Combines vector similarity, tree-based reasoning, and keyword search in a single retrieval pipeline. Exposes hybrid search operations where HNSW handles broad queries, LLM-guided tree navigation handles structured documents (think 10-Ks or technical specs), and BM25 catches exact matches. Results get fused with reciprocal rank scoring. Also includes agent memory primitives (episodic, semantic, procedural, shared) and multi-agent orchestration. The MCP server surfaces these as tools Claude can call. Runs locally with Ollama or mock embeddings. Reach for this when vector search alone keeps returning similar-but-wrong results and you need reasoning about document structure. Particularly relevant for financial docs, compliance materials, or any corpus where precision beats semantic fuzziness.
claude mcp add --transport stdio fusionpacttech-fusionpact-vectordb -- npx -y fusionpact