What is clinical-rag?

CLINICAL-RAG is a question and answer assistant that lets healthcare workers like clinicians, nurses, and care coordinators ask plain-English questions about patient discharge summaries and get accurate, source-cited answers instead of digging through long unstructured notes themselves. It is built as a retrieval-augmented generation system, meaning it first finds the most relevant pieces of a patient's notes, then hands only those pieces to an AI language model to write the actual answer, rather than letting the model answer purely from its own training. For finding relevant note sections, the project uses a technique called TF-IDF, a method that matches exact terms and phrases rather than general meaning. The README explains this choice by noting that clinical notes are full of precise terminology, like drug names, lab values, and medical codes, where exact matching works better than general-purpose approaches, and it requires no GPU and no extra API calls to run. Once the right note chunks are found, they are sent to Anthropic's Claude language model, which is instructed to answer strictly from those retrieved chunks so it does not invent clinical details that are not actually in the records. The project ships with 8 realistic but entirely synthetic patient discharge summaries covering conditions like diabetes, heart attack, pneumonia, stroke, and hip fracture, written in a style similar to real hospital documentation but containing no real patient data. A built-in evaluation panel tests the retrieval system against a set of sample questions and the README reports 100 percent top-1 retrieval accuracy on that test set. To run it, you install the Python dependencies, run two scripts to generate the sample notes and build the search index, then launch a Streamlit web app. The app needs a free Anthropic API key entered in its sidebar to generate AI answers, though the retrieval evaluation panel works without one.