Corpus-KB

Local end-to-end RAG system for agentic code editors.

Exposes retrieval-augmented generation capabilities via MCP (Model Context Protocol) to any MCP-compatible client: OpenCode, Claude Code, Cursor, Codex, and others.

Features

• Hybrid Search — Vector similarity + full-text search fused via Reciprocal Rank Fusion

• Structure-Aware Chunking — AST-aware for code (tree-sitter, 40+ languages), heading-aware for markdown, semantic for plain text

• Hierarchical Context — Parent/child/sibling chunk relationships for context-aware retrieval

• Relational SQL — DuckDB queries over document metadata and chunk contents

• Entity Graph — SQLite-backed graph with entity/relation tracking, upgradable to GraphQLite (Cypher) or LatticeDB

• Data Versioning — LanceDB's built-in Git-like versioning with branches, tags, and time-travel

• 100% Local — Ollama for embeddings, no cloud dependencies, no Docker required

Related MCP server: agentmako

Quick Start

# Install dependencies
pip install -e .

# Pull embedding model
ollama pull nomic-embed-text

# Run the server
python -m src.server

MCP Configuration

See mcp-configs/ for client configuration files.