Clinical RAG System
Production RAG for Multiple Myeloma (medRxiv, 36 citations)
Overview
A production Retrieval-Augmented Generation (RAG) system for clinical decision support in multiple myeloma.
The Problem
Clinicians needed rapid access to the latest multiple myeloma research across thousands of documents while ensuring accuracy and traceability for clinical decision-making.
Technical Architecture
RAG Pipeline
Built using LangGraph for sophisticated orchestration:
- Embeddings: BAAI/bge-large-en-v1.5 (state-of-the-art dense retrieval)
- Generation: Mistral-7B (optimized for clinical text)
- Vector Store: Amazon OpenSearch with hybrid search
- Reranking: Cohere reranker for precision
Document Processing
- Processed 5,000+ clinical documents
- Multi-format support (PDFs, clinical notes, research papers)
- Chunk optimization for clinical context preservation
Results
| Metric | Value |
|---|---|
| Clinical Effectiveness | 88% |
| Documents Processed | 5,000+ |
| Active Clinical Users | 2-3 oncologists |
| Citations | 36 (medRxiv) |
Key Features
- Source Attribution: Every response linked to source documents
- Clinical Guardrails: Input/output guardrails prevent hallucination
- Query Understanding: Handles complex clinical terminology
- Hybrid Search: Combines semantic and keyword matching
Related Publications
This project resulted in the following publications:
- (Quidwai & Lagana, 2024) - Preprint with 36 citations
- (Quidwai et al., 2024) - Conference abstract at IMS 2024
Impact
First prototype of an AI-powered clinical decision support system deployed in active clinical use at a major cancer center, demonstrating the feasibility of production RAG systems in healthcare.
Technical Stack
Orchestration: LangGraph
Embeddings: BAAI/bge-large-en-v1.5
Generation: Mistral-7B
Vector Store: Amazon OpenSearch
Reranking: Cohere
Deployment: AWS (Bedrock, EC2)
Monitoring: MLflow
