h7lu/RiMCP_hybrid

Combines lexical (Lucene/BM25), semantic (vector embedding), and graph-based retrieval to navigate RimWorld's C# source code and XML definitions, enabling AI assistants to answer cross-layer queries like "which Defs use this C# component." Built as an MCP server with stdio/JSON-RPC transport for Claude Desktop and VS Code integration, it uses SIMD-accelerated vector operations and compressed sparse matrix storage (reducing 6.2GB to 400MB) to efficiently manage ~80,000 indexed code elements and relationship graphs. The hybrid search strategy uses BM25 for fast initial filtering across full source text, then semantic reranking, avoiding naive fusion while maintaining precision—subsequent iteration applies edge weighting (combining PageRank, occurrence frequency, and relation type) to surface the most relevant dependency paths.