iaakashkr/Medical_BOT

Simple Medical chatbot using RAG + LLM. Ask questions and get structured responses.

Medical Chatbot RAG

A Retrieval-Augmented Generation (RAG) chatbot designed for answering medical FAQs. It leverages semantic embeddings (FAISS) and syntactic search (BM25) to fetch relevant context, and Google Gemini LLM to generate accurate, context-aware responses.

Table of Contents

1. 1.Overview
2. 2.Features
3. 3.Architecture
4. 4.Folder Structure
5. 5.Setup
6. 6.Usage
7. 7.Generating Resources
8. 8.Environment Variables
9. 9.Pipeline Details
10. 10.FAQ
11. 11.Contributing
12. 12.License

Overview

This chatbot combines a hybrid retrieval system with LLM-based generation:

• —Semantic Search: FAISS embeddings for vector similarity.
• —Syntactic Search: BM25 keyword matching for exact phrase relevance.
• —RAG Integration: Combines retrieved examples with user questions for LLM input.
• —Conversational Memory: Keeps chat history for context-aware responses.
• —Structured Output: Returns JSON containing answer, source_examples, and usage stats.

Ideal for hospitals, clinics, medical students, or educational platforms.

Try It Online

You can try the chatbot live without installing anything:

Click here to test the Medical BOT

Sample: <img width="1918" height="923" alt="image" src="https://github.com/user-attachments/assets/7e6de580-483a-4c1c-9f9d-ee55f58671b6" /> :

**RESPONSE**

<img width="1919" height="924" alt="image" src="https://github.com/user-attachments/assets/0275ea2c-705f-4d91-9809-2c83f360f706" />

Features

• —Retrieval-Augmented Generation (RAG)
• —Few-shot example selection using FAISS + BM25
• —Context-aware responses from Google Gemini LLM
• —JSON output with answer and source_examples
• —Chat history management
• —Modular, reproducible pipeline
• —Token usage tracking

Architecture

User Input
   │
   ▼
[Few-Shot Retrieval] ←─ FAISS Embeddings + BM25 ──→ Candidate Examples
   │
   ▼
[Context Builder] → Merge examples + chat history
   │
   ▼
[LLM Generation] → Google Gemini LLM
   │
   ▼
[JSON Output] → Answer + Source Examples + Token Usage
   │
   ▼
User receives response via CLI or API

Folder Structure

medical-chatbot-rag/
│
├── app/                     
│   ├── MED_CHATBOT.py
│   └── dto/                 # DTO folder
│       └── dto.py
│
├── pipeline/                # Core logic for retrieval & LLM calls
│   ├── embedder.py
│   ├── llm.py
│   ├── retrieval.py
│   ├── token_counter.py
│   └── token_tracker.py
│
├── resources/               # Datasets & precomputed embeddings
│   ├── train.csv
│   ├── embeddings/
│   │   └── med_embeddings.faiss
│   └── pickles/
│       └── syntactic_model_med.pkl
│
├── logs/                    # Log files
│
├── .env                     # API keys (ignored in git)
├── requirements.txt         # Dependencies
└── README.md

Setup

1. 1.Clone Repo

git clone https://github.com/iaakashkr/medical-chatbot-rag.git
cd medical-chatbot-rag

1. 1.Install Dependencies

pip install -r requirements.txt

1. 1.Create `.env`

GEMINI_API_KEY=your_google_gemini_api_key

1. 1.Generate FAISS Embeddings & BM25 Pickle

python pipeline/embedder.py

(if resources already exist, skip this step)

Usage

CLI

python app/MED_CHATBOT.py

Type medical questions and get structured answers with source examples.

Generating Resources

If you want to update your embeddings or BM25 models:

1. 1.Load dataset: resources/train.csv
2. 2.Run embedding script:

python scripts/generate_embeddings.py

1. 1.FAISS index saved to resources/embeddings/med_embeddings.faiss
2. 2.BM25 pickle saved to resources/pickles/syntactic_model_med.pkl

Environment Variables

• —GEMINI_API_KEY: Your Google Gemini API key (required)
• —Optionally, adjust model name or thresholds in pipeline/retrieval.py

Pipeline Details

• —Few-Shot Retrieval: Combines FAISS semantic similarity and BM25 syntactic scores to select top-K examples.
• —Exact Match Bonus: Slightly increases score if user query exactly matches example questions.
• —LLM Call: Uses llm_medical.py to send context and question to Gemini, returns structured JSON.
• —Chat History: Maintains last N turns for continuity.

FAQ

Q: Do I need API keys to run locally? A: Yes, .env must contain your Gemini API key.

Q: Can I use precomputed embeddings? A: Yes, keep med_embeddings.faiss and BM25 pickle in resources/.

Q: How many examples does the bot fetch? A: Top 2 by default, configurable in fetch_few_shots().

Contributing

• —Fork repo
• —Create branch: git checkout -b feature/your-feature
• —Commit changes
• —Push: git push origin feature/your-feature
• —Open PR

License

This project is licensed under the Apache 2.0 License © 2025 Akash Kumar