Open Reaction Database

The Open Reaction Database (ORD) is a shared, open home for chemical reaction data, created so that machine learning models have consistent, structured examples to learn from. Each record captures a reaction in full: the inputs and their amounts, conditions such as temperature and time, the workup and product, and the measured outcomes, all held in a common schema rather than free-text lab notes. That structure is the whole point, because reaction data is usually trapped in incompatible formats scattered across different labs.

You can browse and search reactions through the web interface, and for bulk work the full dataset is published openly: the records are stored as structured protocol-buffer messages in a public repository, with a Python package for reading and writing them. For RAG, the practical route is to render each reaction into a readable passage (reactants, reagents, conditions, and yield in plain language) so a retriever can match a natural-language question to a concrete example, while keeping the structured original alongside for anything that needs exact values.

It suits chemistry and drug-discovery systems: tools that propose synthesis routes, answer questions about how a reaction is run, or help design experiments by surfacing precedents. Because contributions come from pharmaceutical companies, universities, and technology firms, it is meant to be a neutral, community-owned resource rather than one group's private stash.

Two caveats matter. Coverage is uneven and still growing, so it is nowhere near a complete record of known chemistry, and the level of detail varies between contributed batches; treat a missing reaction as no evidence rather than proof it will not work. On licensing, the database is under CC BY-SA 4.0: commercial use is fine and attribution is required, but the share-alike clause means any derivative database you distribute must carry the same licence, so factor that in before you redistribute.

Pair it with sources like ChEMBL or the Materials Project when you need bioactivity or materials data; ORD is the one to reach for when the reactions themselves are the subject.