June 21, 2026

•

XX

min read

Prior art search for artificial intelligence and machine learning inventions is one of the hardest retrieval problems in patent work, for reasons specific to how AI knowledge is produced and disclosed. Prior art search establishes whether an invention is novel by finding any earlier disclosure that describes it. In most fields, the relevant disclosures are predominantly patents. In AI and machine learning, the most relevant and most recent disclosures are predominantly non-patent literature: preprints on arXiv, proceedings from conferences such as NeurIPS and ICML, open-source code and model documentation, and technical reports. These sources are published quickly and openly, often well ahead of any corresponding patent, so a prior art search confined to patent databases misses the state of the art.

The volume compounds the difficulty. AI scientific publications more than doubled from about 102,000 in 2013 to more than 242,000 in 2023, growing nearly 20 percent in the final year alone.¹ Patenting has grown even faster from a smaller base: AI patents granted worldwide rose from 3,833 in 2010 to 122,511 in 2023, an increase of almost 30 percent in the last year measured.¹ Generative AI illustrates the velocity of the literature most sharply, with related scientific publications rising from 116 in 2014 to more than 34,000 in 2023 while generative-AI patent families grew more than 800 percent over roughly the same period.² A prior art searcher in this field is therefore working against both a large and a rapidly expanding corpus, split across patent and non-patent sources.

Retrieval quality falls exactly where AI prior art needs it most. Patent retrieval is already harder than general-domain information retrieval, and controlled evaluation shows that cross-domain retrieval, finding relevant art outside the query's own technology area, performs several times worse than in-domain retrieval; one recent family-level benchmark found out-of-domain retrieval roughly five times worse than in-domain across hundreds of controlled configurations.³,⁴ AI and machine-learning methods are applied across many application domains, so relevant prior art for an AI invention is frequently located in a different field than the invention's stated use, which is precisely the cross-domain case where conventional retrieval degrades. This is the technical reason keyword and classification search alone are insufficient for AI prior art, and why dense, semantic methods have become the focus of research on patent prior art retrieval.⁵,⁶

Why AI prior art is distinctively hard

Non-patent literature dominates. The most relevant and most recent AI disclosures appear first in preprints, conference proceedings, and open-source code, so a patent-only search misses the state of the art.

Exploding volume. AI publications more than doubled to over 242,000 in 2023, and AI patents granted rose to 122,511, so the corpus a searcher must cover is both large and expanding rapidly.¹

Cross-domain dispersion. AI methods are applied across many fields, so relevant prior art is often in a different technology area than the invention, which is where retrieval degrades most.³

Fast obsolescence of terminology. AI vocabulary evolves quickly, so keyword search misses conceptually identical work described in newer or different terms.

Software-claim breadth. Algorithmic and software claims can be drafted broadly and abstractly, which makes matching a claim to its closest prior art a conceptual rather than a lexical task.

How semantic search closes the gap

Semantic search addresses each of these problems. It retrieves conceptually relevant disclosures regardless of terminology, which handles both fast-evolving vocabulary and broadly drafted software claims. Applied across both patents and scientific literature in one corpus, it covers the non-patent literature where AI prior art concentrates rather than patents alone. And because dense retrieval encodes meaning rather than surface form, it is better positioned than keyword search for the cross-domain case, retrieving relevant art from a different application area than the invention. Combined with an ontology that organizes retrieval by concept, semantic search returns a structured, high-recall view of the prior art rather than a keyword-limited sample.

Where Cypris fits

Cypris runs semantic prior art search across a corpus of more than 500 million patents and scientific papers, organized through a proprietary R&D ontology. Because the corpus spans both patents and scientific literature, Cypris covers the non-patent literature where AI and machine-learning prior art concentrates, rather than patents alone. Semantic search retrieves conceptually relevant disclosures regardless of terminology, which handles the fast-evolving vocabulary and broadly drafted software claims characteristic of AI inventions, and the ontology organizes retrieval by concept so cross-domain prior art in a different application area is surfaced rather than missed. Cypris Q, the platform's agentic layer, lets teams run and chain prior art and novelty analysis conversationally, and Agentic Monitoring tracks a technology area over time so newly published disclosures are surfaced as they appear, which matters in a field moving as fast as AI. Cypris provides enterprise API partnerships with OpenAI, Anthropic, and Google, and is built with enterprise-grade security. Cypris serves hundreds of enterprise customers across pharmaceuticals, chemicals, advanced materials, energy, and other regulated industries.

FAQ

Why is prior art search hard for AI and machine learning inventions?

Prior art search is hard for AI and machine learning inventions because the most relevant and most recent disclosures are predominantly non-patent literature, such as preprints, conference proceedings, and open-source code, which a patent-only search misses. The corpus is also large and expanding rapidly, and AI methods are dispersed across many application domains. These factors make high-recall, cross-domain retrieval essential.

Why does non-patent literature matter so much for AI prior art?

Non-patent literature matters for AI prior art because AI research is published quickly and openly, often well ahead of any corresponding patent, so the state of the art appears first in preprints, conference papers, and code. A search confined to patent databases misses these disclosures. Effective AI prior art search must cover both patents and scientific literature.

How large is the AI prior art corpus?

The AI prior art corpus is large and growing quickly. AI scientific publications more than doubled from about 102,000 in 2013 to over 242,000 in 2023, and AI patents granted worldwide rose from 3,833 in 2010 to 122,511 in 2023. Generative-AI publications alone grew from 116 in 2014 to more than 34,000 in 2023.

What makes AI prior art retrieval technically difficult?

AI prior art retrieval is technically difficult because AI methods are applied across many domains, so relevant prior art is often in a different technology area than the invention, and cross-domain retrieval performs several times worse than in-domain retrieval. One benchmark found out-of-domain retrieval roughly five times worse than in-domain. Fast-evolving terminology and broadly drafted software claims add further difficulty.

Why is keyword search insufficient for AI prior art?

Keyword search is insufficient for AI prior art because AI terminology evolves quickly and software claims are often drafted broadly and abstractly, so conceptually identical work is described in different terms. Keyword search matches surface form and misses these. Semantic search retrieves by meaning, which is what the task requires.

How does semantic search improve AI prior art search?

Semantic search improves AI prior art search by retrieving conceptually relevant disclosures regardless of terminology, across both patents and scientific literature, and by handling the cross-domain case where relevant art is in a different field. It encodes meaning rather than surface form. Combined with an ontology, it returns a structured, high-recall view of the prior art.

Does AI prior art search need to cover scientific literature?

AI prior art search needs to cover scientific literature because the most relevant and most recent AI disclosures appear there first, in preprints, conference proceedings, and technical reports. Covering patents alone leaves the state of the art unretrieved. Cypris searches both across more than 500 million patents and scientific papers.

Which teams run AI prior art search?

AI prior art search is run by IP, R&D, and patent teams at technology companies and across industries adopting AI, as well as by patent professionals assessing novelty. It is increasingly important as AI patenting grows. Cypris serves hundreds of enterprise customers across research-intensive and regulated industries.

How current does AI prior art search need to be?

AI prior art search needs to be continuously current, because AI research and filings publish constantly and the state of the art shifts quickly. A one-time search reflects only the moment it was run. Cypris uses Agentic Monitoring to track a technology area and surface newly published disclosures as they appear.

Endnotes

Stanford Institute for Human-Centered Artificial Intelligence (2025). Artificial Intelligence Index Report 2025, Chapter 1. arXiv:2504.07139. https://doi.org/10.48550/arxiv.2504.07139
World Intellectual Property Organization (2024). Patent Landscape Report: Generative Artificial Intelligence. Geneva: WIPO. https://doi.org/10.34667/tind.49740
Cavallucci, N., Chibane, I. & Ayaou, M. (2026). DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross-domain patent retrieval. Array. https://doi.org/10.1016/j.array.2026.100720
Lupu, M. (2013). Patent Retrieval. Foundations and Trends in Information Retrieval. https://doi.org/10.1561/1500000027
Stamatis, V. (2022). End to End Neural Retrieval for Patent Prior Art Search. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-030-99739-7_66
Zihayat, M. & Etwaroo, R. (2021). A non-factoid question answering system for prior art search. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2021.114910

‍

Prior Art Search for AI and Machine Learning Inventions in 2026

Writen By:

Cypris Research Team

Table of contents

Register here

Subscribe to receive the latest blog posts to your inbox every week.

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Prior art search for artificial intelligence and machine learning inventions is one of the hardest retrieval problems in patent work, for reasons specific to how AI knowledge is produced and disclosed. Prior art search establishes whether an invention is novel by finding any earlier disclosure that describes it. In most fields, the relevant disclosures are predominantly patents. In AI and machine learning, the most relevant and most recent disclosures are predominantly non-patent literature: preprints on arXiv, proceedings from conferences such as NeurIPS and ICML, open-source code and model documentation, and technical reports. These sources are published quickly and openly, often well ahead of any corresponding patent, so a prior art search confined to patent databases misses the state of the art.

The volume compounds the difficulty. AI scientific publications more than doubled from about 102,000 in 2013 to more than 242,000 in 2023, growing nearly 20 percent in the final year alone.¹ Patenting has grown even faster from a smaller base: AI patents granted worldwide rose from 3,833 in 2010 to 122,511 in 2023, an increase of almost 30 percent in the last year measured.¹ Generative AI illustrates the velocity of the literature most sharply, with related scientific publications rising from 116 in 2014 to more than 34,000 in 2023 while generative-AI patent families grew more than 800 percent over roughly the same period.² A prior art searcher in this field is therefore working against both a large and a rapidly expanding corpus, split across patent and non-patent sources.

Retrieval quality falls exactly where AI prior art needs it most. Patent retrieval is already harder than general-domain information retrieval, and controlled evaluation shows that cross-domain retrieval, finding relevant art outside the query's own technology area, performs several times worse than in-domain retrieval; one recent family-level benchmark found out-of-domain retrieval roughly five times worse than in-domain across hundreds of controlled configurations.³,⁴ AI and machine-learning methods are applied across many application domains, so relevant prior art for an AI invention is frequently located in a different field than the invention's stated use, which is precisely the cross-domain case where conventional retrieval degrades. This is the technical reason keyword and classification search alone are insufficient for AI prior art, and why dense, semantic methods have become the focus of research on patent prior art retrieval.⁵,⁶

Why AI prior art is distinctively hard

Non-patent literature dominates. The most relevant and most recent AI disclosures appear first in preprints, conference proceedings, and open-source code, so a patent-only search misses the state of the art.

Exploding volume. AI publications more than doubled to over 242,000 in 2023, and AI patents granted rose to 122,511, so the corpus a searcher must cover is both large and expanding rapidly.¹

Cross-domain dispersion. AI methods are applied across many fields, so relevant prior art is often in a different technology area than the invention, which is where retrieval degrades most.³

Fast obsolescence of terminology. AI vocabulary evolves quickly, so keyword search misses conceptually identical work described in newer or different terms.

Software-claim breadth. Algorithmic and software claims can be drafted broadly and abstractly, which makes matching a claim to its closest prior art a conceptual rather than a lexical task.

How semantic search closes the gap

Semantic search addresses each of these problems. It retrieves conceptually relevant disclosures regardless of terminology, which handles both fast-evolving vocabulary and broadly drafted software claims. Applied across both patents and scientific literature in one corpus, it covers the non-patent literature where AI prior art concentrates rather than patents alone. And because dense retrieval encodes meaning rather than surface form, it is better positioned than keyword search for the cross-domain case, retrieving relevant art from a different application area than the invention. Combined with an ontology that organizes retrieval by concept, semantic search returns a structured, high-recall view of the prior art rather than a keyword-limited sample.

Where Cypris fits

Cypris runs semantic prior art search across a corpus of more than 500 million patents and scientific papers, organized through a proprietary R&D ontology. Because the corpus spans both patents and scientific literature, Cypris covers the non-patent literature where AI and machine-learning prior art concentrates, rather than patents alone. Semantic search retrieves conceptually relevant disclosures regardless of terminology, which handles the fast-evolving vocabulary and broadly drafted software claims characteristic of AI inventions, and the ontology organizes retrieval by concept so cross-domain prior art in a different application area is surfaced rather than missed. Cypris Q, the platform's agentic layer, lets teams run and chain prior art and novelty analysis conversationally, and Agentic Monitoring tracks a technology area over time so newly published disclosures are surfaced as they appear, which matters in a field moving as fast as AI. Cypris provides enterprise API partnerships with OpenAI, Anthropic, and Google, and is built with enterprise-grade security. Cypris serves hundreds of enterprise customers across pharmaceuticals, chemicals, advanced materials, energy, and other regulated industries.

FAQ

Why is prior art search hard for AI and machine learning inventions?

Prior art search is hard for AI and machine learning inventions because the most relevant and most recent disclosures are predominantly non-patent literature, such as preprints, conference proceedings, and open-source code, which a patent-only search misses. The corpus is also large and expanding rapidly, and AI methods are dispersed across many application domains. These factors make high-recall, cross-domain retrieval essential.

Why does non-patent literature matter so much for AI prior art?

Non-patent literature matters for AI prior art because AI research is published quickly and openly, often well ahead of any corresponding patent, so the state of the art appears first in preprints, conference papers, and code. A search confined to patent databases misses these disclosures. Effective AI prior art search must cover both patents and scientific literature.

How large is the AI prior art corpus?

The AI prior art corpus is large and growing quickly. AI scientific publications more than doubled from about 102,000 in 2013 to over 242,000 in 2023, and AI patents granted worldwide rose from 3,833 in 2010 to 122,511 in 2023. Generative-AI publications alone grew from 116 in 2014 to more than 34,000 in 2023.

What makes AI prior art retrieval technically difficult?

AI prior art retrieval is technically difficult because AI methods are applied across many domains, so relevant prior art is often in a different technology area than the invention, and cross-domain retrieval performs several times worse than in-domain retrieval. One benchmark found out-of-domain retrieval roughly five times worse than in-domain. Fast-evolving terminology and broadly drafted software claims add further difficulty.

Why is keyword search insufficient for AI prior art?

Keyword search is insufficient for AI prior art because AI terminology evolves quickly and software claims are often drafted broadly and abstractly, so conceptually identical work is described in different terms. Keyword search matches surface form and misses these. Semantic search retrieves by meaning, which is what the task requires.

How does semantic search improve AI prior art search?

Semantic search improves AI prior art search by retrieving conceptually relevant disclosures regardless of terminology, across both patents and scientific literature, and by handling the cross-domain case where relevant art is in a different field. It encodes meaning rather than surface form. Combined with an ontology, it returns a structured, high-recall view of the prior art.

Does AI prior art search need to cover scientific literature?

AI prior art search needs to cover scientific literature because the most relevant and most recent AI disclosures appear there first, in preprints, conference proceedings, and technical reports. Covering patents alone leaves the state of the art unretrieved. Cypris searches both across more than 500 million patents and scientific papers.

Which teams run AI prior art search?

AI prior art search is run by IP, R&D, and patent teams at technology companies and across industries adopting AI, as well as by patent professionals assessing novelty. It is increasingly important as AI patenting grows. Cypris serves hundreds of enterprise customers across research-intensive and regulated industries.

How current does AI prior art search need to be?

AI prior art search needs to be continuously current, because AI research and filings publish constantly and the state of the art shifts quickly. A one-time search reflects only the moment it was run. Cypris uses Agentic Monitoring to track a technology area and surface newly published disclosures as they appear.

Endnotes

Stanford Institute for Human-Centered Artificial Intelligence (2025). Artificial Intelligence Index Report 2025, Chapter 1. arXiv:2504.07139. https://doi.org/10.48550/arxiv.2504.07139
World Intellectual Property Organization (2024). Patent Landscape Report: Generative Artificial Intelligence. Geneva: WIPO. https://doi.org/10.34667/tind.49740
Cavallucci, N., Chibane, I. & Ayaou, M. (2026). DAPFAM: A Domain-Aware Family-level Dataset to benchmark cross-domain patent retrieval. Array. https://doi.org/10.1016/j.array.2026.100720
Lupu, M. (2013). Patent Retrieval. Foundations and Trends in Information Retrieval. https://doi.org/10.1561/1500000027
Stamatis, V. (2022). End to End Neural Retrieval for Patent Prior Art Search. Lecture Notes in Computer Science. https://doi.org/10.1007/978-3-030-99739-7_66
Zihayat, M. & Etwaroo, R. (2021). A non-factoid question answering system for prior art search. Expert Systems with Applications. https://doi.org/10.1016/j.eswa.2021.114910

‍

Keep Reading

July 12, 2026

•

XX

min read

An ontology is a formal, machine-readable specification of the concepts in a domain and the relationships among them. The term has a precise meaning in knowledge representation: an explicit specification of a conceptualization,¹ that is, a defined vocabulary of entity types, attributes, and relations, together with constraints on how they may be combined. This distinguishes an ontology from a flat taxonomy, which only arranges terms hierarchically; an ontology also encodes non-hierarchical relations, such as a material being used in a process or a method being applied to a claim. In R&D and patent intelligence, the ontology defines the domain schema: the technologies, materials, methods, claims, organizations, and research areas that matter, and the relationship types that connect them.²

A knowledge graph instantiates that schema over real data. It represents information as a graph of nodes and typed edges, commonly expressed as subject-predicate-object triples, linking specific patents, scientific papers, assignees, inventors, technologies, and materials as connected entities rather than isolated documents. Building the graph requires several engineering steps that determine its quality: named-entity recognition and relation extraction to convert unstructured patent and paper text into triples; entity resolution to normalize the many surface forms of an organization, inventor, or compound to a single canonical node; and provenance tracking so every assertion in the graph traces back to the source document that supports it. The result is a structure that can be queried declaratively, for example with a graph query language, and that supports multi-hop traversal, so a question can follow chains of relationships rather than matching a single string.

This structure matters because patents and scientific literature become intelligence only when their relationships are made explicit. A ranked list of relevant documents does not state how a technology area is organized, which organizations are active, how research connects to patents, or where the graph is sparse. An ontology-backed knowledge graph makes those relationships first-class and queryable. A team can ask how two technologies relate, which body of research underpins a patent cluster, which assignees co-file in an area, or where a domain is unclaimed, and receive an answer computed over structured connections rather than assembled by reading.

The 2026 relevance is that structured knowledge is the most reliable way to ground generative AI. Large language models produce fluent output but can assert unsupported claims when they generate from parametric memory over unstructured text. Retrieval-augmented generation (RAG), which conditions a model's output on retrieved external evidence, was introduced to address this and improves factual accuracy on knowledge-intensive tasks.³,⁴ Graph retrieval-augmented generation (GraphRAG) extends RAG by retrieving connected subgraphs rather than isolated passages, so the model reasons over entities and their relationships and can answer questions that require traversing multiple hops.⁵,⁶ Grounding a system on an ontology-backed knowledge graph constrains its outputs to real, connected entities, which is essential for patent and R&D work where every conclusion must trace to actual patents and papers, and where retrieval quality directly governs the reliability of downstream generation.⁷ It is also what makes agentic workflows dependable: an agent reasoning over a structured, provenance-tracked graph produces results a team can verify against sources.

What an ontology and knowledge graph add to patent intelligence

Multi-hop reasoning over relationships. A knowledge graph answers relational and multi-hop questions, such as how two technologies connect through shared materials or which research a patent cluster builds on, rather than only returning documents that match a query string.

Concept-organized semantic search. Dense semantic retrieval returns conceptually relevant documents; the ontology organizes that retrieval within a domain schema, improving both recall and the interpretability of results by grouping them under defined concepts.

White space analysis. White space analysis depends on clustering activity by concept to expose genuine gaps. Clustering patents and papers over the ontology's relationship structure exposes real conceptual sparsity rather than gaps that are artifacts of divergent terminology.

Entity-resolved attribution and competitive intelligence. Entity resolution normalizes assignee and inventor variants to canonical nodes, which lets the graph attribute filings and research accurately and build co-assignee and citation networks rather than a document list.

Provenance-grounded AI. The ontology and knowledge graph give AI agents a structured, provenance-tracked foundation to reason over, which improves the accuracy of agentic analysis and makes its results traceable to the specific patents and papers that support them.

Where Cypris fits

Cypris organizes a corpus of more than 500 million patents and scientific papers through a proprietary R&D ontology. That ontology is the core of the platform: it defines how technologies, claims, materials, methods, and research relate, so Cypris reasons over an entity-resolved relationship structure rather than only matching keywords. This structure powers dense semantic retrieval organized by concept, white space analysis that exposes genuine conceptual gaps, and competitive intelligence that attributes activity to canonical organizations and technology areas. Cypris Q, the platform's agentic layer, reasons over this provenance-tracked foundation, which is what makes its multi-step analysis both reliable and traceable to real patents and papers, consistent with graph-grounded retrieval approaches. Agentic Monitoring tracks a technology area over time against the same structure. Cypris provides enterprise API partnerships with OpenAI, Anthropic, and Google, so AI agents can query the structured corpus programmatically, and it is built with enterprise-grade security. Cypris serves hundreds of enterprise customers across pharmaceuticals, chemicals, advanced materials, energy, and other regulated industries.

FAQ

What is an ontology in R&D and patent intelligence?

An ontology in R&D and patent intelligence is a formal, machine-readable specification of the concepts in the domain and the relationships among them, defined as an explicit specification of a conceptualization. It sets out the entity types, such as technologies, materials, methods, and claims, and the relations that connect them. This lets a platform reason over connections between patents and scientific literature rather than treating documents as isolated.

How is an ontology different from a taxonomy?

An ontology differs from a taxonomy in expressiveness. A taxonomy arranges terms in a hierarchy, while an ontology also encodes non-hierarchical, typed relationships and constraints, such as a material being used in a process. This richer structure is what allows multi-hop reasoning across patents and research rather than simple category lookup.

What is a knowledge graph for patents?

A knowledge graph for patents represents patents, scientific papers, assignees, inventors, technologies, and materials as nodes connected by typed edges, commonly expressed as subject-predicate-object triples. It applies an ontology's schema to real data so relationships are explicit and queryable. This turns a document collection into a structure that supports declarative queries and multi-hop traversal.

How is a knowledge graph built from patent text?

A knowledge graph is built from patent text through named-entity recognition and relation extraction to convert unstructured text into triples, entity resolution to normalize variant names to canonical nodes, and provenance tracking so each assertion links back to its source document. The quality of these steps determines the reliability of the graph. Poor entity resolution, for example, fragments an organization across many nodes and distorts attribution.

Why do knowledge graphs matter for AI in patent research?

Knowledge graphs matter for AI in patent research because they ground generative models on real, connected entities, which improves accuracy and traceability. A model generating from unstructured text alone can assert unsupported claims, whereas one conditioned on a provenance-tracked graph constrains its answers to actual patents and papers. This is essential where conclusions must be verifiable.

What is GraphRAG and how does it differ from standard RAG? GraphRAG is graph retrieval-augmented generation. Standard RAG retrieves isolated text passages to condition a model's output, while GraphRAG retrieves connected subgraphs, so the model reasons over entities and their relationships and can answer multi-hop questions. This suits patent intelligence, where questions often require traversing links between technologies, research, and organizations.

How does an ontology improve white space analysis?

An ontology improves white space analysis by clustering patents and papers over defined relationships rather than by exact keywords, which exposes genuine conceptual gaps instead of gaps that are only artifacts of differing terminology. Because the sparsity reflects the domain structure, the identified white space corresponds to unclaimed technical territory. Cypris organizes its corpus of more than 500 million patents and scientific papers through a proprietary R&D ontology for this purpose.

How do knowledge graphs reduce AI hallucination in patent work?

Knowledge graphs reduce AI hallucination in patent work by constraining a model's outputs to real, connected entities with tracked provenance rather than letting it generate from unstructured text. Retrieval-augmented approaches, and graph-based retrieval in particular, condition generation on retrieved evidence, which improves factual accuracy and lets conclusions be traced to sources. This makes results verifiable against the underlying patents and papers.

Is a knowledge graph the same as a vector database?

A knowledge graph is not the same as a vector database. A vector database supports semantic similarity search over embeddings, while a knowledge graph represents explicit, typed relationships between entities. They are complementary: dense retrieval finds relevant documents, and the graph structures how those documents and entities relate. Cypris combines semantic retrieval with a proprietary R&D ontology.

Which teams benefit from ontology-based patent intelligence?

‍Ontology-based patent intelligence benefits R&D, innovation, IP, and strategy teams that need to understand how technologies relate, attribute activity to organizations, and find genuine white space. It is most valuable in research-intensive industries such as pharmaceuticals, chemicals, advanced materials, and energy. Cypris serves hundreds of enterprise customers across these industries.

Endnotes

Gruber, T. R. (1993). A translation approach to portable ontology specifications. Knowledge Acquisition. https://doi.org/10.1006/knac.1993.1008
Gruber, T. R. (1995). Toward principles for the design of ontologies used for knowledge sharing. International Journal of Human-Computer Studies. https://doi.org/10.1006/ijhc.1995.1081
Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS.
Gao, Y. et al. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv:2312.10997. https://doi.org/10.48550/arxiv.2312.10997
Procko, T. & Ochoa, O. (2024). Graph Retrieval-Augmented Generation for Large Language Models: A Survey. https://doi.org/10.1109/aixset62544.2024.00030
Han, S. et al. (2025). A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models. arXiv:2501.13958. https://doi.org/10.48550/arxiv.2501.13958
Chen, J. et al. (2024). Benchmarking Large Language Models in Retrieval-Augmented Generation. AAAI. https://doi.org/10.1609/aaai.v38i16.29728

‍

Knowledge Graphs and Ontologies for R&D and Patent Intelligence in 2026

Blogs

July 11, 2026

•

XX

min read

Patent filings are a leading indicator of competitor R&D direction, and the lead time is a structural consequence of how the patent system operates. An application is filed at its priority date, well before the corresponding product reaches the market, and under the standard 18-month publication rule reflected in USPTO practice and PCT Article 21,⁵ it is not published until roughly eighteen months after that priority date. The interval between when a competitor commits R&D and when the public can observe it is therefore built into the system. The International Energy Agency treats patenting as a leading indicator of technological change in its innovation analysis,¹ and the same logic holds across sectors: a competitor's published filings reveal committed R&D direction ahead of the market, and studies of the linkage between scientific publication and patenting document a measurable lag between the two that compounds the observable lead time.² For R&D and competitive intelligence teams, this makes patents one of the most reliable forward-looking competitive signals available.

Reading that signal well requires structured analysis rather than filing counts, and several technical steps determine its accuracy. First, the unit of analysis should be the patent family, not the individual document, because a single invention generates multiple applications across jurisdictions; counting documents rather than families overstates activity and double-counts international coverage. Second, filings must be located in the technology space using classification codes, principally the Cooperative Patent Classification and International Patent Classification systems, which assign standardized technology categories independent of the applicant's terminology. Third, activity must be attributed through assignee disambiguation, normalizing the many name variants, subsidiaries, and transliterations of an organization to a single canonical entity, because unresolved assignee names fragment a competitor's portfolio and distort the picture. Fourth, the analysis should read the trend over time rather than the latest counts, because the most recent eighteen-to-twenty-four months of data are systematically under-represented by publication lag, so apparent recent declines are usually artifacts rather than real slowdowns.

Two network structures add depth beyond volume. Forward and backward citation analysis situates a competitor's filings in the flow of prior art: backward citations reveal the foundations a filing builds on, and forward citations indicate influence and where a technology is being extended. Co-assignee and knowledge-search network analysis reveals partnerships, academic-industry pipelines, and the coupling between organizations, which shape a competitor's future direction; network-embedding methods over these structures are an established competitive-intelligence technique.³ Scientific literature strengthens the signal further, because research is published before it is patented and patents are filed before products ship, so combining the two sources extends the observable lead time; the scientific footprint within a competitor's filings can be traced through their non-patent references.⁴

What competitor filings reveal

Technology direction. The classification areas where a competitor is filing show where R&D is being committed, often well before those commitments appear in products.

Intensity and momentum. The distribution and rate of change of filing activity across technology areas indicate priorities, and shifts in filing momentum signal changes in strategy earlier than raw counts.

Adjacent moves. Filings in classifications adjacent to a competitor's current products can signal diversification or expansion before it is announced.

Research foundations. The non-patent references and scientific literature a competitor's filings build on show the research base behind their direction, and rising related research is an earlier signal still.

Collaboration structure. Co-assignee patterns and citation coupling reveal partnerships and academic-industry pipelines; network analysis of these relationships is an established competitive-intelligence method.³

How to read competitor R&D direction

Define the competitors and the technology space, scoping the latter with classification codes so the boundary is standardized and reproducible.

Resolve assignees to canonical entities and aggregate to the patent-family level, so activity is attributed accurately and international coverage is not double-counted.

Cluster filings by concept using semantic analysis over the classification and text, so related work groups together regardless of terminology.

Analyze filing momentum as a time series, discounting the most recent windows for publication lag, since direction is visible in trends rather than in the latest bar.

Connect filings to their non-patent references and to the scientific literature, to extend the lead time and expose the research foundations.

Monitor continuously, because competitor direction is revealed by how activity shifts, and continuous monitoring captures those shifts as they publish.

Where Cypris fits

Cypris supports competitive intelligence across a corpus of more than 500 million patents and scientific papers, organized through a proprietary R&D ontology. The ontology and its entity resolution are what turn filings into direction: they normalize assignees to canonical organizations, aggregate to the family level, and cluster activity by concept, so a team sees where a competitor is moving rather than a list of documents. Dense semantic search across patents and scientific literature connects filings to their research foundations, which extends the lead time on the signal, and citation and co-assignee structures expose collaboration and influence. Cypris Q, the platform's agentic layer, lets teams analyze competitor direction conversationally and chain the attribution, clustering, and time-series analysis. Agentic Monitoring is central to this use case: it tracks defined competitors and technology areas over time and flags new filings and research as they publish, so competitive intelligence is continuous rather than a one-time report. Cypris provides enterprise API partnerships with OpenAI, Anthropic, and Google, so AI agents can query the corpus programmatically, and it is built with enterprise-grade security. Cypris serves hundreds of enterprise customers across pharmaceuticals, chemicals, advanced materials, energy, and other regulated industries.

FAQ

How do patent filings reveal competitor R&D direction?

Patent filings reveal competitor R&D direction because an application is filed at its priority date, before the product ships, and is published only about eighteen months later under the standard publication rule. This built-in lag means published filings show committed R&D ahead of the market. Reading the direction requires attributing filings to competitors and technology areas and analyzing where activity concentrates and shifts.

What is the 18-month publication rule?

The 18-month publication rule is the standard practice, reflected in USPTO procedure and PCT Article 21, under which a patent application is published approximately eighteen months after its earliest priority date. It creates a predictable interval between filing and public visibility. It is also why the most recent windows of filing data are under-represented and should not be read as slowdowns.

Why analyze patent families instead of individual documents?

Analyzing patent families instead of individual documents avoids double-counting, because a single invention generates multiple applications across jurisdictions. Counting documents overstates activity and conflates international coverage with genuine volume. The family is the correct unit for measuring how much distinct R&D a competitor is committing.

What role do classification codes play?

Classification codes, principally the Cooperative Patent Classification and International Patent Classification systems, assign standardized technology categories to filings independent of the applicant's wording. They let an analyst locate and compare activity in a technology space reproducibly. This is more reliable than keyword filtering, which varies with drafting style.

Why is assignee disambiguation important?

Assignee disambiguation is important because organizations appear under many name variants, subsidiaries, and transliterations, and unresolved names fragment a competitor's portfolio across multiple entities. Normalizing these to a single canonical entity is what makes attribution and trend analysis accurate. Poor disambiguation systematically distorts competitive intelligence.

How do citation networks support competitive intelligence?

Citation networks support competitive intelligence by situating filings in the flow of prior art. Backward citations reveal the foundations a filing builds on, and forward citations indicate influence and where a technology is being extended. Co-assignee and knowledge-search network analysis additionally reveals partnerships and academic-industry pipelines.

Why combine patents with scientific literature?

Combining patents with scientific literature extends the observable lead time, because research is published before it is patented and patents precede products. Rising research associated with a competitor, followed by early filings, is an earlier and stronger signal than filings alone. The scientific footprint within filings can be traced through their non-patent references.

Why not just count competitor patent filings?

Counting filings alone is misleading because recent counts are depressed by publication lag and raw volume does not indicate direction. The informative signal is which classification areas activity concentrates in and how that distribution changes over time. Family-level aggregation, classification analysis, and time-series momentum are what reveal direction.

Why is continuous monitoring important for competitive intelligence? Continuous monitoring is important because competitor direction is revealed by how activity changes, which a one-time report cannot capture, and because new filings and research publish constantly. A shift in a competitor's focus is only visible if the area is tracked over time. Cypris uses Agentic Monitoring to track competitors and technology areas and flag new activity as it publishes.

Which teams read competitor R&D direction from patents?

Reading competitor R&D direction from patents is done by competitive intelligence, R&D, innovation, strategy, and corporate development teams that need forward-looking awareness of competitor moves. It is most valuable in research-intensive industries such as pharmaceuticals, chemicals, advanced materials, and energy. Cypris serves hundreds of enterprise customers across these industries.

Endnotes

International Energy Agency (2026). The State of Energy Innovation 2026. https://www.iea.org/reports/the-state-of-energy-innovation-2026
Fukuzawa, N. & Ida, T. (2015). Science linkages between scientific articles and patents for leading scientists in the life and medical sciences field. Scientometrics. https://doi.org/10.1007/s11192-015-1795-z
Yang, X. et al. (2024). Predicting patent transaction behaviour based on embedded features of knowledge search networks. Journal of Knowledge Management. https://doi.org/10.1108/jkm-12-2023-1220
Callaert, J., Grouwels, J. & Van Looy, B. (2011). Delineating the scientific footprint in technology: identifying scientific publications within non-patent references. Scientometrics. https://doi.org/10.1007/s11192-011-0573-9
World Intellectual Property Organization, PCT Article 21 (International Publication), and USPTO Manual of Patent Examining Procedure, on patent publication timing.

‍

Reading Competitor R&D Direction from Patent Filings: A Technical Guide for R&D Teams in 2026

Blogs

July 1, 2026

•

XX

min read

Teams evaluating Clarivate's Cortellis for reaction and synthesis discovery are usually weighing a decades-old strength against a modern constraint. Cortellis is deep, trusted, and thorough. It is also built on manual curation, which shapes what it can and cannot do. Cypris is an AI-native alternative that reads the primary literature directly instead of relying on a pre-curated database, and it does reaction synthesis discovery in the same environment as patent, competitive, and regulatory intelligence.

What Cortellis does

Cortellis Drug Discovery Intelligence is Clarivate's flagship preclinical platform, built on the legacy of the Integrity database. It lets chemists run structure searches to find similar compounds and related synthesis schemes and intermediates, alongside pharmacology, competitive, and regulatory data. Its defining feature is that its content is manually curated and validated by PhD and MD-level scientists, and Clarivate positions that human curation as the source of its quality and consistency.

That curation is a real strength. It is also the constraint that leads teams to look for an alternative.

Why teams look for an alternative

Manual curation has three properties built into it. It is slow, because a person reads each source. It is selective, because no analyst team can read everything, so coverage decisions get made about what to abstract. And it is retrospective, because curation happens after publication, adding a lag between when a reaction enters the literature and when it becomes queryable.

For reaction synthesis discovery, those compound. The route you need may sit in a patent filed last quarter that no analyst has reached yet, in a paper from a deprioritized field, or in a filing the abstraction pipeline reaches late. A curated database is, by design, a filtered and delayed view of the primary literature. For most of the last thirty years that was the best available option. It no longer is.

What Cypris does differently

Cypris ingests chemical structure data alongside a corpus of more than 500 million patents and scientific datasets, and its agentic system, Cypris Q, works against the full text of that corpus rather than a pre-abstracted summary of it. Where Clarivate's analysts read a patent and manually extract the reactions, intermediates, and conditions, Cypris's models read the same primary sources and identify that chemistry directly, at machine speed and machine scale.

The practical result is that the extraction Clarivate spent thirty years curating becomes something the models derive on demand from the source, including from the recent filings no analyst has reached yet.

Structure search

Structure search is central to reaction discovery, and Cortellis provides it through exact, similarity, and substructure matching against its curated compound set. Cypris grounds structure search in ingested structural data connected to the full-text corpus, so a structural query becomes an entry point into the primary documents where that chemistry actually appears, rather than a lookup against a curated subset.

One layer instead of a suite of modules

A discovery program does not run on reaction data alone. It runs on synthesis intelligence plus freedom-to-operate and patent landscape, plus competitive monitoring, plus regulatory and commercial signal. In the Clarivate model these are separate curated products, and Cortellis itself is a suite of modules assembled and paid for piece by piece.

Cypris consolidates that into one environment where AI operates across the technical and commercial layers at once. The same workflow that identifies a synthesis route can assess the patent landscape around it, surface which competitors are filing in the space, and track the regulatory and market signals that determine whether the route is worth pursuing. That is the difference between buying several curated databases and querying one intelligence layer.

Where Cortellis still fits

The honest boundary: if a workflow depends on a specific proprietary dataset that exists nowhere in the public or patent literature, a curated platform remains the right tool, and Cypris does not claim otherwise. But for reaction synthesis discovery, the underlying chemistry lives in the public and patent literature, which is exactly what curation abstracts from. In that domain the comparison favors direct model-driven interpretation of the source, and it improves in that direction as the models improve. A curated database advances at the speed of its curation team. An AI-native layer advances at the speed of its models.

The short version

For reaction synthesis discovery run alongside the patent, competitive, and regulatory intelligence that determines whether a route matters, Cypris is the AI-native alternative to Cortellis: it reads the primary literature directly, grounds structure search in the full corpus, and does the technical and commercial work in one layer instead of a stack of curated modules.

FAQ

Is Cypris a direct alternative to Clarivate Cortellis?
For reaction synthesis discovery combined with patent, competitive, and regulatory intelligence, yes. Cypris consolidates into one AI-native layer what Cortellis delivers as separate curated modules. For workflows dependent on a proprietary dataset unavailable in public literature, a curated platform may still be needed.

What is the core difference between Cypris and Cortellis?
Data model. Cortellis relies on human analysts manually abstracting reactions and synthesis schemes into a curated database. Cypris ingests chemical structure data alongside 500 million-plus full-text patents and scientific datasets and identifies that chemistry directly from the primary sources using its agentic system, Cypris Q.

Does Cypris support chemical structure search?
Yes. Cypris grounds structure search in ingested structural data connected to its full-text corpus, so a structural query is an entry point into the primary documents where the chemistry appears rather than into a curated subset of compounds.

What does Cortellis do for reaction synthesis?
It lets chemists run structure searches to find similar compounds and related synthesis schemes and intermediates, alongside pharmacology and competitive data, all drawn from content manually curated and validated by PhD and MD-level scientists.

Why would a team move off a curated database?
Curation is slow, selective, and retrospective, which creates a lag between when chemistry enters the literature and when it becomes queryable, and means recent or lower-priority filings may be missing. Reading the primary corpus directly removes that lag.

Is manual curation still valuable?
For datasets that exist nowhere in public or patent literature, yes. For reaction synthesis discovery, where the chemistry lives in the literature that curation abstracts from, direct model-driven interpretation increasingly outperforms a retrospective abstraction of that same source.

How does Cypris handle recent filings better?
Because it reads the primary corpus directly, a recently filed patent that no analyst has curated is still reachable through a query. Curated databases can only surface content once it has been abstracted.

What does the "single layer" advantage mean in practice?
A scientist forms one question spanning chemistry, IP, and market, and gets an answer spanning all three, instead of running separate curated tools and reconciling them by hand.

Which teams is Cypris the better fit for?
Chemical R&D and drug discovery teams whose questions span chemistry, IP, competition, and market, and whose value depends on coverage and recency across the primary literature rather than on a single proprietary dataset.

What is Cypris Q?
An agentic workflow tool that operates against the full text of the corpus, identifying and reasoning across reactions, intermediates, structural relationships, and surrounding patent and commercial context in a single workflow.

‍

Cypris: An AI-Native Alternative to Clarivate Cortellis for Reaction Synthesis Discovery

Blogs