Clinical RAG — Medical Case Retrieval System

A Retrieval-Augmented Generation (RAG) system that takes an unstructured patient complaint as input, retrieves similar past clinical cases from a corpus of 30,000 notes, and generates a grounded clinical response.

---

Architecture

Patient Query (plain text)
        ↓
   FastAPI /query
        ↓
  Embed query → Qdrant pre-filter + semantic search
        ↓
  Deduplicate → Fetch full notes from document store
        ↓
  LLM generation (NVIDIA / Llama 3.1 70B)
        ↓
  Grounded clinical response

Small-to-Big Retrieval Strategy

Layer	Field	Role
Filter	summary	Pre-filter by age, sex
Search	conversation	Small chunks — precise matching
Context	full_note	Full document — rich LLM context

---

Tech Stack

Component	Technology
Backend	FastAPI + Uvicorn
Vector DB	Qdrant
Embeddings	sentence-transformers/all-MiniLM-L6-v2
LLM	meta/llama-3.1-70b-instruct (NVIDIA Build)
Dataset	Vinay393/augmented-clinical-notes (30k rows)
Containerization	Docker + Docker Compose

---

Project Structure

RAG/
├── Dockerfile
├── docker-compose.yml
├── requirements.txt
├── .env.example
└── app/
    ├── main.py
    ├── config.py
    ├── ingestion/
    │   ├── ingest.py
    │   ├── chunker.py
    │   ├── document_store.py
    │   ├── indexer.py
    │   └── summary_parser.py
    ├── services/
    │   ├── retriever.py
    │   ├── generator.py
    │   └── rag_pipeline.py
    ├── routes/
    │   └── query.py
    └── models/
        └── schemas.py

---

Quickstart

1. Clone and configure

git clone <your-repo>
cd RAG
cp .env.example .env
# Add your NVIDIA_API_KEY to .env

2. Start Qdrant

docker compose up -d qdrant

3. Run ingestion

docker compose run --rm ingest

4. Start the API

docker compose up -d api

5. Test

http://localhost:8000/docs     ← Swagger UI
http://localhost:8000/health   ← Health check

---

Example Request

POST /query
{
  "patient_query": "severe knee pain for two weeks, worse when walking",
  "filters": {
    "patient_sex": "Male",
    "patient_age_group": "middle_aged"
  },
  "top_k": 3
}

---

Environment Variables

Variable	Description
NVIDIA_API_KEY	NVIDIA Build API key
NVIDIA_MODEL	LLM model name
QDRANT_URL	Qdrant connection URL
EMBEDDING_MODEL	Sentence transformer model
DOCUMENT_STORE_PATH	Path to document store JSON

See .env.example for all variables.

---

Dataset

AGBonnet/augmented-clinical-notes — 30,000 clinical case rows with fields: idx, note, full_note, conversation, summary. License: MIT.