bill, one key: this runs on the same Brainiall API key and usage-based billing as PDF → Markdown and the rest of the catalog — no separate IDP vendor to procure.

Latency profile

Document Intelligence is a thin orchestration layer — it does not host a model itself. End-to-end timing is dominated by the underlying document reader and, where applicable, the OCR step. Plan around these envelopes.

• Native-text input (the fast path): a document image with sharp printed text or a PDF page with an embedded text layer recognises in ~1.3 s/page. /extract, /query and /tables all sit comfortably inside a synchronous HTTP call here.
• Scanned or image-only input (the OCR path): camera scans, photos of paper, faxed documents — the document model has to run full OCR on CPU. Budget ~20–30 s/page. A two-page scanned contract through /extract or /v1/document/layout can run ~25 s end-to-end.
• Translation flow: /v1/document/translate (available via the Document AI Expansion bundle) routes text through the translation model after extraction; on CPU this is around 12–15 s end-to-end, almost all of it the translation step.
• Recommended pattern for long or scanned input: treat the request as an async job, the same way dubbing and speech-to-speech already do — submit, render a progress indicator, fetch the result when ready.

Press kit & resources

What reviewers, integrators and procurement teams typically ask for.

More document & text APIs

Same single API key, same usage-based pricing, different problem solved.

PDF → MarkdownFull-document layout to clean Markdown

Document AI (OCR)Printed-text recognition with a confidence gate

Text IntelligenceSummarization & grounded Q&A over text

Knowledge APIManaged RAG — ingest docs, ask questions