AgentRAG — Multi-Source Agentic RAG with MCP

English | 中文

An intelligent research assistant that combines Retrieval-Augmented Generation with an autonomous agent powered by Model Context Protocol (MCP).

The agent reasons about your question, selects the right retrieval tools, fetches information from multiple sources, and synthesizes a coherent answer — all through a CLI, REST API, or Web UI.

Architecture

User Query
    │
    ▼
┌─────────────────────────────────┐
│  LangGraph ReAct Agent          │
│  (DeepSeek / GPT / Claude)      │
│                                 │
│  Thinks → Selects Tools → Acts  │
└────────────┬────────────────────┘
             │ MCP Tool Calls
             ▼
┌────────────────────────────────────────────┐
│  MCP Servers (FastMCP)                     │
│                                            │
│  📄 docs-rag     — Index & search docs     │
│  🔍 web-search   — DuckDuckGo + fetch      │
│  💻 code-index   — Semantic code search     │
│  🧠 memory       — Persistent memory       │
└────────────┬───────────────────────────────┘
             │
             ▼
┌────────────────────────────────────────────┐
│  Vector Store (Qdrant local)               │
│  Embeddings: text-embedding-3-large        │
└────────────────────────────────────────────┘

Features

• Multi-source retrieval: Documents, code, web, and persistent memory
• Autonomous agent: LangGraph ReAct pattern — the agent decides which tools to use
• MCP integration: Tools exposed as MCP servers (FastMCP), composable and extensible
• Multiple interfaces: CLI (click), REST API (FastAPI), Web UI (Streamlit)
• Flexible LLM backend: Works with any OpenAI-compatible API (DeepSeek, GPT, Qwen, etc.)
• Local vector store: Qdrant in local mode — no external services needed

Quick Start

# Clone and install
git clone https://github.com/URaux/agentrag.git
cd agentrag
pip install -e ".[dev]"

# Configure API keys
cp .env.example .env
# Edit .env with your API keys

# Index a document
agentrag index ./pyproject.toml --type doc

# Ask a question (agent auto-retrieves)
agentrag ask "What dependencies does this project use?"

# Search directly (no agent reasoning)
agentrag search "dependencies" --source docs

# Start the API server
agentrag serve

CLI Commands

Command	Description
agentrag ask <query>	Ask the agent — it retrieves and reasons
agentrag index <path>	Index a file or directory
agentrag search <query>	Direct semantic search (no agent)
agentrag serve	Start REST API server
agentrag status	Show system status

REST API

# Start server
agentrag serve --port 8000

# Ask the agent
curl -X POST http://localhost:8000/ask \
  -H "Content-Type: application/json" \
  -d '{"query": "What is this project about?"}'

# Index a document
curl -X POST http://localhost:8000/index \
  -H "Content-Type: application/json" \
  -d '{"path": "./README.md", "doc_type": "doc"}'

# Direct search
curl -X POST http://localhost:8000/search \
  -H "Content-Type: application/json" \
  -d '{"query": "dependencies", "source": "docs"}'

MCP Servers

Each retrieval source runs as an independent MCP server:

Server	Tools	Description
docs-rag	index_document, search_documents, list_indexed	Index and search local documents
web-search	web_search, fetch_page	Real-time web search via DuckDuckGo
code-index	index_repo, search_code	Semantic code repository search
memory	save_memory, recall, list_memories	Persistent conversation memory

Tech Stack

• Agent: LangGraph (ReAct pattern)
• LLM: Any OpenAI-compatible API
• MCP: FastMCP 2.x
• Vector Store: Qdrant (local mode)
• Embeddings: text-embedding-3-large (3072 dim)
• CLI: Click + Rich
• API: FastAPI + Uvicorn
• Web UI: Streamlit

Configuration

All settings via environment variables (prefix AGENTRAG_):

Variable	Default	Description
AGENTRAG_API_KEY	—	LLM API key
AGENTRAG_API_BASE	https://api.deepseek.com/v1	LLM API base URL
AGENTRAG_MODEL	deepseek-chat	Chat model name
AGENTRAG_EMBEDDING_API_KEY	—	Embedding API key (falls back to API_KEY)
AGENTRAG_EMBEDDING_API_BASE	—	Embedding API base (falls back to API_BASE)
AGENTRAG_EMBEDDING_MODEL	text-embedding-3-large	Embedding model

License

MIT