Cypris

Innovation Intelligence, Briefly

• Cypris is a research intelligence platform that searches patents, scientific papers, chemical data and market sources together, and applies task-specific AI agents to the results — built for R&D and IP teams rather than for general search.
• The corpus is the differentiator: 180M+ patent documents from 150+ countries, 280M+ papers from 20,000+ journals, 120M+ indexed chemical compounds, and 1M+ tracked market sources.
• Agents are scoped by task — technology scouting, literature review, competitive intelligence, prior art, regulatory research and white space identification — rather than one general chat box.
• It joins internal and external knowledge, which is the part most competitors do not attempt: your own project history alongside the global record.
• Named customers include Johnson & Johnson, Honda, Yamaha, NASA and Philip Morris International, and it sells contact-sales to enterprises.

Corpus, Agents And Assurance

IP And R&D Terms Used Here

Five terms carry the weight in this field and mean something precise rather than general.

Prior art

Any existing public disclosure — patent, paper, product, conference talk — that may invalidate a patent claim. Missing one is the defining failure of a search.

Freedom to operate

An assessment of whether commercialising something would infringe existing rights. A legal opinion, informed by search but not produced by it.

White space

An area of technology with little existing patent or research activity, sought as an opportunity. Reasoning about absence, which is much harder than retrieval.

Technology scouting

Systematically looking outside your organisation for relevant developments, partners or acquisition targets.

Markush claim

A chemistry patent claim covering a family of related compounds through a general structure rather than listing each one — a well-known reason chemical patent search is specialised.

Why Patent Search Is Genuinely Hard

It is worth explaining what makes this a real problem, because from outside it looks like search with a bigger index.

Patents are written to be simultaneously legally precise and as broad as the examiner will allow. That produces prose written by lawyers to avoid the obvious word. A patent covering a bicycle brake may never use the phrase bicycle brake — it describes a friction-based deceleration mechanism for a wheeled personal conveyance.

So keyword search fails structurally rather than incidentally. The document you most need is the one that never used your term, and you cannot know it exists to look for it.

The second difficulty is scale with consequences. 180 million documents is not a corpus you sample; in prior art work, missing one document is the entire failure mode. A patent granted over undisclosed prior art can be invalidated later, which is a far more expensive outcome than not filing.

Literature tools like Elicit exist for exactly that half of the problem. The third is that the answer often is not in the patents at all. A published paper, a conference abstract or a product datasheet can be prior art. Searching patents alone leaves a gap that only shows up in litigation.

That combination — obfuscated language, unforgiving scale, and evidence scattered across document types — is exactly the shape of problem semantic search and a domain ontology are suited to, and it is why this category can charge enterprise rates.

The document that sinks your patent is the one that never used your words. Keyword search cannot find it, and that is the whole reason this category exists.

Task-Scoped Agents Beat One Chat Box

Cypris splits its AI into agents named for specific jobs rather than offering a single assistant, and that is a more meaningful design decision than it first appears.

These tasks look similar and differ sharply in what a good answer means. Technology scouting rewards breadth — you want to see everything adjacent, including things you would not have thought to ask about. Prior art rewards exhaustiveness in a narrow band, where a miss is the failure. White space requires reasoning about absence, which is a genuinely different operation from retrieval.

A single general assistant tuned for helpfulness will optimise for a satisfying answer, and a satisfying answer is the wrong target for prior art. What you want there is an uncomfortable one.

The retrieval-augmented approach with sentence-level grounding matters for the same reason. In this field an unsourced assertion is not merely unhelpful — it is unusable, because the output frequently ends up supporting a filing decision or a legal position.

The stated official API partnerships with OpenAI, Anthropic and Google — the providers behind Claude and its peers — are worth noting for a specific reason: enterprise R&D buyers care about where their queries go. Formal partnerships with zero-retention terms are a different posture from consumer API access, and it is the sort of detail that clears a procurement review.

Joining Internal Knowledge To The Global Record

The knowledge management piece is the most underrated part of the platform and the hardest to replace with anything else.

Large R&D organisations have a chronic and expensive problem: they do not know what they already know. Work is done, written up, filed in a project system, and the person who did it moves team or leaves. Three years later a different group runs the same experiment because there was no way to discover the first one.

External databases cannot help with this. No amount of patent coverage tells you that a team in another division tried this in 2021 and abandoned it for a reason that is still valid.

Combining both in one search is a genuinely different capability from either alone. It means a technology scouting query can return the global state of the art alongside your own organisation's prior attempts, which is the actual context a decision needs.

It also raises the question to settle first in any deployment: getting internal knowledge into the system requires it to exist in a usable form. Organisations whose institutional memory is in individual inboxes and personal drives will find that the data preparation, not the software, is the project.

1. Pick a technology question your team already researched thoroughly and knows the answer to.
2. Run it through the platform and check whether it finds what your analysts found — and what they missed.
3. Deliberately test a prior art query in an area where you know of an obscure non-patent reference.
4. Assess how internal project data would actually be ingested, and who would maintain it.
5. Confirm the data residency and retention terms with your own legal team, not from the marketing page.

Procuring It

This site does not publish plan prices, and Cypris does not either. The structure is what you can plan around.

The model is contact-sales enterprise licensing, appearing to be seat-based with tiered levels of access. There is no self-serve entry point, which is consistent with the customer list and the data licensing underneath.

That last point explains the shape. A platform indexing 180 million patents and 280 million papers is paying for data access, and those upstream costs are why nothing in this category is cheap or metered casually.

Two structural variables usually drive the quote: how many seats, and which data collections and agent capabilities are included at each tier. Chemical and regulatory data in particular are often separately licensed rather than universally included.

The Research Briefs offering is worth understanding as a distinct line item — analyst-supported custom reports are a service, not software, and they are priced and scheduled accordingly.

For anyone comparing this against a general research tool, the honest framing is that you are not paying for the AI. You are paying for licensed access to corpora that are expensive to assemble and maintain, with the AI as the interface to them.

What Still Needs A Human

Three things, and the first is a legal distinction that matters enormously.

A search is not an opinion. Drafting the filing itself is a separate product again — Solve Intelligence occupies that end. Freedom-to-operate and patentability searches inform a legal conclusion; they are not the conclusion. That judgement is made by a qualified attorney who takes professional responsibility for it, and no platform output substitutes for that.

Absence of evidence is the hardest result to trust. White space analysis finds gaps, and a gap in the data can mean nobody has done this, or it can mean everyone tried and failed quietly, or that the work exists in a form the corpus does not cover. Distinguishing those requires domain expertise the tool does not have.

Coverage is never total. Trade secrets are, by design, not in any database. Recent filings are unpublished for eighteen months. Non-English and non-patent literature coverage varies. A confident empty result is the output to be most suspicious of.

Used properly, this compresses the weeks of searching that precede expert judgement — which is a large and genuine saving. It does not compress the judgement itself, and organisations that treat it as though it does are taking a risk the software cannot carry.

• Freedom-to-operate and patentability conclusions remain attorney work
• Recent filings stay unpublished for around eighteen months and are invisible to everyone
• Trade secrets are absent from every database by definition
• An empty white space result may reflect a coverage gap rather than an opportunity
• Internal knowledge is only searchable if it was written down in the first place