DeepRetrieval

Project Overview

DeepRetrieval is a minimal, production-oriented implementation of a Retrieval-Augmented Generation (RAG) agent exposed through a Model Context Protocol (MCP) server, built without using any external systems.

✨ Key Features

• 🧠 Custom RAG Pipeline

  • Document chunking
  • TF-IDF / classical vector search
  • Context injection into prompts

• 🔌 MCP Server (JSON-RPC)

  • Fully compliant MCP tool server
  • Stdio-based transport

• 🛠 Agentic Tool Calling

  • Explicit tool routing
  • Deterministic execution
  • No hidden orchestration

• 🔍 Local & Web Search Tools

  • Local document search
  • Optional web search integration

• ⚙️ Zero Framework Dependency

• 🧱 Use of ollama that is local llm use for seamless streaming of queries

User Query
    ↓
Agent Controller
    ↓
Tool Router
    ├── Document Search Tool (TF-IDF)
    ├── Web Search Tool
    └── Utility Tools
    ↓
Context Composer
    ↓
LLM (Answer strictly from provided context)

🏆 Why This Project Matters

Most production AI systems do not rely on public agent frameworks. This project demonstrates how to build:

• Reliable agents

• Transparent reasoning

• Auditable tool execution

• Vendor-neutral architectures

—all from scratch.

Example

1. Built from scratch - No LangChain dependency shows deep understanding
2. Proper architecture - Clean separation: chunking, embedding, retrieval, inference
3. Working semantic search - Similarity scores (0.83, 0.74, 0.71) show it's finding relevant content
4. Database integration - Vector storage with proper indexing
5. Token-aware chunking - Smart overlap strategy preserves context

👤 Author

Aditya Katkar