Patent search & prior-art analysis

Deciding whether an invention is new, whether a product infringes someone's rights, or whether a competitor's patent can be attacked requires searching tens of millions of documents in many languages, mostly through classification codes and boolean queries. AI adds semantic retrieval from a plain-language description, cross-lingual matching, reranking with reasoning and draft claim-to-passage mappings. It speeds up the search; the legal assessment stays with a professional.

Problem & value

• Today: patent attorneys, in-house IP teams and search firms build boolean queries over CPC/IPC classes and keywords in commercial databases or free ones (Espacenet, DEPATISnet, Google Patents, Lens, PATENTSCOPE). Vocabulary differs between applicants, fields and languages, so recall is hard to guarantee and the work is slow. SMEs and start-ups often skip searches entirely or only use a free tool.
• AI value: (1) describe the invention in prose and get semantically similar documents, (2) match across languages (CN/JP/KR/DE documents), (3) summarise and compare hits against claim features, (4) keep a documented, repeatable search log, (5) monitor new publications.
• Official offerings confirm the direction: WIPO announced "PATENTSCOPE AI-Assisted Search" on 2 July 2026 [2]; Google Patents Public Datasets provide worldwide bibliographic data for more than 90 million publications from 17 countries in BigQuery (figure at launch in 2017), plus a research table with English machine translations of titles/abstracts [5].
• No reliable public figure exists for time saved or recall gained that holds across tools; vendor numbers are marketing and should be tested on your own past searches (see Getting started).

Variants

• Novelty / patentability search before filing (inventor, SME, university tech-transfer office).
• Freedom-to-operate (FTO) for a product: claim-focused, jurisdiction-specific, legal-status aware.
• Invalidity / opposition search against a third-party patent (earlier art, non-patent literature).
• Landscape and competitor monitoring: clustering, white-space analysis, applicant trends, M&A due diligence.
• Examiner / patent-office side (public sector): AI-assisted classification and search tools inside offices (WIPO, national offices).
• Private inventors / makers: free tools plus a general LLM to formulate queries and explain hits.
• Trademark and design clearance is a separate use case: KNOW-022. Case law and statutes: KNOW-006.

How it works

1. invention description / claims / patent number
2. query understanding (LLM extracts features, synonyms, CPC candidates, multilingual keywords)
3. retrieval: hybrid (boolean/BM25 + dense embeddings + classification filters) over patent corpus + NPL
4. reranking (cross-encoder or LLM judging claim-feature coverage)
5. analysis: per-reference passage extraction, draft claim chart, family/legal-status check
6. human review: attorney/searcher accepts, rejects, refines query (iterate)
7. documented search report + monitoring

• Corpus: full text and claims of patents (EPO DOCDB/OPS, USPTO, DPMA, WIPO), CPC classification, legal status and non-patent literature (papers, standards, product manuals; see KNOW-008, KNOW-009).
• Retrieval: patent-specific or multilingual embedding models (e.g. PatentSBERTa [3], BGE-M3 with dense, sparse and multi-vector modes and 8192-token inputs [4]) combined with classic boolean search, because searchers must be able to justify and reproduce a query.
• Reasoning layer: an LLM reads claims and candidate passages and explains which feature is disclosed where. Long claims and cross-references to figures are the hard part; drawings need a VLM.
• Human-in-the-loop: always. Relevance decisions, X/Y/A categorisation, and any statement about novelty, infringement or validity are made by a qualified person.

Data preparation

Two kinds of data: external patent and literature databases (EPO, WIPO, USPTO, commercial providers) that are already structured, and the company's own material: invention disclosures, draft claims, lab reports and product descriptions. Those live in Word files on file shares, the DMS, e-mail and sometimes paper lab notebooks.

Typical gaps: invention disclosures that are vague or scattered over several documents, no uniform vocabulary between R&D and patent attorneys, scanned notebooks, and own portfolio data kept in Excel with inconsistent family and status entries.

Preparation steps: - Write each disclosure as a short, structured problem-solution-features description; this is the best query input. - Clean the own portfolio list into a tidy table (PREP-011) and add technical field tags (PREP-010). - Build a glossary with synonyms and technical terms in German and English (PREP-015). - Digitise relevant notebooks (PREP-006). - Create a test set of known relevant prior art (PREP-018).

Effort is low: the large public corpora come ready-made, so about 10 to 20 percent of the effort goes into preparation. Confidential disclosures must never reach tools before novelty is protected; see PREP-024.

Implementation paths

Which path for whom

• SMB: Path A, because specialised patent-search platforms with AI semantic search are mature and priced per seat, and a small R&D team cannot run its own index of tens of millions of documents.
• Enterprise: Path B, because IP departments want to combine commercial patent data, own portfolio data and reasoning models via APIs and keep search strategies in-house.
• Developer: Path C, because open patent data (e.g. EPO, USPTO bulk files) plus open embedding and reranker models allow a self-built search on own hardware.
• Public: Path B on a sovereign EU cloud, because research institutions and technology-transfer offices need EU hosting for unpublished invention disclosures.

Path A — Buy (recommended for most SMBs and IP departments)

• Commercial patent platforms with AI search: Questel Orbit (France, EU), IPRally (Finland, EU; graph-based semantic search on claims), Minesoft PatBase (UK), Clarivate Derwent/DWPI, PatSnap, LexisNexis PatentSight, Anaqua, XLSCOUT. Vendor features and pricing change fast, so request a pilot on your own cases.
• Free official tools: Espacenet and Patent Translate (EPO) [1], DEPATISnet and DPMAregister (DPMA) [6], PATENTSCOPE with AI-assisted search and WIPO Translate [2], Google Patents, Lens.org (non-profit). Good enough for orienting searches by non-experts.
• Check: coverage by jurisdiction, legal-status data, family handling, CN/JP/KR full text, export of the search history, EU hosting and confidentiality terms (see Risks), seat/pricing model.
• Choose A when: you search regularly but do not have an ML team. Not a substitute for a professional search for filing decisions.

Path B — Build on cloud APIs

• Data: EPO Open Patent Services (OPS), PatentsView / USPTO bulk data, Google Patents Public Datasets in BigQuery [5], national office bulk downloads; commercial data licences for legal status.
• Models: an embedding API or open embedding model for the index, a reranker, and a frontier LLM with long context (Claude, Gemini, GPT, Mistral from the EU) for claim analysis and chart drafting; use EU regions or EU providers when disclosures are unpublished.
• Orchestration: an agent loop that proposes queries, runs boolean and semantic searches, deduplicates by family, scores hits and writes the search log. Frameworks: LangGraph, LlamaIndex, plain code.
• Choose B when: you need a tailored workflow (e.g. internal invention-disclosure triage) and have developers.
• Warning: never send unpublished invention disclosures to a provider without a no-training, no-retention contract; early disclosure can destroy novelty in absolute-novelty jurisdictions.

Path C — Self-host (open source)

• Models: BGE-M3 (MIT licence) [4] or PatentSBERTa (MIT) [3, 7] for embeddings; a cross-encoder reranker (e.g. bge-reranker family); an open-weight LLM of 30–120B class for claim analysis; Docling or PaddleOCR for scanned PDFs and drawings; Argos/NLLB-class models for translation.
• Stack: OpenSearch or Elasticsearch (hybrid BM25 + kNN), Qdrant/pgvector, vLLM or Ollama. The patent open data of EPO/USPTO/DPMA/WIPO can be loaded into an own index, though full global full-text coverage is a large ETL project and legal-status data is usually licensed.
• Hardware: HW3 for a usable reasoning LLM; embedding and index serving alone fit on HW2. Storage in the TB range for a global corpus.
• Choose C when: invention disclosures are highly confidential (defence-adjacent, pharma pre-filing), or a large search volume makes SaaS seats costly.

Path D — Custom training

Worth it mainly for large IP departments, patent offices or vendors: fine-tune an embedding model on citation pairs (examiner X/Y citations, family citations) to learn domain-specific similarity, or train a CPC classifier. Public citation data is plentiful; a few hundred thousand pairs is typical for noticeable gains. Always evaluate against your own historical searches before replacing an off-the-shelf model.

State of the art (2026)

• Hybrid retrieval wins: dense plus lexical plus classification filters plus reranker consistently beat any single method; multilingual embeddings (BGE-M3 class) make cross-lingual first-pass recall practical [4].
• Official bodies are shipping AI search: WIPO's PATENTSCOPE AI-assisted search (July 2026) [2]; patent offices use AI for classification and search internally.
• LLM claim analysis (feature extraction, claim charts, explanation of why a document is relevant) is useful as a draft but hallucinates references, mis-reads "comprising" vs. "consisting of" scope and misses figure-only disclosure. Every cited passage must be verified in the source.
• Limits and failure modes: completeness cannot be proven (a missed X-document is the costly error); legal status and family data must come from authoritative sources, not from an LLM; machine translation of CJK documents loses technical terms; unpublished applications (18-month rule) are invisible by design; non-patent literature and public prior use are poorly covered; benchmark numbers from vendors are rarely comparable.

Shared resources & combinations

• Same retrieval stack (embedding-text + reranker + vector-db + search-engine) as KNOW-001 enterprise search, KNOW-002 RAG and KNOW-006 legal research; one hybrid index service and one GPU server (HW2/HW3) can serve all.
• Neighbours: KNOW-008 (scientific literature as non-patent prior art), KNOW-022 (trademark/design clearance), KNOW-015 (deep-research agent for market context), KNOW-013 (graph of inventors, applicants, citations), KNOW-009 (standards as prior art / standard-essential patents), DOC-014 (parsing PDFs and tables), LANG-001 (translation).
• Upstream: invention disclosure workflows and R&D document archives; downstream: patent drafting assistance and portfolio analytics, which are separate use cases.

Effort, cost & KPIs

• Path A: days to a few weeks (pilot, access, training); licences are typically per seat or per search volume.
• Path B: 1–3 months for a working agent plus evaluation; data licences and LLM tokens (long claims plus many references) drive cost.
• Path C: 2–4 months; GPU server, ETL of patent data and ongoing index updates.
• KPIs: recall of known relevant documents on a held-out set of past searches (the key metric), precision at top-20/50, time per search, share of searches with documented strategy, number of verified citations rejected for hallucination, monitoring alert relevance.

People & change

Involve management for the IP strategy, the head of R&D or IP as process owner, the patent attorneys and engineers as users, IT and information security (unpublished inventions are trade secrets) and legal for freedom-to-operate liability. Bring in the external patent attorney early, because their workflow and professional duties shape the tool choice.

Works council: normally no. The users are a small expert group, and tools used per search project rarely monitor individuals; if the platform logs user activity at employee level, §87(1) no. 6 BetrVG can still apply, so check it. Inform the works council under §90(1) no. 3 when work processes change.

Data protection: mostly company data; inventor names are personal data, so involve the DPO briefly and sign an AVV (Art. 28 GDPR). A DPIA is usually not needed.

AI literacy: support staff with a workshop on semantic search limits and checking claim charts (Art. 4 AI Act, GUIDE-006). Pilot with two or three real searches run in parallel with the classic method and compare recall (GUIDE-009). Typical concern: missed prior art creates liability; address it by keeping AI as an addition to the documented search strategy and having a professional sign off. Security: GUIDE-013.

Risks, regulation & compliance

• EU AI Act: minimal risk for patent search and analysis; no Annex III category applies. AI literacy duty (Art. 4) for staff applies. Offices using AI in administrative decisions should check their own context.
• GDPR: patents name inventors and applicants (personal data, but published); monitoring individuals or building inventor profiles needs a lawful basis (KNOW-012-style expert finding is a separate topic). Invention disclosures may contain employee data.
• Confidentiality / novelty: unpublished inventions are trade secrets (Directive (EU) 2016/943) and can lose novelty if disclosed; use no-training contracts, EU hosting or on-prem (Path C).
• Professional duties: FTO and validity opinions are legal advice; in Germany the reserved activities of patent attorneys (Patentanwälte) and lawyers apply. AI output is a research aid, not an opinion.
• Mitigations: documented search logs, verification of every cited passage, authoritative legal-status sources, independent human second search for filing and litigation decisions.

Getting started

1. Pick 5–10 past searches with known relevant documents and use them as a test set (recall baseline).
2. Try the free tools (Espacenet, Google Patents, PATENTSCOPE AI search) on these cases to see what semantic search already delivers.
3. Pilot 2–3 commercial platforms from Path A on the same test set; compare recall, legal-status quality and export.
4. Define a data-handling rule: which disclosures may leave the company, which need EU-only or on-prem tools.
5. If building: start with hybrid search over a single jurisdiction plus a reranker, add LLM claim charts last and keep a mandatory verification step.

Sources

1. EPO — Espacenet and Patent Translate overview — https://www.epo.org/en/searching-for-patents/technical/espacenet
2. WIPO — PATENTSCOPE (news item "PATENTSCOPE AI-Assisted Search Now Available", 2 July 2026; WIPO Translate) — https://www.wipo.int/en/web/patentscope and https://www.wipo.int/en/web/patentscope/news
3. AI-Growth-Lab — PatentSBERTa model card (Hugging Face) — https://huggingface.co/AI-Growth-Lab/PatentSBERTa
4. BAAI — BGE-M3 model card (100+ languages, 8192 tokens, MIT, dense/sparse/multi-vector) — https://huggingface.co/BAAI/bge-m3
5. Google Cloud — Google Patents Public Datasets on BigQuery — https://cloud.google.com/blog/topics/public-datasets/google-patents-public-datasets-connecting-public-paid-and-private-patent-data
6. DPMA — search services DEPATISnet, DPMAregister — https://www.dpma.de/english/search/index.html
7. PatentSBERTa repository (MIT licence) — https://github.com/AI-Growth-Lab/PatentSBERTa
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj