Documentation

Extraction · Extract tables

Pull every recoverable table out of a PDF as data: rows plus a CSV rendering, each tagged with the page it came from. Deterministic by default — no model is invoked and nothing is guessed.

Endpoint: POST https://inkrun.dev/api/v1/extract/tables, authenticated with Authorization: Bearer sk_....
MCP tool: extract_tables.
Cost: deterministic tables are 1 credit per 20 pages — the deterministic tier's price (see Extract → Tiers & credits).

Request body (JSON)

Field	Type	Required	Description
`file_base64`	string	one of	Standard base64 of the PDF bytes. (`file_path` works only on the local stdio MCP server, where the file is read on your machine.)
`file_url`	string	one of	An `http(s)` URL the server downloads the PDF from — cheaper than base64 for agents, which would otherwise emit the whole file as tokens. Must be publicly reachable: private, loopback and link-local addresses are refused, redirects are capped, and the download is size-capped to your plan’s document limit.
`ocr_fallback`	boolean	no	Read tables off scanned pages' images too. Defaults to `false`; OCR pricing and caveats are identical to Extract.
`mime`	string	no	Content type hint; PDFs are detected without it.

It takes no preset or schema — it always returns all tables (sending one is a 400). Want the same tables inside the extract envelope? Use Extract with preset table_dump — it honours ocr_fallback identically.

bash

curl "https://inkrun.dev/api/v1/extract/tables" \
  -H "Authorization: Bearer $INKRUN_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{ \"file_base64\": \"$(base64 -i price-list.pdf)\" }"

Response

json

{
  "id": "ext_71d0b2ce",
  "tables": [
    {
      "page": 1,
      "index": 0,
      "rows": [
        ["SKU", "Description", "Unit price"],
        ["A-100", "Widget", "12.50"],
        ["A-200", "Gadget", "34.00"]
      ],
      "csv": "SKU,Description,Unit price\nA-100,Widget,12.50\nA-200,Gadget,34.00"
    }
  ],
  "extraction_method": "deterministic",
  "warnings": [
    "A table on page 3 could not be deterministically reconstructed and was skipped."
  ],
  "credits_charged": 1,
  "pages": 4,
  "size_bytes": 96411,
  "duration_ms": 233
}

Field	Description
`tables[]`	Every recovered table: `rows` (arrays-of-arrays of cell strings, row-major), a `csv` rendering, and `{ page, index }` provenance.
`warnings`	A table whose grid can't be recovered from positioned text (borderless merged cells, wrapped multi-line rows) produces a warning naming its page — instead of a half-guessed grid.
`extraction_method`	`deterministic`, or `ocr` when `ocr_fallback` read scanned pages.
`credits_charged`	What this call actually cost.

Scanned pages

Works on ruled tables and whitespace-aligned tables in the document's text layer. A scanned page has no positioned text at all, so it yields no tables by default — the response tells you so in warnings. Pass ocr_fallback: true to read its tables off the page image instead; call Inspect first (free) to see which pages are scanned and what that would cost.

The document is processed in memory and never stored — page images included. See Data handling.

Errors

Identical to Extract → Errors, plus 400 invalid_request if a preset or schema is sent.

Machine-readable spec

This page is the human reference; the machine-readable OpenAPI spec for the v1 API is planned as its own piece of work.