Empower Mental Health Chatbot

A culturally sensitive AI-powered mental health chatbot using LangChain, GPT-4, Pinecone, and Flask. Developed as a master's capstone project at San José State University.

---

Project Overview

This chatbot addresses global mental health challenges by:

• Offering culturally personalized responses
• Supporting crisis detection and escalation
• Integrating retrieval-augmented generation (RAG) using Pinecone
• Enabling real-time voice/text interaction with TTS/STT
• Logging sessions for therapist reviews and continuity

Final Report Reference:
See Chapter 1: Introduction, Goals, and Contributions

---

Architecture Overview

Refer to Report: Chapter 2 – Project Architecture

Frontend (React) <--> Flask Backend (LangChain + GPT-4) <--> PostgreSQL + Pinecone
                    ↳ Tools: PineconeSearch | WebSearch | ChatSummary | NearestTherapist

• Routing Agent: Dynamically selects tools or raw LLM
• Tool-Enhanced Responses: Used for verified advice, cultural grounding, therapist lookup
• Raw GPT-4 Responses: Used when personal empathy is more helpful
• Dual Path Ranking: GPT-4 ranks both responses for final reply

---

Component Mapping to Report

Component	Description	Code Path	Report Chapter
Frontend	Chat UI with voice input, TTS playback, therapist map	frontend/components/	Ch. 4.1, 5.1
Chatbot Logic	Core LLM flow, RAG, prompt building, ranking	backend/chatbot_logic.py	Ch. 2.2.3, 5.2.1
Crisis Detection	BERT + GPT logic to detect emotional danger	chatbot_logic.py	Ch. 2.2.2, 4.2.3
Router Agent	Orchestrates tool usage with LangChain	backend/agents/router_agent.py	Ch. 2.2.4, 4.2.1
Pinecone Search Tool	Vector DB lookup using OpenAI embeddings	backend/tools/pinecone_search_tool.py	Ch. 4.2.2, 5.3.2
Web Search Tool	DuckDuckGo real-time fallback	tools/web_search_tool.py, web_search_beautiful.py	Ch. 3.3.3
Chat Summary Tool	Fetches previous summaries for continuity	tools/chat_summary_tool.py	Ch. 5.2.3
Therapist Locator	Google Maps API to suggest local help	tools/nearest_therapist_tool.py	Ch. 2.2.6, 5.3.3
Memory Buffer	Stores prior chat context in session	utils/memory_manager.py	Ch. 3.3.1
Session API	User auth, chat summary, therapist API	backend/api_for_db.py	Ch. 5.2.3
Data Embedding	Indexes JSON, PDF, scraped content	pinecone_rag.py	Ch. 5.3.5
PostgreSQL Setup	Stores users, chats, crisis logs, therapists	backend/connection.py	Ch. 4.3.1
Speech Modules	STT (browser-based), TTS (gTTS)	ChatArea.jsx, chatbot_logic.py	Ch. 5.1.2
Security	bcrypt password hashing, CORS, JWT-ready	api_for_db.py, .env	Ch. 5.3.4

---

Features

Feature	Description
Cultural Personalization	Tailors prompts based on selected user culture
RAG + Pinecone	Search vector DBs for CBT, PHQ-9, Indian context, SAMHSA
TTS & 🎙️ STT	Accessible speech input/output
Crisis Detection	Escalates emergencies with 988 lifeline + therapist map
Therapist Locator	Uses Google Maps API to suggest nearby help
Session Summaries	GPT-4 creates summaries for therapist review
Agent Tool Routing	LangChain agent selects tools dynamically
Secure Auth	Login/signup with hashed password & safe routing

---

Setup Instructions

1. Backend Setup (Flask)

cd backend
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt

Environment variables (.env):

OPENAI_API_KEY=your_key
PINECONE_API_KEY=your_key
VITE_GOOGLE_MAPS_API_KEY=your_key

Start the API:

python api_for_db.py

---

2. Frontend Setup (React + Vite)

cd frontend
npm install
npm run dev
npm install rehype-raw

Make sure the Flask backend is running on localhost:5001.

---

Testing

• ✅ 50 Test cases for Routing Agent – 88% accuracy
• ✅ 50 Crisis Detection inputs – 94% accuracy
• ✅ GPT-4 Evaluator scores (Empathy, Safety, Fluency)
• ✅ Manual testing: TTS, STT, Therapist Locator
• ✅ UI/UX tested across browsers and devices

---

Dataset Overview

Mapped to Report Section 5.3.5

Dataset Type	Description	Pinecone Namespace
🇮🇳 Indian Cultural Context	Family, beliefs, collectivism	indian_culture
SAMHSA Guidelines	Clinical treatments & meds	samhsa_guidelines
Forum Chat Logs	Empathetic support examples	chat_logs
CBT Techniques	15 strategies for mental health	cbt_techniques
PHQ-9	9-point depression scale	phq9_intents
Web-Scraped Info	WHO, NIMH, VeryWellMind	web_anxiety, etc.

---

Deployment Notes

• Currently tested on localhost
• Future-ready for deployment on Heroku, Render, or AWS
• Requires .env for OpenAI, Pinecone, and Google Maps API keys

---

Authors & Credits

Team Members:
Sai Pranavi Kurapati, Sayali Bayaskar, Divija Choudhary, Armaghan Abtahi

Advisor:
Professor Kaikai Liu

Final Report (PDF)
GitHub Repo: https://github.com/SayaliVB/MentalHealthChatbot