Knowledge MCP

A high-precision local knowledge base server (RAG) that implements the Model Context Protocol (MCP). It enables AI agents (Codex, Claude Code, Gemini CLI, etc.) to navigate, search, and reason about complex codebases using a hybrid approach combining semantic, lexical, and structural analysis.

🚀 Key Features

1. Hybrid Triple Search: Fuses three distinct retrieval channels through Reciprocal Rank Fusion (RRF) for maximum recall:

   • Full-Text Search (FTS5): Handles exact name matches and specific keywords.

   • Semantic Vector Search (sqlite-vec): Understands concepts and natural language intent.

   • Graph-Based Retrieval: Provides a 2x relevance boost to actual code symbols and their relationships.

2. Deep Semantic Indexing:

   • C# / .NET: Integrated Roslyn analysis for precise symbol extraction and dependency graphs.

   • Polyglot Support: Tree-sitter integration for high-quality parsing of TS, JS, Python, Go, and more.

   • Markdown: Section-aware chunking for documentation.

3. Knowledge Graph: Tracks relationships between symbols: CALLS, INHERITS, IMPLEMENTS, and IMPORTS. Supports recursive Impact Analysis to estimate the blast radius of code changes.

4. Autonomous Embeddings: In-process generation using sentence-transformers (mpnet-base-v2). Works natively inside Docker without external API dependencies.

5. Incremental Sync: Delta-sync mechanism via mtime and SHA-256 ensures only modified files are processed, significantly speeding up updates.

🛠 Available MCP Tools

Search & Retrieval

• knowledge_search: Triple hybrid search across all repositories. Returns chunks with trust levels (verified for code, hint for docs).

• knowledge_get_chunk: Retrieve detailed content and metadata for a specific knowledge chunk.

Symbol & Graph Navigation

• knowledge_find_symbol: Locate classes, methods, and interfaces using wildcards (e.g., *Repository).

• knowledge_get_callers / knowledge_get_callees: Navigate the call graph of any symbol.

• knowledge_get_hierarchy: Explore inheritance and interface implementations.

• knowledge_impact_analysis: Perform recursive dependency analysis to find everything affected by a symbol change.

Management

• knowledge_sync_repo: Trigger a background delta-sync for a repository to update the AI's "memory" after code changes.

• knowledge_delete_repo: Wipe all indexed data for a specific repository.

📋 Requirements

• Python 3.10+

• .NET 8.0 SDK (Required for Roslyn-based C# analysis)

• Docker (Recommended for easiest deployment)

🐳 Quick Start (Docker)

1. Configure your repository path

Create a .env file in the project root:

# The host directory that will be mounted as /repos inside the container.
# Set this to the PARENT folder of your repositories.
REPOS_DIR=C:\Repos

2. Build and start the container

docker compose up -d

3. Index a specific project

Option A — PowerShell script (recommended, see Scripts):

.\scripts\Reindex-Repo.ps1 -Wait

Option B — direct HTTP call:

curl -X POST http://localhost:8000/sync \
     -H "Content-Type: application/json" \
     -d '{"repo_id": "my-app", "repo_path": "/repos/my-app"}'

⚠️ repo_path is the path inside the container (e.g., /repos/my-app), not the host path.

4. Connect your MCP Client

Add this to your AI client's config (e.g., mcp.json):

{
  "mcpServers": {
    "knowledge-mcp": {
      "command": "docker",
      "args": ["exec", "-i", "knowledge-mcp", "python", "-m", "knowledge_mcp.main", "mcp"]
    }
  }
}

🔧 Scripts

scripts/Reindex-Repo.ps1

A PowerShell helper script that triggers re-indexing of a repository by sending a POST request to the running knowledge-mcp server.

Parameters:

Usage examples:

# Trigger re-indexing with default settings (ImpactOS.Core.Lib)
.\scripts\Reindex-Repo.ps1

# Trigger and watch progress in real time
.\scripts\Reindex-Repo.ps1 -Wait

# Index a different repository
.\scripts\Reindex-Repo.ps1 -RepoId "MyOtherLib" -RepoPath "/repos/MyOtherLib"

# Point to a remote server
.\scripts\Reindex-Repo.ps1 -ServerUrl "http://192.168.1.100:8000" -Wait

How it works:

1. Verifies the server is reachable at ServerUrl.

2. Sends a POST /sync request with repo_id and repo_path.

3. Indexing runs in the background inside the container.

4. With -Wait, streams live logs via docker logs -f.

💡 Path mapping: If REPOS_DIR=C:\Repos, then C:\Repos\MyLib on the host is accessible inside the container as /repos/MyLib.

📐 Architecture & Decisions

• Roadmap: Autonomous Sync & Webhooks

For deep dives into the technical design, see our Architecture Decision Records:

• ADR-001: Local MCP RAG Foundation

• ADR-002: Hybrid Semantic & Graph Indexing