malishilpa9119/health-care-chatbot

Purpose and Runtime Entry Points

This repository is a document-grounded healthcare clinic chatbot. It answers clinic-policy and medical-document questions through a RAG workflow, while routing emergency, medication/dosage, and unsupported requests to safe responses.

• Web runtime: app.py

• Indexing/batch ingestion runtime: store_index.py

• Core application package: src/

Document Ingestion and Index Construction

Clinic and medical PDFs are the authoritative knowledge source. The indexing process loads PDFs, chunks their content, produces HuggingFace embeddings, and writes vectors to Pinecone. Re-indexing is required after document changes.

• Source corpus: data/HealthFirst_Clinic_Info.pdf, data/Medical_book.pdf

• Loading, splitting, and embedding utilities: src/helper.py

• Index build entry point: store_index.py

Retrieval Layer

The retrieval boundary converts a user question into grounded evidence from the vector store. It combines semantic Pinecone search with BM25 keyword retrieval, then applies cross-encoder re-ranking to prioritize the most relevant document chunks.

• Hybrid retrieval and re-ranking: src/retriever.py

• Retriever/tool implementation is consumed by the workflow: src/agent.py

• Persistent external retrieval service: Pinecone

LangGraph Conversation Workflow

The central control flow is a LangGraph agent pipeline:

1. Receive current user question plus multi-turn conversation context.

2. Apply safety guardrails.

3. Retrieve and rank document evidence when permitted.

4. Generate an answer constrained to retrieved content.

5. Use a fallback branch when evidence is absent or the request is unsafe/out of scope.

• Workflow nodes, routing, and agent state: src/agent.py

• Grounding and behavioral instructions: src/prompt.py

Generation, Safety, and Fallback Boundaries

Groq-hosted Llama 3.3 70B performs answer generation. Prompting and graph routing enforce the core product constraint: do not invent clinic or medical facts beyond supplied documents. Emergency and medication/dosage requests are safety boundaries rather than normal retrieval questions; unsupported requests receive a defined fallback.

• Agent orchestration and guardrail/fallback handling: src/agent.py

• System prompt contract: src/prompt.py

• External LLM provider: Groq; optional image-analysis provider: OpenAI

Flask Interface and Session Interaction

Flask exposes the chat application on the local web server and connects browser requests to the agent workflow. The HTML template provides the conversation interface, with static CSS for presentation. The application maintains multi-turn context so follow-up questions can depend on prior turns.

• HTTP application entry point: app.py

• Chat UI: templates/chat.html

• Client styling: static/style.css

Deployment and Dependencies

The runtime is Python, with container deployment support. Required credentials are environment-provided API keys for Pinecone and Groq; OpenAI is optional for image analysis. The architecture depends on LangChain/LangGraph for orchestration, Pinecone for vector persistence, HuggingFace embeddings for vectorization, and Flask for serving.

• Dependency declaration: requirements.txt

• Container runtime definition: Dockerfile

• Python runtime version declaration: runtime.txt