RAG-Powered First-Aid Chatbot for Diabetes, Cardiac & Renal Emergencies

Mission Statement

Build a patient-safety–aware chatbot that combines local knowledge embeddings with real-time web evidence to deliver actionable first-aid guidance, always prefaced with a clinical disclaimer.

Disclaimer: "This information is for educational purposes only and is not a substitute for professional medical advice."

---

Resume-Ready Project Summary (Bullet Points):

• Developed a patient-safety–focused RAG (Retrieval-Augmented Generation) chatbot for first-aid guidance in diabetes, cardiac, and renal emergencies.
• Achieved 100% compliance with clinical disclaimer and answer structure requirements, as verified by automated tests on 10+ sample queries.
• Architected a hybrid retrieval system combining FAISS vector search and real-time web search (Serper.dev) to maximize answer relevance.
• Integrated a HuggingFace LLM (Mistral/Mixtral) for concise, actionable response generation.
• Deployed a secure, user-friendly Streamlit web interface with strict domain and safety controls.

---

Core Objectives

1. Triage / Diagnosis: Infer the most likely condition from free-text symptoms.
2. Hybrid Retrieval:
   • Local semantic search over the 60 provided medical snippets.
   • Serper.dev web search for fresh evidence.
   • Fuse results (semantic + Serper.dev results) and rank by relevance (keyword search).
3. Answer Generation: Produce ≤ 250-word response with: condition, first-aid steps, key medicine(s), and source citations. All answers are always prefixed with a clinical disclaimer.

Focus Areas

The chatbot is designed to answer any question about the following emergencies:

• Diabetes: Type 1, Type 2, Gestational, ketoacidosis, hypoglycaemia.
• Cardiac: Myocardial infarction, angina, arrhythmia, heart failure.
• Renal: AKI, CKD, hyperkalaemia, dialysis crises.

Project Structure

• src/ — Main source code (including medibot.py for Streamlit UI, retrieval, and LLM logic)
• tests/ — Automated tests (pytest, includes 10 sample queries)
• data/ — Local data (e.g., Assignment Data Base.xlsx)
• vectorstore/ — FAISS vector database
• requirements.txt — Python dependencies
• README.md — This file
• architecture.md — System architecture and design

Setup Instructions

1. Clone the repository

   git clone <repo-url>
   cd <repo-directory>

2. Install dependencies

   pip install -r requirements.txt

3. Set up environment variables

   • Create a .env file in the root directory with:

     HF_TOKEN=your_huggingface_token
     SERPER_API_KEY=your_serper_dev_key
     

4. Prepare the vectorstore

   • Place your 60 medical snippets in /data/Assignment Data Base.xlsx (already present).
   • Run:

     python src/create_memory_for_llm.py

5. Run the chatbot

   streamlit run medibot.py
   # or, alternatively
   python -m streamlit run src/medibot.py

6. Run automated tests

   pytest tests/test_sample_queries.py

Usage Instructions

• Open the Streamlit app in your browser (usually http://localhost:8501).
• Enter a query related to diabetes, cardiac, or renal emergencies (see sample queries below).
• The chatbot will:
  • Preface every answer with a clinical disclaimer (enforced in all code paths)
  • Infer the most likely condition
  • Provide first-aid steps as a numbered list
  • List key medicines as bullet points (if relevant)
  • Cite sources as a bulleted list of proper references/links (no Local[x] or Web[x] artifacts)
  • Output answers in a structured format:
    • Inferred Condition:
    • First-Aid Steps: (numbered)
    • Key Medicine(s): (bulleted)
    • Sources: (bulleted, with links)
  • Keep answers concise and ≤ 250 words

Sample Queries:

• "I'm sweating, shaky, and my glucometer reads 55 mg/dL—what should I do right now?"
• "Crushing chest pain shooting down my left arm—do I chew aspirin first or call an ambulance?"
• (See /tests/test_sample_queries.py for all 10 test queries)

Design Trade-offs

• Retrieval: Combines local semantic search (FAISS + MiniLM) and Serper.dev web search for up-to-date evidence.
• Fusion & Ranking: Uses keyword overlap for simple, transparent ranking. More advanced semantic ranking could be added.
• LLM: Uses Mistral-7B-Instruct-v0.3 via HuggingFace Endpoint for cost and performance balance.
• Safety: Always includes a clinical disclaimer to reinforce that the information is not a substitute for professional medical advice.
• Testing: Automated pytest script for the 10 sample queries ensures core requirements are met.

Known Limitations

• No advanced semantic reranking (just keyword overlap).
• No user feedback loop or learning from corrections.
• Dependent on external APIs (HuggingFace, Serper.dev) for inference and web search.
• Performance and accuracy may vary with API/model changes.