# Extract data from a batch of PDFs

**Outcome:** a CSV with one row per paper and a citation for every cell, built from a folder of PDFs.

## What you need

Key with `files:write`, `documents:write`, `extractions:write`; the PDFs; a field list.

## Approach

Upload → ingest → define schema once → run → export. The schema is reusable; write it carefully and
reuse it across projects.

## Walkthrough

1. `POST /v1/files` for each PDF (bounded concurrency; respect rate-limits)
2. `POST /v1/documents` per file; wait for `succeeded`
3. `POST /v1/extraction-schemas` — descriptions matter more than names (extraction-schema)
4. `POST /v1/extractions` with all `document_ids`
5. Poll or subscribe; read partial rows as they land
6. `POST /v1/extractions/{id}/export` → CSV

## Handling the hard parts

- **Unparseable PDFs.** Scanned-without-OCR and encrypted files fail at ingest. Collect them and report; do not silently drop.
- **Partial failure.** Per-row `error` means the run succeeds with holes. Reconcile explicitly.
- **Nulls are the honest answer.** A `null` value means the extractor did not find it. Do not backfill with a guess.
- **Verification.** Keep the per-cell citation in your output so a human can spot-check any row.

## Cost

Extraction costs **2 credits per source** at `standard` (5 at `high_quality`), independent of field
count. Document ingest costs 1 credit per 10 pages.

For 5,000 papers averaging 12 pages: ingest ≈ 10,000 credits, extraction = 10,000 credits, exports
free — about 20,000 credits, or roughly $200 at top-up rates. Runs are capped at 500 sources, so this
is 10 runs; they queue against your concurrency limit (rate-limits).

## Production checklist

Concurrency cap · resume from partial state · store `extr_` IDs for audit · check
content-licensing before redistributing extracted values

## Related

extractions · extraction-schema · files · documents