🏥 Emergency Medical Triage – Custom RAG Model

This project implements a Retrieval-Augmented Generation (RAG) system for emergency medical triage using Haystack and OpenAI’s GPT models. The model retrieves relevant clinical content from a curated document store and generates structured triage decisions based on user queries.

⚕️ Use Case

Emergency triage is a high-stakes process where rapid, accurate decisions are critical. This project explores how RAG-enabled large language models (LLMs) can support clinical decision-making by:

• Assigning triage levels based on symptoms
• Providing rationale for decisions
• Referencing reliable medical sources (e.g., WHO triage manuals, medical textbooks)

🧠 Tech Stack

• Framework: Haystack (Python)
• LLM: OpenAI GPT (via API)
• Retriever: BM25 (InMemory), with support for FAISS
• Evaluation: Ragas framework for LLM performance assessment
• Pipeline components: Retriever, DocumentJoiner, PromptBuilder, OpenAI Generator, AnswerBuilder
• Optional: RAGAS Evaluator

⚙️ How It Works

1. Query Input: User enters a triage-related question (e.g. “A 70-year-old male with chest pain and shortness of breath. What should be done?”)
2. Document Retrieval: Top-k relevant chunks are retrieved from the document store (BM25 or FAISS).
3. Prompt Construction: Retrieved content is injected into a prompt template.
4. Generation: OpenAI GPT generates a triage recommendation and rationale.
5. Evaluation: RAGAS evaluates outputs on metrics such as: Answer Relevancy, Faithfulness, Context Precision, Custom metric: Patient Triaged (binary)

📊 Sample Output

{ "triage_decision": "Red (Critical)", "rationale": "Patient is elderly, experiencing chest pain and shortness of breath — possible cardiac event. Requires immediate intervention." }

✅ Key Learnings

• Prompt engineering and document chunking are critical to RAG performance.
• Evaluation without ground-truth labels is possible using LLM-based metrics.
• Model interpretability is vital in clinical settings — black-box models are not ideal.