Dataset one live

Patents & prior art

Prior-art retrieval over 10.5M+ patents. Not a search box; a retrieval layer your platform calls, benchmarked, with the source passages attached.

0.0M+

US patents in the live index

Currently processing

EP 4.7M·JP 144k·KR 40k·CN 1.1M translating·48.5M NPL references

0

filing languages resolved

0%

examiner citations in top-100 (97,913-query holdout)

Ground truth

Trained on what examiners cited.

Every patent examiner leaves a record: the prior art they cited against each claim. That is millions of labeled claim-to-art pairs, and it is the reason our numbers can be checked rather than asserted.

We structured a decade of it into training and evaluation data, then tuned two retrieval stages on it. The result walks from a claim to the nearest existing art the way the examiner would, across fields and across languages.

How you reach it

One retrieval layer, three ways in.

The product is the retrieval, not the interface. Your code calls the MCP server; your platform calls the API; you check the results by hand in the sandbox and the studio.

The product

MCP server soon

Point your platform at one endpoint. It plans the retrieval, runs both neural stages, compares claim elements, and returns an answer traced to the source. One call.

locatecomparetrace
HTTP

Retrieval API soon

The same two-stage retrieval over REST. Send a claim or a disclosure, get a ranked, scored list with source passages. For teams building their own layer on top.

To test it

Sandbox & studio

The sandbox runs a live query and shows the call that produced it. The studio renders the same index as a space you can fly through. Check a result by hand in either before you wire the API into anything.

The call

A claim in. Ranked art, with receipts, out.

No query syntax. Pass the claim text; get back scored matches, examiner-citation lineage, and the exact passage each match rests on.

  • locatenearest art, ranked and scored
  • compareclaim elements against a reference
  • tracethe source passage behind a match
request
{
  "tool": "locate",
  "input": {
    "claim": "A phone grip with a flexible
      accordion member affixed to the rear
      housing, expandable between a collapsed
      and an extended state."
  }
}
response
{
  "results": [
    { "id": "US 15,923,629", "score": 0.731,
      "examiner_cited_in": 3 },
    { "id": "US 16,777,383", "score": 0.731 },
    { "id": "US 17,021,128", "score": 0.720 }
  ],
  "trace": "specification ¶[0027], verbatim"
}

Coverage

Six offices. One index.

The decisive reference is often filed in another jurisdiction and another language. All six major offices resolve into a single searchable index.

USPTO

United States

EPO

Europe

WIPO

PCT / international

JPO

Japan

KIPO

Korea

CNIPA

China

What comes next

Patents are dataset one.

The same pipeline runs on the rest of the legal record. Each dataset is its own product, reached the same way: an MCP server and an API over structured, benchmarked retrieval.

Case law & statutes

Strategy

Litigation & dockets

Strategy

Corporate filings

Strategy

Put it in your product.

Request access to the MCP server and the API, or test the retrieval first in the app.