MemKB - Memory Knowledge Base

A Rust-based MCP (Model Context Protocol) server that provides AI-powered semantic search over markdown documentation using local LLMs all in memory!

Features

• Semantic Search: Uses embeddings for intelligent content retrieval
• AI-Generated Answers: Synthesizes information from multiple sources into coherent responses

Quick Start

Basic Usage

# With embedding and generation servers
cargo install memkb
memkb --directory ./docs -e http://127.0.0.1:9095/v1 -g http://127.0.0.1:9091/v1

Command Line Options

-p, --port <PORT>                      MCP server port [default: 8080]
-H, --host <HOST>                      Server host [default: localhost]
-d, --directory <DIRECTORY>            Directory with .md files [default: .]
-e, --embedding-url <URL>              Embedding server endpoint
-g, --generation-url <URL>             Generation server endpoint  
-c, --chunk-size <SIZE>                Chunk size in characters [default: 1000]
-o, --overlap <OVERLAP>                Chunk overlap in characters [default: 200]
    --test                             Enable web test interface on port+1

How It Works

1. Indexing: Scans directory for .md files and chunks them
2. Embedding: Generates embeddings for each chunk (if embedding server available)
3. Query: When asked a question:
   • Generates query embedding
   • Finds most similar chunks
   • Uses generation server to create contextual answer
4. Response: Returns AI-generated answer or falls back to raw chunks

Web Test Interface

When using --test, a web interface is available at http://localhost:8081 (port+1) for easy testing.

Key Libraries

• rmcp: Rust MCP (Model Context Protocol) framework for building AI tool servers
• simsimd: SIMD-accelerated similarity calculations for fast embedding comparisons
• text-splitter: Intelligent markdown-aware text chunking with configurable overlap

MCP Integration

The server exposes an ask tool that can be used by MCP-compatible clients like Claude Desktop.