Prior art analysis
A search box tells you what looks similar.This tells you which parts are already taken.
Describe an invention in plain English. Patently breaks it into the discrete features a novelty search has to clear, then shows which of them the prior art already teaches — every finding backed by a passage quoted verbatim from the patent it came from.
Four things a similarity search cannot give you
Claim-element decomposition
Your description is split into the discrete technical features an independent claim's limitations decompose into — not treated as one blob of text.
Multi-angle retrieval
One embedding of a paragraph averages away the specifics that determine novelty. Patently searches 4–6 targeted angles and fuses the rankings.
Evidence you can check
Every filled cell carries a quote verified character-for-character against the source abstract. A citation that can't be located is downgraded automatically.
It admits when it can't tell
When the corpus holds nothing close to your field, it says so rather than reporting a clean score. An absence of evidence is not evidence of novelty.
How it works
- 01
Decompose
One model call turns the disclosure into claim elements and search angles.
- 02
Retrieve
Angles are embedded and searched in parallel, then fused by Reciprocal Rank Fusion.
- 03
Assess
One model call maps every element against every candidate, with a quote for each.
- 04
Synthesise
Grounding checks, whitespace, and §103 pairs — computed, not asked.
Two model calls per analysis, regardless of how many candidates come back. Everything that can be computed is computed rather than asked.
What it is not
- It searches an indexed corpus of US patent abstracts ending around 2014, not the full patent literature. A recent idea will honestly come back inconclusive.
- Coverage is judged from abstracts, not full claim text. An abstract can omit something the claims teach.
- It is a research tool, not a freedom-to-operate opinion, and not legal advice.