ChEMBL

ChEMBL connects three things that usually live in separate databases: chemical structures, their measured biological activity, and the protein targets and genes those measurements relate to. It holds close to 2.5M compound records on nearly 2M unique structures, along with tens of millions of activity readings, drawn mainly from the medicinal chemistry literature. New releases come out roughly every three to four months.

You can download the full database from EMBL-EBI as a relational dump for Oracle, PostgreSQL, or (easiest for a laptop) SQLite, and there is also a set of flat files, an RDF version, and a REST API for targeted lookups. For RAG the relational shape is the thing to plan around: the interesting facts live in joins between compounds, assays, and targets, so you will usually write queries that flatten those relationships into readable sentences before you chunk and embed them, rather than dumping raw tables at your model.

ChEMBL suits drug-discovery and medicinal-chemistry RAG: answering what activity a compound has shown against a given target, which molecules hit a target of interest, or how a series behaves across assays. Because every data point is abstracted from the literature and curated by hand, confidence per record is high, which matters when the answers feed clinical or chemistry decisions.

The trade-off is coverage. Hand curation makes ChEMBL smaller than automated collections such as PubChem, so reach for it when trust beats breadth and combine the two when you need both. On licensing, CC BY-SA 3.0 lets you use the data commercially with attribution, but the share-alike term means any derivative database you distribute has to carry the same licence. Whether embeddings (numeric vectors derived from the text) count as a derivative under share-alike is genuinely unsettled, so factor that ambiguity into your plans.

Pair ChEMBL with ChEBI for consistent compound naming, with PubChem for raw coverage, and with DrugBank when you need approved-drug and target detail.