GRAPHRAG & COGNITION PRODUCTION READY LOCAL AIR-GAPPED
Hypertensor GraphRAG Operator Manual
Deploy, configure, and operate the in-memory Hypertensor GraphRAG engine, SIMD vector index, and FastMCP server.
Hypertensor GraphRAG Operator Manual
Hypertensor GraphRAG is an in-memory typed property-graph store designed for sub-millisecond retrieval, hybrid dense and BM25 search, and multi-hop graph traversal. It bridges directly to AI agents via a 65+ tool FastMCP server.
Architectural Highlights
- In-Memory Property Graph: Zero disk I/O during query execution; typed nodes and edges with attribute indices.
- Hybrid Dense & BM25 Search: Reciprocal Rank Fusion (RRF) combining dense vector embeddings with sparse BM25 inverted indices.
- SIMD TensorPrimitives: Vector distance computations (cosine, dot, L2) accelerated on CPU vector registers or Level Zero GPU.
- FastMCP Protocol: Standardized Model Context Protocol (MCP) server exposing tools, prompts, and resources to desktop HUDs and agent runners.
Quickstart Commands
1. Launch In-Memory GraphRAG Server
dotnet run --project src/HFLabs/Products/HypertensorGraphRAG/Server/Server.csproj -c Release2. Run Headless Query
dotnet run --project src/HFLabs/Products/HypertensorGraphRAG/App/App.csproj -c Release -- --headless --query "Director"3. Launch FastMCP SSE Bridge
python -m mcp_graphrag.server --transport sse --port 8000Configuration Options
Configure appsettings.json or environment variables:
| Setting | Default | Description |
|---|---|---|
GraphRAG:VectorDimensions | 1536 | Embedding vector dimensionality |
GraphRAG:SIMDAcceleration | true | Enable hardware SIMD TensorPrimitives |
GraphRAG:BM25:k1 | 1.2 | BM25 term frequency saturation parameter |
GraphRAG:BM25:b | 0.75 | BM25 document length normalization parameter |
GraphRAG:MaxHopDepth | 4 | Maximum multi-hop graph exploration depth |
FastMCP:Port | 8000 | Port for FastMCP SSE / HTTP endpoint |
FastMCP Tools Catalog
The built-in FastMCP server exposes over 65 tools for agent reasoning, including:
graph_query_nodes: Retrieve nodes by type, property filters, or semantic proximity.graph_find_paths: Find shortest or lowest-cost causal paths between entity nodes.graph_hybrid_search: Execute combined BM25 and vector nearest-neighbor search.graph_expand_subgraph: Extract k-hop neighborhood ego-graphs for context injection.graph_tri_beam_verify: Verify entity candidate strings against graph token trie.