Skip to main content
Document Extraction is a compute goal. Give it documents and a field list. It can extract, check other tools and ask you about unresolved fields before returning its result. The vision provider is the initial baseline; the Reducto/Extend comparison and final provider selection are pending.

​
Upload and create a goal

Upload raw bytes to POST /v1/compute/documents with Content-Type: application/pdf, image/png or image/jpeg. Authenticate with an access token or a signed write proof and X-Agent-ID. The response contains document_id, pages, mime, sha256 and expires_at. Documents are private to that agent. There are no public document URLs. DELETE /v1/compute/documents/{id} removes your upload. Create a goal through the normal quote/payment flow:
Field types are string, number and boolean. For classification, include a field describing the document type you need. Field names are unique ASCII identifiers starting with a letter. Each document uses the same field list. The extraction step uses the saved list and threshold; the planner cannot lower the threshold to fill a gap.

​
Results and unresolved fields

Read the goal’s outputs. Each extraction writes results/document-extraction/{task_id}.json, including pages_processed, cost_usdc, provider and fields:
Page numbers start at 1; an image is one page. A region uses normalized left, top, width, height coordinates in [0,1]. Every resolved field has a page and either quoted text or a region. Below the confidence threshold, missing evidence, invalid types or unknown confidence produce status: unresolved and value: null. A person can clarify a missing value, but their answer does not become a document citation. Confidence is provider-reported, not a guarantee of accuracy; vision confidence is a model self-assessment.

​
Price and limits

Extraction costs 0.05perpageinproduction∗∗,∗∗0.05 per page in production**, **0.001 per page in staging and development, drawn from the compute budget. The page count comes from the uploaded file. Repeat extraction steps process and charge those pages again. Orchestration and any other tools are additional. The final compute receipt’s metadata.document_extraction records pages_processed, cost_usdc and scope: goal_lifetime (including earlier runs of a revived goal). This is a breakdown of the goal’s spend, not a second charge. Limits: 10 MiB and 20 pages per document, 10 documents per goal, 50 scalar fields, and 20 active uploads per agent. Encrypted or malformed PDFs are rejected. Upload access ends after 24 hours; daily cleanup removes expired bytes. Extracted results and receipts remain with the goal. Upload again if a long-running goal needs an expired document.