Find your next useful lead.
$20/month
100 monthly credits for standard patent queries. Up to 20 searches a day, with optional top-ups.
Explore BasicPatent research for your AI assistant
Invention.
In perspective.
Bring G62 patent research into your AI assistant.
Search 25.5 million US and European patent documents.
Powered by G64. Built for the ideas words miss. ↗
Developer MCP for clients with Bearer-token support.
Explore the optional visual workspace ↗
Opening the interactive exhibit…
Illustrative reconstruction from the original patent. Not
engineering CAD or a search result.Read the original patent ↗
25.5MUS & European patent documents in one index
1.41Mexaminer citation pairs used to train the model
3.27×BM25 citation recall in our
US development benchmark
The product
Your question.
The earlier work.
A patent, a claim, or an idea.
Real documents to explore.
G62
Search
Recorded example / US index
A foundational communications patentQUALCOMM · US 4,901,307
Many mobile calls.
Many mobile calls.
One radio band.
A different code for each.
Explore the earlier patent record around Qualcomm’s CDMA communications patent.
Whole patentUnited StatesBefore 17 Oct 1986
Selected matches from the G62 example
Selected earlier publications
02
Spread spectrum communications method and apparatus
1981 → cutoff 1986Publication precedes the example’s cutoff.
Spread spectrum apparatus for cellular mobile
communication systems
1980 → cutoff 1986Publication precedes the example’s cutoff.
Orthogonal code division multiple access communications
systems
1985 → cutoff 1986Publication precedes the example’s cutoff.
Three selected results. Original ranks shown. Open any patent to read the source.
Recorded example, not live search. Similarity and earlier
publication do not establish invalidity.Connect your AI assistant ↗
Meet G64 · The research engine behind G62
The power to see
what words miss.
Different language. Different decades. Related engineering. G64 connects the ideas behind the words—and leads you back to the evidence.
Find the connection.
Not just the same phrase.
Trained on 1.41 million citation-derived pairs, G64 learns relationships from the patent record. That helps it surface related inventions a keyword match can miss.
Patent-specific learning. Measured against keyword search.One invention.
More than one way in.
Start with a patent. Focus on a claim. Or describe the mechanism in plain English. Different views can surface different references—so an important search deserves more than one angle.
Patent · claim · plain-English ideaA lead you can
actually investigate.
Follow the patent number, inspect the dates, and open the original document. G64 gives you candidates to investigate—not a generated verdict about whether a patent is valid.
Real documents. Visible dates. Your judgment.Find the engineering.
Learned from invention.
Patent citations teach relationships that generic text similarity can overlook.
A substantial search space.
25.5 million US and European publications in the core index.
Evidence you can follow.
Source documents and dates keep the research grounded in the patent record.
The next frontier · Deeper reading
The detail that matters
may be pages inside.
A title tells you what a patent is about. The description can reveal how it actually works. We’re building deeper retrieval so those useful passages have another way to be found.
Why this changes the search
A front-page match can miss a small but important mechanism described later. Searching description passages adds another view; it does not replace the original ranking.
Description search currently covers selected US collections, not all 25.5 million documents. More coverage and better ranking are research goals—not guarantees of finding every relevant reference.
Measured progress, not just a bigger model
The ambition is big.
The evidence is specific.
On our US development benchmark, the tuned model recovers 3.27× as many examiner-cited references as full-text keyword search. Here is exactly what that means.
Examiner-citation recall @1003.27×More of the cited record.
More of the cited record.
On the same test.
30.42% versus 9.29% for BM25.
17.8M US documents. 2,000 queries.
The same date rules.
Cited references recovered in the top 100
0–100% scaleRecall is the share of labeled citations recovered—not the share of results that invalidate a claim. The remaining references matter too.
A separate European-query test77.3%A deeper reading list.
A deeper reading list.
More of the cited record.
Cited EP documents recovered in the top 1,000 for 1,000 European queries against the combined US + EP index.
Different queries and reference labelsRecall at different reading depths
0–100% scaleThese are different depths of one European-query test, not competing models. The scores cannot be compared directly with the US benchmark.
The benchmark, without the fine-print surprises
What was measured
Examiner-citation recall, averaged across queries. The US comparison uses 2,000 queries over 17.8 million documents. Published encoder results, not a new evaluation of every current retrieval feature.
What this does not prove
The US set informed development decisions and some answer families appeared in training. It is not a blind final test, legal precision, a claim-invalidity success rate, or proof of superiority over commercial search products.
Why the tests stay separate
The European-query benchmark uses 1,000 EP queries over the combined US + EP index. Different queries, references and reading depths mean its percentages answer a different question.
Where the numbers come from
G64 research record, 4 September 2026. Full-US encoder comparison and separate European-query evaluation. Rounded chart labels; exact US values determine the ratio. Detailed methodology available on request.
The model’s journey
Learn from more invention.
Earn the improvement.
Three checkpoints on the same full-US development benchmark. Model changes stay subject to evaluation—not optimism.
- 01 · Stock patent encoder23.1%
A patent-specific starting point.
- 02 · US citation training29.3%
Broader technical examples.
- 03 · US + EP citation training30.4%
G64’s retained pworld encoder.
Examiner-citation recall @100 · full-US development test · not three independent confirmations.
Your assistant. G62 patent research.
Start with curiosity.
Build toward conviction.
Start with MCP connection setup. Use the optional visual workspace to inspect patents, and the dashboard to manage access, credits and your plan.
Give the question more thought.
$200/month
1,000 shared monthly credits. Standard search, Deep thinking and source-linked Frontier analysis.
Explore DeepSTART WITH YOUR ASSISTANT
Check account access, create a scoped developer token, and copy your client’s MCP configuration. Supports clients with Bearer-token settings; G62’s ChatGPT browser OAuth connection is not available yet.
Connect your AI assistant ↗Private alpha
Put G62 to work
on your next search.
Bring a public patent or claim. We’ll help you explore the earlier work.
Request a patent search Opens Gmail to vatsal@ultrahuman.com.Founder-led searches during the private alpha.
Start with public information. Please don’t email confidential,
privileged, export-controlled, or unpublished invention
details.Your search inputs are not used for training
unless you explicitly opt in.
unless you explicitly opt in.