Product

The product

MCP server liveOne call; a cited answer comes back Retrieval API liveBoth neural stages over REST

See it first

API sandbox Run a live query; read the call behind it Data visualization studio The same corpus rendered, three ways

Datasets

Patents & prior art live10.5M+ filings, six languages

Case law & statutes soon9M+ precedential decisions

Litigation & dockets soonWhere a claim was litigated, and how it landed

Corporate filings soonAssignees resolved to registrants

Resources

Benchmarks Recall, the holdout, and the wins vs the field Methodology & harness liveHow we score; rerun it on one GPU Research notes Dated write-ups of the runs, negative results included Documentation The REST API, the MCP server, and their limits How usage is counted The credit tariff, the ceilings, and the headers Status What is deployed, and what it promises

Changelog soonWhat shipped, when

Pricing Customers

Research

Research

Notes on measuring prior-art retrieval, written from the runs rather than about them. Every figure names the artifact it came from, so a reader who disagrees has something to check.

RSS · Reference pages live under benchmarks and methodology, and are kept current. These are dated and are not rewritten.

August 26, 20269 min readLandry Bielefeld
How we measure prior-art retrieval
The harness behind the 45.3% recall@100 figure: where the ground truth comes from, how the holdout is split, and which experiments failed.
retrievalbenchmarksevaluation