1Patsnap Eureka: AI agents that return a report, not just a result list

Typically used by: in-house IP teams, patent agencies and R&D groups who want the analysis produced, not only the documents found.

What the AI actually does

Eureka’s IP search agents, Novelty Search, FTO Search and Design FTO Search, take a technical disclosure as input and work through the sequence a searcher would: summarise the technical solution, extract the distinguishing features, derive search elements, build several strategies in parallel (semantic, Boolean, classification and citation tracking, plus paper and web routes), screen the results, then produce a feature-by-feature comparison table with each reference tagged for its citation role.² Every step is visible and editable before it runs, which is what separates an auditable agent from a black box.

The drafting agents, Invention Disclosure, Patent Drafting and Office Action Response, work against CNIPA, USPTO and EPO document expectations and can call the search agents mid-workflow to pull family documents, citations and legal events into the draft.³

On the engineering side, Eureka Engineering applies the same AI evidence base to R&D questions rather than legal ones: Find Solutions with TRIZ, Technical Q&A and Quick Research, producing DFMEA reports, technology roadmaps and feasibility assessments. Patsnap states that outputs carry “a traceable citation path to the underlying sources” back to the 200M+ patents and scientific literature behind the platform.⁴ For an engineering team that means the corpus clearing a design for freedom to operate also answers how other people have solved the same problem.

The benchmark, and how to read it

Patsnap publishes a benchmark methodology, PatentBench, rather than only a headline number. The novelty search round uses 340 cross-jurisdiction patent family samples. Ground truth is the set of X and Y references cited by examiners at different receiving offices, deduplicated and normalised by patent family. Two metrics are defined: X Hit Rate, meaning a correct answer appears in the top 1, 3 or 5 results, and X Recall Rate, the share of correct answers found within the top 100. On that round, tested in July 2026, the Patsnap novelty search agent records an 85% X Hit Rate and a 37% X Recall Rate.¹ A separate design patent round reports a 77% high-risk hit rate in the top 200 across 261 test cases.²

These are vendor-published results on a vendor-selected sample. The methodology is disclosed, which is unusual in this category, but it is not an independent evaluation and should be read as such.

Putting patent search inside your own agent

The part most teams miss is that Patsnap exposes this as infrastructure. The Patsnap Open Platform publishes 31 MCP servers, described as a way to “connect Patsnap’s domain-specific AI agents to LLM platforms such as Claude or custom models using the Model Context Protocol.”⁵ They include Patsnap Patent Research (novelty, invention FTO, design FTO), a cross-domain TRIZ case library, a workspace reader and 3GPP TDoc tools. Setup is an API key plus a generated connection link in Cursor, Claude Desktop or any MCP-compatible client, and the Starter tier is free with 10,000 credits valid for 90 days.⁶

Things to be aware of

• It is a commercial platform. Entry credits are free and the paid tiers are published on the product page, with enterprise terms on quote.²
• Agent outputs still require professional review. Patsnap’s own terms state outputs “may be inaccurate, incomplete, misleading” and that you “must obtain independent professional review before filing, publishing, or taking action based on them.”⁷
• The published benchmark is run against general-purpose AI models, not against other professional patent search systems.

2Espacenet: the broadest free worldwide collection

Espacenet is the European Patent Office’s free service and a standard starting point for manual searching. The EPO announced in February 2024 that it “now offers more than 150 million patent documents” from “more than 100 patent authorities around the world,” and it remains the easiest free route to simple and INPADOC family views plus INPADOC legal events, which is what a freedom-to-operate question ultimately turns on.^{8, 9} Worth knowing before you plan around it: exports are capped at “the first 500 results in xlsx and csv format,” and the EPO’s official release notes page has no entries after October 2024.¹⁰

3WIPO PATENTSCOPE: multilingual and PCT searching

PATENTSCOPE lets you “search 128.8 million patent documents including 5.5 million published international patent applications (PCT)” across 123 national and regional collections, free of charge.¹¹ Its differentiators are Cross Lingual Expansion across 14 languages and a chemical structure search for logged-in users, which together make it the strongest free option when the relevant art is Chinese, Japanese or chemical.¹² In July 2026 WIPO added an AI assistant that translates natural language into structured queries, hosted on WIPO premises; result caps sit at 10,000 for logged-in users and bulk download is prohibited.^{13, 14}

4Google Patents: quick orientation plus non-patent literature

Google Patents “includes over 120 million patent publications from 100+ patent offices around the world, as well as many more technical documents and books indexed in Google Scholar and Google Books, and documents from the Prior Art Archive,” which makes it the fastest free route to non-patent prior art.¹⁵ Two details are easy to miss: only 22 offices are indexed in full text, and the “Similar Documents” list is described in Google’s own help as “based on text similarity” rather than as a semantic or generative feature.¹⁶ CSV download is capped at the top 1,000 results, and Google states plainly that it “cannot guarantee complete coverage.”¹⁷

5USPTO Patent Public Search: examiner-grade operators, US only

The USPTO’s free public tool covers US documents only, with “full text of most patents issued from 1971 to present,” pre-grant publications from March 2001, and OCR text back to 1836, searchable with the graded proximity operators examiners use.^{18, 19} Two sentences from its own FAQ define it: “Unlike third party proprietary patent search databases, Patent Public Search does not use semantic searching,” and “At this time, Patent Public Search does not have access to foreign patent databases.”¹⁸ The USPTO’s AI-based Similarity Search does exist, but it lives in the internal PE2E examiner system, where an October 2025 memorandum made its use mandatory on plant and utility applications.²⁰

6The Lens: patents joined to the scientific literature

The Lens indexes 153.2 million patent records alongside 266 million scholarly works, and its PatCite tool mines patent literature for citations of published research so you can trace which papers influenced which patents.²¹ Check your licence before planning around it: on 1 June 2026 Cambia announced the sale of The Lens to a newly established Australian non-profit described as “a wholly owned subsidiary of SPIE,” and fee-waived accounts are now limited to academics, with commercial users required to subscribe at US$1,000 a year.^{22, 23} The Lens documents no AI search feature, though it removed its AI-use restriction clause in July 2026.²⁴

7PQAI: open-source AI prior-art search you can inspect

PQAI is an MIT-licensed, AT&T-initiated project that takes “plain language invention description as input” and “uses a number of machine learning (ML) models” trained “on past patent examination records” to surface similar art across patents from 68 countries and 100M+ research papers.^{25, 26, 27} It is a useful, inspectable second opinion rather than a system under heavy active development: the free plan returns 20 results per query, PQAI+ is $20 a month, and several supporting repositories have not been updated since 2022 and 2023.²⁸

Three more independent tools worth a look

IPRally (Finland, 2018) runs a graph neural network trained on patent examiner citations over roughly 125 million publications, and announced an API and an MCP connector in preview in July 2026; plan tiers are named but no prices are published.^{29, 30}

Ambercite (Australia, 2010) works from citation networks rather than keywords, applying network analytics across “over 106 million patents and 175 million patent citations”; pricing is quote-based.^{31, 32}

Patentfield (Kyoto, 2017) offers semantic and drawing-similarity search with a generative AI layer across JP, US, EP, CN, KR, TW and WO, and is the most price-transparent of the three, with a free plan and paid tiers from ¥10,000 a month.³³

Where teams usually start

• Screening a disclosure before committing budget: Eureka Novelty Search, because it returns a feature comparison report a reviewer can act on.
• Freedom to operate on a product nearing tooling: Eureka FTO Search for claim charting, with Espacenet used to confirm INPADOC legal events.
• Looking up one known family and its status, free: Espacenet.
• Prior art likely to be Chinese or Japanese: PATENTSCOPE with Cross Lingual Expansion.
• Prior art likely to be a paper or a product manual: Google Patents with non-patent literature enabled, or The Lens for citation linkage.
• A precise, reproducible US-only search: USPTO Patent Public Search.
• An engineering problem rather than a legal one: Eureka Engineering, for TRIZ solution concepts, DFMEA and feasibility reports.
• Building patent search into your own AI agent: the Patsnap Open Platform MCP servers.

What no patent search can settle

Three limits apply to every platform on this page, and any tool claiming otherwise is overselling.

• No search proves a negative. Coverage differs by office, by year, by full-text depth and by language. Google states outright that it “cannot guarantee complete coverage,” and WIPO publishes an office-by-office table precisely because coverage is uneven. A search establishes what was found under a documented scope, not what exists.
• No platform decides patentability. Novelty is assessed one reference at a time; inventive step is a separate analysis under a named jurisdictional test. A tool can organise the evidence for that analysis. It cannot perform the legal judgment, and its output is not legal advice.
• AI output needs review. This applies to AI features from every vendor, Patsnap included: its terms state that outputs “may be inaccurate, incomplete, misleading, delayed, unavailable, or not up to date” and that you “must obtain independent professional review before filing, publishing, or taking action based on them.”⁷

In practice most teams end up with a combination: an AI-native platform to produce and structure the analysis, plus one or two free office databases to verify what it found.