About

A prior-art search engine that reads the references it returns. Built on our own corpus of 4,983,657 publications in Postgres with pgvector, an agentic retrieval cascade, and a full-text reading stage that grounds every statement it makes in a verbatim quote.

The system, in one page

Database

PostgreSQL 17 with the pgvector extension, on a dedicated 8-core, 64 GB machine. Not a vector database product: an ordinary relational database that also holds embeddings, so a single query can combine semantic similarity, full-text search, CPC classification, citation edges and legal dates. One HNSW index over the passage embeddings.

What is indexed

4,983,657 publications, split into 27.6 million separately embedded passages: each claim on its own, each claim again with its parent folded in, each description paragraph, each figure caption, the abstract, and the whole document. Chunking per claim is what lets a result cite "claim 7" rather than "this document looks similar". Coverage by office and by text depth is in the scope panel below.

Embeddings

Google Vertex AI gemini-embedding-001 at 768 dimensions. Benchmarked against 1024 and 3072 on the same evaluation set: identical scores, so the larger vectors buy nothing here and are not paid for.

Where the data comes from

Bulk ingest from Google's patents-public-data on BigQuery, refreshed weekly; the corpus currently runs to 2026-08-18. Text and drawings that Google lacks for DE, EP and WO documents are recovered from the European Patent Office's OPS service.

External sources

Every search also fans out live to SerpApi Google Patents, BigQuery, PQAI, EPO OPS, USPTO ODP, Lens.org, OpenAlex and IP Australia. Each reports its own state on the report page, and a source that fails or times out degrades the search rather than breaking it. External sources supply breadth and display detail; ranking is always done here, because ranking needs the full text and most APIs return only bibliography.

Reading and ranking

Retrieval produces a few thousand candidate families. Roughly 600 are screened by a language model, the strongest ~180 are read in full, and the ranking is computed from what each was measured to disclose: a verbatim quote per feature, located by code, refuted by an independent pass, and weighted by how rare that disclosure is across the references actually read. The step-by-step version.

Language models

Gemini 2.5 Flash on Vertex AI for extraction, query expansion, screening, reading and refutation. A bge-reranker-v2-m3 cross-encoder orders the shortlist. Dates, filtering, deduplication, quote grounding and quote location are deterministic code, never a model.

Search scope and measured reliability

This is a bounded technical-field search tool, not a clearance search. It searches one indexed field, to a fixed date, with measured recall well below completeness. Read these limits before relying on a result.

Corpus 4,983,657 publications Limited to 8 CPC classes covering vacuum gripping, lifting and handling. CN, US, JP, DE, KR, EP, GB, FR, WO, CA, AU, TW, ES, AT only. No IN, RU, BR or other national collections — a search here cannot see art published only in those offices. (85 further offices appear in trace amounts via family and citation expansion and are not indexed coverage.) No non-patent literature.

Current to 2026-08-18 Nothing published after this date is searched. For a novelty or validity question this is a live gap, and it widens every day until the next ingest.

Measured recall 19% @100 On a 11-query internal benchmark, 19% of known-relevant families appeared in the top 100 results — roughly four in five known-relevant families were not surfaced.

AI output accuracy verify every cell Rationales and claim charts are machine-generated drafting aids, not findings. Rationales overclaimed in 8% of 40 audited cases. Retrieval-map cells asserting coverage were wrong in 58% of 12 cells before a verification pass was added. After it, an independent review re-read the 5 surviving confirmed cells against the verbatim cited passages and judged 2 of them (40%) to be overclaims. The verification pass's own tally was 0 of 5 rejected, but that is the verifier grading its own work. The samples are small and the verifier shares a model family with the auditing judge, so treat the independent figure as a lower bound. Assume residual error and check each citation against the reference text.

Absence of results is not evidence of absence of prior art. A short result list means this corpus did not surface art — it does not mean none exists. Given the recall figure, the date ceiling, and the class and jurisdiction limits above, a negative result here cannot support a clearance, freedom-to-operate, or patentability opinion. Those require a professional search across full collections.

What is and is not indexed

Indexed CPC classes (8):

B66C1/02 — suction lifting devices
B66C1/0225 — handheld suction lifting devices
B25J15/0616 — robotic vacuum grippers
B65G47/91 — suction transfer devices
B65G49/061 — suction handling of fragile sheets
B25B11/005 — vacuum work holders
F16B47/00 — suction cups
B65G7/12 — carrying objects by hand

Genuine strengths, so you can weigh them fairly: within these classes the corpus is deep — full claim text as structured rows, paragraph- and figure-level coordinates, EN/DE cross-lingual retrieval, real legal-event and family data, and prior-art dating implemented against EPC Art. 54(2), 54(3) and 56 rather than approximated. A reference this tool does return is well-evidenced and correctly dated. The limitation is coverage and ranking recall, not the quality of what it holds.

Not indexed: anything outside the classes listed above; jurisdictions other than CN, US, JP, DE, KR, EP, GB, FR, WO, CA, AU, TW, ES, AT; non-patent literature, standards, product manuals and datasheets; and anything published after 2026-08-18.

Recall basis: 11-query internal benchmark with human-marked relevant families; fraction of known-relevant families appearing in the top 100 results. Rationale basis: independent LLM examiner audit of 40 generated 'why relevant' rationales, graded against the exact reference text the generator was shown. Retrieval-map basis: independent LLM examiner audit of coordinate-backed claim-chart cells: the fraction whose cited passage does not actually disclose the element. Independent re-read: independent review of all 18 covered cells of the grabo_gripper_novelty report; the 5 cells the verification pass had confirmed as 'discloses' were re-read against the verbatim cited passages. 2 were judged clear overclaims -- e.g. EP-0176125-A1, an adhesive wall-fixing patent, matched to 'driver pin for mechanical coupling' on the words 'pin' and 'clamped' alone -- and one further cell's verdict was defensible but cited the wrong coordinate.

What this is not

A prior-art retrieval and drafting aid. Not a search opinion, not a freedom-to-operate clearance, not legal advice. It does not state a conclusion about novelty, inventive step, validity or infringement, and the report says so on its face and in every export.

A filled cell in the element by reference chart means a passage of that reference was quoted verbatim, located, and survived a pass asked to refute it. It does not mean the element is anticipated. That judgement is the reader's.

Create a free account How it works