Google Patents Public Datasets

Google Patents Public Datasets is one of the broadest patent collections you can query in a single place, gathering more than 120M patent documents from over 100 offices worldwide. It spans applications, pre-grant publications, and granted patents, with the bibliographic data, full text, classifications, and citations attached. On top of the documents sit hundreds of millions of USPTO events, including the office actions exchanged between applicants and examiners, so you can trace how a patent moved through examination rather than just reading its final form.

You reach all of it through BigQuery, Google's cloud data warehouse, where the collection lives as SQL-queryable tables. For RAG that is a genuine advantage: you can narrow to exactly the patents you want by office, date, classification code, assignee, or keyword, then export a focused slice instead of wrestling with raw bulk files. A common pattern is a single BigQuery job that pulls the abstracts and claims for your domain, which you then chunk and embed for a vector database.

The data is free under CC BY 4.0, so commercial use is fine provided you attribute it. The cost to watch is compute, not licensing: BigQuery bills for the bytes each query scans, and these tables are large, so an unfiltered `SELECT *` can run up a bill quickly. Select only the columns and partitions you need, and lean on preview and sampling while you develop.

Two more caveats. Patent prose is dense, formulaic legalese that does not always chunk or embed cleanly, so budget real time for preprocessing. And both coverage and text quality vary by office and era, with older and non-English documents patchier than recent US and European filings.

For focused US machine-learning work, the Harvard USPTO Patent Dataset (HUPD) is a tidier, ready-made alternative. For the widest international coverage in one queryable place, or for studying examination history, this is the stronger pick.