Patent and intellectual property collections: a large, structured, largely public-domain corpus that is under-used in RAG applications.
6sources
Patents are one of the most under-used sources in RAG, and that is a shame, because they are a huge, carefully structured, and largely public-domain body of technical writing. Every granted patent describes an invention in detail, with a title, an abstract, formal claims, and consistent metadata like inventors, dates, and classification codes. For a RAG system doing prior-art searches, competitive research, or answering deep technical questions, that combination of depth and structure is hard to find anywhere else.
Choosing within this category depends on whether you want reach or convenience. Google Patents Public Datasets and PatentsView give you bulk, query-ready access to United States and international data with the fields already parsed, which is ideal for building at scale. Official offices like the EPO’s Espacenet (the European Patent Office) and WIPO’s PATENTSCOPE (the World Intellectual Property Organization) offer the broadest and most authoritative coverage, but often through APIs with rate limits or credentials rather than a simple bulk download. Research sets like HUPD are smaller and pre-shaped for machine learning.
Two quirks are worth knowing. First, the text of a granted patent is generally not copyrighted, but the drawings, cited non-patent literature, and some database compilations can carry their own restrictions, so read the access terms rather than assuming everything is free. Second, patents use dense legal-technical language and their own classification schemes, so plan to keep the classification codes as metadata for filtering, since they are far more reliable than keyword matching.
The sources below range from bulk public datasets to the official patent office platforms.
A corpus of 1.3 million US utility patents filed between 1971 and 2018, each paired with its human-written abstract as a gold-standard summary and organised by Cooperative Patent Classification code. A large, clean patent text corpus built for abstractive summarisation and other patent NLP work.
Free access to over 140M patent documents from the European Patent Office (EPO). Its Open Patent Services text analysis tools have made patent full texts much easier to reach, though the terms are open access rather than an open reuse licence.
A query-based collection of over 120M patent documents from more than 100 patent offices worldwide, including applications, pre-grant publications, and granted patents. Accessible through Google BigQuery, with hundreds of millions of USPTO events also queryable.
A large-scale, structured corpus of US patent applications built specifically for machine learning and natural language processing research. It fills the gap left by mainstream patent search tools, which are not designed with the ML and NLP community in mind.
A platform from the US Patent and Trademark Office (USPTO) for exploring US patent data, with visualisation and analysis tools. It connects patents, inventors, organisations, and locations in a regularly updated database, and offers both an API and complete bulk data tables.
A search service for international patent collections, maintained by the World Intellectual Property Organization (WIPO). It lets you search across patent applications filed under international treaties alongside national collections from many countries.