How it works

Tell it what the invention is

Type a description, drop a PDF, or paste a patent link. From a document we pull out the text and the drawings, rebuild the claim set, and show you an editable summary before anything is searched. Correct a claim and the search re-embeds your correction, so it never searches text you have just rejected.

It searches with many queries, not one

A long description makes one blurry query. We turn your invention into a set: a one-sentence essence, five alternative phrasings a different inventor would have used, every technical element, and every independent claim. Each searches on its own and the results are fused.

Measured: on a real search, the same reference sat at position 528 when the full description was used as one query and at position 2 when a single 30-word essence sentence was used instead.

Eleven channels run in parallel

Semantic search over 28 million text passages, claim-only semantic search, keyword and exact-phrase search, CPC classification, the citation and patent-family graph, query-by-example, cross-lingual English and German, your document's own passages, your drawings against our figure index, and a live fan-out to eight external patent APIs.

Channels are fused by rank, never by raw score: a cosine, a keyword rank and a classification count are not comparable numbers, and fusing them as if they were is how a broad, noisy channel ends up outvoting a strong match.

Prior art is filtered by law, not by feel

Give it a subject patent and a mode and only citable art comes back. Novelty includes secret prior art that was filed before your priority date and published after it (EPC Art. 54(3)); inventive step is public art only. The dates are computed by code, never by a model. Results collapse to one card per patent family, so the same invention filed in six countries is one result and not six.

The strongest candidates are read in full

Around 600 candidates get a fast screen. The best of those, roughly 180, are then read end to end: every claim and every description paragraph, hundreds of thousands of characters each. For each reference, every feature of your invention gets a verdict, a verbatim quote, and its exact location, resolved by code rather than written by a model. A quote that is not really in the document is thrown away. Every "discloses" is then handed to a second pass whose only job is to refute it.

The reading is the ranking

The score on each card is what that reference was measured to disclose, weighted by how rare the disclosure is. A feature that 150 of 180 references share counts for little; one that only 18 share is what a novelty argument turns on. A reference we could not read can never outrank one we quoted.

This is the part most tools skip. A similarity score tells you a document looks close. It cannot tell you which of your claim features it teaches, or where.

Then you write the application, with the art in front of you

Finding the art is half of it. From any report — or from nothing but a description of your invention, or a draft you already wrote — a drafting agent writes the specification, the drawings and the claims, and you talk to it about them. Point out a fact it got wrong, upload another reference, ask why a limitation is in claim 1, or ask for the whole thing rewritten in different terms.

It reads every reference before it writes, and it says which feature carries each independent claim clear of which document. After every single iteration a separate reviewer checks the draft it has never argued with: that the reference numerals, the drawings and the text agree, that the claims are supported by what was actually disclosed, and that every citation resolves to a publication that really exists and says what the draft says it says.

The mechanical checks are decided in code and mean what they say. The reviewer's judgements are shown separately, each with the sentence it is about, so you can disagree with one without doubting the other.