MIMIC-III / MIMIC-IV
MIMIC (Medical Information Mart for Intensive Care) is one of the most widely used open clinical datasets in the world. It brings together de-identified records from tens of thousands of intensive care unit stays at a single Boston teaching hospital, covering vital signs sampled through a stay, laboratory results, prescribed medications, procedures, billing codes, and, most valuable of all, the free-text notes that nurses and doctors wrote at the bedside. Those notes (discharge summaries, radiology reports, and nursing observations) are the part most people come for, because they are exactly the messy, abbreviation-heavy clinical prose a healthcare RAG system has to reason over.
There are two generations. MIMIC-III is the long-established release that underpins thousands of published papers, so if you want to reproduce or compare against prior work it is the safe choice. MIMIC-IV, hosted at https://physionet.org/content/mimiciv/, has an updated relational structure, more recent admissions, and a cleaner split between hospital and ICU modules. Both arrive as a set of CSV tables that load straight into PostgreSQL or DuckDB. For RAG, the usual pattern is to pull the note tables, chunk each note by section, embed the passages, and index them so a model can retrieve a patient's history at query time.
Access is the real hurdle. This is not a straight download: you must register as a credentialed PhysioNet user, complete recognised human-subjects research training (the CITI programme is the common route), and sign a data-use agreement. Approval can take days to weeks, so start it before you need the data. The licence forbids any attempt to re-identify patients and forbids redistributing the raw records, so you cannot ship the dataset inside a product or feed raw notes to an external LLM API without checking the terms first.
For clinical work, pair it with a controlled vocabulary such as UMLS or SNOMED CT for entity linking, and treat MIMIC as a development and evaluation corpus rather than a knowledge base you deploy verbatim.
Related sources
AlphaFold Protein Structure Database
EMBL-EBI and Google DeepMind's open database of AI-predicted protein structures, covering over 200 million sequences across almost all of UniProt. Each model ships with per-residue confidence and predicted aligned error scores, and is reachable by website, API, FTP and Google Cloud bulk access. Released under CC BY 4.0 and designed to pair with UniProt and the experimental structures in the PDB.
ClinicalTrials.gov
A US National Institutes of Health database of privately and publicly funded clinical studies. It holds structured records on more than 450,000 studies, including protocols, conditions, interventions, outcomes, and results, all in the public domain.
DrugBank
A freely accessible database that combines detailed drug data with drug target information, covering more than 15,000 drug entries. It links chemistry, pharmacology, and biology in one place, which makes it a strong knowledge source for medical and pharmaceutical RAG.
MIRIAD
A million-scale medical instruction and retrieval dataset of roughly 5.8 million question-answer pairs, each grounded in a passage from peer-reviewed biomedical literature. Built for medical RAG, retrieval, and hallucination detection.