AI Chatbot for Mitek

Intelligent customer support chatbot using Retrieval-Augmented Generation (RAG) for Mitek J.S.C products and services.

Technology Stack

• Backend: FastAPI + Uvicorn
• Frontend: Streamlit
• Vector Database: FAISS (399 chunks from 80 pages)
• LLM: Google Gemini 2.5 Flash
• Embeddings: multilingual-e5-base (768 dimensions)

Performance

• Average response time: 12 seconds
• Retrieval accuracy: 85%
• Index size: 500KB
• Memory usage: 500MB

---

Quick Start

1. Installation

git clone https://github.com/mapleleaflatte03/ai-chatbot.git
cd ai-chatbot
python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -r requirements.txt

2. Configuration

Create .env file in project root:

GOOGLE_API_KEY=your_google_gemini_api_key
DEMO_SEEDS=https://mitek.vn/

3. Run Application

.\run.ps1     # Start both API and UI servers
.\stop.ps1    # Stop all services

Access Points

• Streamlit UI: http://127.0.0.1:8501
• FastAPI Docs: http://127.0.0.1:8000/docs
• API Health: http://127.0.0.1:8000/health

---

API Reference

Endpoints

POST /ask

Request:  {"question": "MiPBX là gì?"}
Response: {
  "answer": "MiPBX là giải pháp tổng đài ảo...",
  "sources": [{"url": "https://mitek.vn/mipbx", "title": "..."}]
}

GET /health

Response: {"ok": true}

GET /metrics

Response: {
  "p95_latency": 12.5,
  "counts": {"cntt": 10, "software": 5}
}

---

Project Structure

ai-chatbot/
├── app/
│   └── main.py              # FastAPI backend with RAG pipeline
├── ui/
│   └── app.py               # Streamlit chat interface
├── scripts/
│   ├── 01_crawl.py          # Web scraper for mitek.vn
│   └── 02_build_index.py    # FAISS vector index builder
├── data/
│   └── faq.csv              # Crawled data (80 pages)
├── storage/
│   ├── index.faiss          # FAISS vector index (399 chunks)
│   ├── meta.json            # Chunk metadata
│   └── log.jsonl            # Query logs
├── .env                     # Environment variables (not in git)
├── requirements.txt         # Python dependencies
├── run.ps1                  # Startup script
└── stop.ps1                 # Shutdown script

---

How RAG Works

The system implements a standard Retrieval-Augmented Generation pipeline:

1. Query Embedding: Convert user question to 768-dimensional vector using multilingual-e5-base
2. Vector Search: FAISS searches 399 indexed chunks and returns top-4 most relevant results
3. Context Assembly: Combine retrieved chunks (max 800 characters)
4. LLM Generation: Send context + question to Google Gemini 2.5 Flash
5. Response Delivery: Return generated answer with source URLs

Special Feature: Greeting detection bypasses RAG for instant responses to simple greetings.

---

Data Pipeline (Optional)

To re-crawl and rebuild the knowledge base:

# Stop services
.\stop.ps1

# Crawl mitek.vn
.venv\Scripts\python.exe scripts\01_crawl.py --seeds "https://mitek.vn/" --limit 80 --out data\faq.csv

# Build FAISS index
.venv\Scripts\python.exe scripts\02_build_index.py --csv data\faq.csv --out storage\index.faiss --meta storage\meta.json

# Restart services (required to load new index)
.\run.ps1

---

License

MIT License - Open source for educational and commercial use.