💬 LiveKit RAG Assistant

AI-powered semantic search and question-answering system for LiveKit documentation

🎯 Features

Dual Search Modes: Documentation (Pinecone) + Real-time Web Search (Tavily)
Standard MCP Server: Async LangChain integration with Model Context Protocol
Fast Responses: Groq LLM (llama-3.3-70b) with ultra-fast inference
Semantic Search: HuggingFace embeddings (384-dim) with vector indexing
Source Attribution: View exact sources for every answer
Chat History: Persistent conversation tracking with recent query access
Query Validation: Prevents invalid inputs with helpful error messages
Copy-to-Clipboard: One-click message sharing

🚀 Quick Start

Prerequisites

Python 3.10+ with conda
API Keys: GROQ, TAVILY, PINECONE, HuggingFace

Installation

# 1. Clone and setup environment cd c:\lg mcp ai conda create -n langmcp python=3.12 conda activate langmcp # 2. Install dependencies pip install -r requirements.txt # 3. Configure .env file echo "GROQ_API_KEY=your_key" >> .env echo "TAVILY_API_KEY=your_key" >> .env echo "PINECONE_API_KEY=your_key" >> .env echo "PINECONE_INDEX_NAME=livekit-docs" >> .env

Running the Application

Terminal 1 - Start MCP Server:

python mcp_server_standard.py

Terminal 2 - Start Streamlit UI:

streamlit run app.py

The app opens at http://localhost:8501

📊 Architecture

Streamlit UI (app.py) ↓ Query Validation → Error Handling ↓ MCP Server (subprocess) → mcp_server_standard.py ↓ Dual Search Layer: ├─ Pinecone (3,007 vectors) - Semantic search └─ Tavily API - Real-time web results ↓ LLM Layer (Groq): ├─ Temperature: 0.3 (detailed, focused) ├─ Max Tokens: 2048 (comprehensive answers) └─ Model: llama-3.3-70b-versatile ↓ Response Display + Source Attribution

🔧 Tech Stack

Component	Technology	Purpose
Frontend	Streamlit 1.28+	Premium glassmorphism UI
Backend	MCP Standard	Async subprocess server
LLM	Groq API	Ultra-fast inference (free tier)
Embeddings	HuggingFace	sentence-transformers/all-MiniLM-L6-v2
Vector DB	Pinecone Serverless	Ultra-fast similarity search (AWS us-east-1)
Web Search	Tavily API	Real-time internet search
Framework	LangChain	LLM orchestration & tools
Language	Python 3.12	Modern syntax & features

📚 Project Structure

c:\lg mcp ai\ ├── app.py # Main Streamlit application (enhanced) ├── mcp_server_standard.py # MCP server with LangChain tools ├── ingest_docs_quick.py # Document ingestion to Pinecone ├── requirements.txt # Python dependencies ├── .env # API keys & configuration ├── README.md # This file └── ingest_docs.py # Legacy ingestion script

🎮 Usage

Ask Questions

Choose Search Mode: Documentation (📚) or Web Search (🌐)
Type Question: Natural language queries work best
Get Answer: AI responds with detailed 3-5 sentence answers
View Sources: Click "View Sources" to see cited documents

Features

Copy Messages: 📋 Click button on any message to copy
Recent Queries: ↻ Quick re-ask from history
Quick Help: 💡 Expandable tips and usage guide
Performance Metrics: 📊 Real-time statistics in System Status
Error Prevention: ✅ Query validation with helpful feedback

⚡ Performance

First Query: ~15-20s (model initialization)
Subsequent Queries: 3-8s (cached LLM)
Copy/History: Instant (client-side)
Metrics Update: Real-time (no overhead)

📖 Example Queries

"How do I set up LiveKit?"
"What are the best practices for video conferencing?"
"How do I implement real-time communication?"
"What authentication methods are available?"
"How do I handle bandwidth optimization?"
"Deploy to Kubernetes - how does LiveKit handle it?"

🛠️ Configuration

Edit .env file:

GROQ_API_KEY=gsk_your_groq_key TAVILY_API_KEY=tvly_your_tavily_key PINECONE_API_KEY=your_pinecone_key PINECONE_INDEX_NAME=livekit-docs HF_TOKEN=optional_huggingface_token

🔄 Ingestion

Populate Pinecone with LiveKit documentation:

python ingest_docs_quick.py

This creates 3,007 searchable vector chunks from LiveKit docs.

📊 Status Checks

View system status in Streamlit sidebar:

✅ MCP Server Ready - Status indicator ✅ Groq LLM - API connection ✅ Pinecone VectorDB - Index status 💬 Messages - Total message count 👤 Questions - User query count 🤖 Responses - AI response count

🚨 Troubleshooting

Issue	Solution
"No relevant documentation found"	Try web search mode or different keywords
"MCP Server not found"	Ensure `mcp_server_standard.py` is running in Terminal 1
Slow first response	Normal - model loads on first query (~15-20s)
API key errors	Check `.env` file and verify all keys are set
Empty Pinecone index	Run `python ingest_docs_quick.py` to populate

📝 Notes

All chat history saved in session state
Supports semantic search with keyword fallback
Responses stored with source attribution
Query validation prevents invalid inputs
Performance optimized for fast inference

👨‍💻 Created By

@THENABILMAN - GitHub

📄 License

Built with ❤️ for developers. Feel free to modify and extend!

Version: Enhanced v1.0 | Status: ✅ Production Ready | Date: November 1, 2025

This server cannot be installed

-

security - not tested

F

license - not found

-

quality - not tested

How are these scores calculated?

remote-capable server

The server can be hosted and run remotely because it primarily relies on remote services or has no dependency on the local environment.

Enables AI-powered semantic search and question-answering for LiveKit documentation using Pinecone vector search and real-time web search with Tavily, providing detailed responses with source attribution.

LiveKit RAG Assistant