Product
Velrim is generally available
Velrim is a document extraction API. You send a JSON Schema, or a Zod or Pydantic model through the SDKs, together with a document. You get back a typed object where every field carries a state, a confidence score, and page anchors that point at where in the document the value came from. As of today it is out of beta.
What changed since the beta
Every present field now carries a grounding member: verified when the value itself was located inside the cited region, cited when an anchor is attached but the value was not confirmed there, ungrounded when no accepted span exists. The confidence score was refit on those grounding signals, and the current curves are published at /reliability under cal-2026.08-4.
Webhook signing secrets can be rotated from the dashboard; for 24 hours after a rotation, deliveries are signed with both secrets so a receiver keeps verifying while it switches. Free accounts run at 2 requests per second and 100 pages per day until a first top-up or a saved card, then at the default limits. The full history is in the changelog.
What we measured
Before calling it generally available, we ran Velrim and five other extraction setups on 124 real documents and counted how often each one invented a value for a field that was not in the document. Ours invented about 11%, inside the noise of the bare model it runs on, and there are two places where we lose. The whole comparison, with the raw outputs and the scorer, is at /research.
Pricing
Flat $0.02 per successfully extracted page, drawn from a prepaid balance you top up from $5. The balance never expires, and nothing recurs unless you turn on auto-recharge. The free grant is about 500 pages and needs no card.
Sign up to start, or read the docs. SDKs: @velrim/sdk for TypeScript and velrim for Python; an MCP server at mcp.velrim.com/mcp. Questions to hello@velrim.com.