Prior art

Search by limitation across more than one source

Patendra can query multiple configured patent sources, merge their rankings, and connect retrieved passages to the claim language they may teach. When retrieval fails or coverage is thin, the matter remains unresolved instead of presenting absence of results as proof of novelty.

4 retrieval sources RRF rank fusion ID-verified citations Works offline (local store)
Macro of stacked document edges, an abstract topography of prior art

The problem

Why source coverage matters

Most "patent novelty search" tools query one database and report whatever comes back. That design fails in three quiet ways: each failure looks exactly like good news.

Failure 01

Coverage gaps look like novelty

Art published only in Europe or Asia, or older families indexed differently, simply never appears in a U.S.-only query. The tool finds nothing and the report reads "no anticipating references", for the wrong reason.

Failure 02

Keywords miss concepts

Patent drafters are professional synonym inventors. A keyword search for your phrasing misses the same idea described in someone else's vocabulary. Semantic search catches it; a single keyword index does not.

Failure 03

Outages look like empty results

When the backing API times out, many tools return an empty list, indistinguishable from a clean search. Patendra records search failure as a distinct outcome from empty results, so a dead backend can never inflate a novelty score.

The sources

Four backends, one fused ranking

Each source has a different blind spot, so Patendra queries all four and lets no single one decide what the examiner simulation sees.

SourceCoverageWhat it contributes
USPTO Open Data Portal U.S. patents and published applications, full text The corpus a U.S. examiner starts from: the baseline any §102/§103 argument must survive.
Espacenet (EPO OPS) Worldwide patent families via the European Patent Office Art published only outside the U.S., surfaced before it can surprise you at examination.
Google Patents Public Data BigQuery-backed public corpus, broad and historical Recall on older filings and non-English families that narrower indexes rank poorly.
Local ChromaDB vector store Semantic index built on your machine from prior runs Conceptual matches that share no keywords with your query, and it keeps answering offline.

Fusion

How the four rankings become one

You don't need to be an engineer to follow this. The whole mechanism is three steps.

  1. Each source votes with a ranked list

    Every backend returns its own ordered list of candidate references. The lists disagree, and that's the point. A reference that only one source knows about still gets a seat at the table.

  2. Reciprocal Rank Fusion merges the votes

    RRF scores each reference by where it ranks in each list, not by each source's incomparable relevance numbers. Appearing near the top of several lists beats appearing first in one, so the fused ranking rewards agreement between independent searches without letting any single backend dominate.

  3. A cross-encoder rereads the finalists (optional)

    Rank fusion is fast but shallow. An optional cross-encoder reranker then reads your claim text and each finalist together, side by side, and reorders the shortlist by how directly the reference actually bears on the claim. It is a slower, closer read reserved for the candidates that matter.

Citation integrity

A rejection needs a real reference

Language models will confidently cite patents that do not exist. Patendra's examination stage is built so that a hallucinated citation is structurally incapable of becoming a rejection.

ID verification

Every citation traces to a retrieval

Every asserted §102 or §103 rejection must cite a reference that was actually retrieved from one of the four backends and verified by its document ID. If the model names a patent number that isn't in the retrieved set, that assertion cannot enter the examination memo. The evidence decides; the model only argues.

Honesty rule

Insufficient search ≠ novel

If zero references were examined (because backends failed, budgets ran out, or queries returned nothing usable) the run renders an explicit insufficient-search warning, never "novel". "We found nothing" and "we couldn't look" are different sentences, and the scoring layer treats them differently too.

Tier 2

Targeted search against the claim

The first search asks "what's out there?". The Tier 2 search asks a nastier question: "what would an opposing counsel cite?"

Claim stress test

A fresh search per surviving claim

In Patendra's three-tier refinement funnel, only claims that pass free structural validation (Tier 0) and a cheap anticipation screen (Tier 1) reach Tier 2, where a new, adversarial prior-art search is formulated specifically to anticipate or render obvious that exact claim, followed by a full simulated examination.

Why adversarial

Find the reference before the examiner does

A search that wants your claim to survive will stop early. A search instructed to destroy the claim keeps digging. If it finds the killer reference, you learn it now, at your desk, instead of eighteen months later in a first Office action. What survives Tier 2 is reported per statute: § 102 § 103 § 112(a) § 112(b)

Local store

The vector store that lives on your machine

The fourth source is a ChromaDB semantic index in your local data directory rather than a remote API, and that location is a feature.

Offline

Searches without a network

Patendra runs offline except for LLM and patent-API calls. The local store keeps semantic retrieval working even when the remote backends are unreachable, and the run records exactly which sources were live.

Private

Your queries stay yours

Searches against the local index never leave your machine. For pre-filing work where the query itself is confidential information, that matters. Full picture on the security page.

Compounding

It learns your domain

Art retrieved across runs accumulates in the index, so the tenth search in your field starts from a richer local corpus than the first. This is part of the same cross-run learning loop that tunes search strategies and scoring weights.

Retrieval is the first half of the story. What happens to the references next (element-by-element anticipation mapping, obviousness combinations, and the deterministic score) is covered in how it works and patentability scoring. Terms like RRF and anticipation are defined in the glossary.

Review the search record, not just the result count

Install the desktop app, add your API key, and run a four-source, ID-verified prior-art search on your own disclosure. Every result requires attorney review. Patendra is not legal advice.

Get started