Hypertensor GraphRAG & SIMD Vectors
Coming soon Hypertensor GraphRAG is not yet released. Details on this page may change before launch.
An in-memory typed property-graph and vector store for autonomous developer agents. It keeps the graph in RAM and serves agents over REST and MCP. That removes disk-bound lookups and external database round-trips.
Engine Architecture & Substrates
Built for low-latency agent workloads where a disk-bound or networked vector database adds too much per-query overhead.
Low-latency in-memory traversal
The graph store is RAM-resident. Measured 2-hop traversal is about 0.15 ms at ~1,000 nodes and 0.16 ms at ~10,000 nodes. It rises to about 1.95 ms at 25,000 nodes / 100,000 edges. A thread-safe worker pool serves concurrent queries.
SIMD vector search
Dense cosine similarity runs on .NET TensorPrimitives with hardware SIMD. Vectors are 128-dimensional by default and configurable per store. It is combined with BM25 for hybrid retrieval. Level Zero GPU offload is available in the High-Assurance tier.
FastMCP protocol server
A Python FastMCP layer bridges MCP clients to the .NET server. It supports stdio, SSE and Streamable HTTP. Agents can query codebase topology, dependency graphs and symbol call-graphs.
Offline signed licensing and MLS
Licenses are verified offline against an issuer public key (RSA), so air-gapped deployments work. High-Assurance adds Bell-LaPadula multi-level security (MLS) clearance enforcement and FIPS-approved HMAC-SHA256 clearance tokens.
Graph-grounded verification
The Grounding Verifier checks claims against graph memory, validates citations and flags hallucinated edges. TAPE (Task-Adaptive Prediction & Explanation) plans exploration and reasoning as action-perception DAGs.
Sharded, hardware-aware memory
The BESS partitioner shards graphs. A hardware-topology model covers NUMA nodes, fabric links and pinned host memory. The largest benchmarked configuration is 1,000,000 nodes / 5,000,000 edges.
51 FEATURE SUB-GRAPHS
Hypertensor GraphRAG โ 10 FREE | 14 PRO | 27 HIGH ASSURANCE
PROOF CONTRACT
โx โ G: Inv(x) โ โกValid limtโโ V(x) = 0
||x - x*|| < ฮต Tier assignment reflects the launch baseline (sales gate 0); features cascade to lower tiers at commercial milestones. Internal bundles are omitted.
Coming soon Hypertensor GraphRAG is not yet released. Tiers and terms may change before launch.
Licensing Tiers
Choose the tier that fits your workload: free single-developer use, per-seat Pro, or team-level High-Assurance.
Free
Free single-developer workstation use
Runtime Bounds & Capacity
- Validated Bound 25,000 nodes / 100,000 edges
- SIMD Vectors 128-d SIMD vectors, CPU only
- 2-Hop Traversal 2-hop traversal โ 1.95 ms
- Worker Pool 16 workers
Included Subsystems
- Codebase Topology
- Continuous Tasks DAG queue
- GitHub Provenance
- Agent Memory (local single-developer store)
- Developer Workspace scanning
- Graph Consistency linting
- Directive/Playbook symlink sync
Pro
Commercial developers & scaling engineering teams
Runtime Bounds & Capacity
- Validated Bound 100,000 nodes / 500,000 edges
- 2-Hop Traversal 2-hop traversal โ 1.47 ms
- Worker Pool 64 workers (parallel acceleration enabled)
Included Subsystems
- Everything in Free tier
- Roslyn AST blast-radius analysis (CodeSemantics)
- Skill and Agent Blender
- Hierarchical Louvain/Leiden clustering
- Repository Restructuring
- Pattern Composition
- Test Verification
- Feature Interaction Matrix
High-Assurance
Mission-critical, defense, regulated enterprise, and sovereign teams
Runtime Bounds & Capacity
- Validated Bound 1,000,000 nodes / 5,000,000 edges
- 2-Hop Traversal 2-hop traversal < 12 ms at that scale
- Worker Pool Worker ceiling of 75,000 (auto-scaled to hardware)
Included Subsystems
- Everything in Pro tier
- Level Zero GPU vector-search acceleration
- Pixie Swarm mesh reasoning
- BESS sharding and reasoning
- Hardware Topology
- Formal Invariants (proof contracts)
- World Model
- TAPE Planning & Grounding Verifier
- Sovereign Air-Gap with MLS enforcement
- Capability Sandbox
- Regulatory Compliance (ISO 26262, DO-178C, EU AI Act, HIPAA)
- Multi-Agent Governance and Decision Ledger
Ready to Deploy Hypertensor GraphRAG?
Developer Pro trial keys are provisioned within 24 hours. High-Assurance Enterprise contracts include cryptographic license generator tooling, custom NUMA build optimization, and air-gapped installation bundles.
Runtime Launch Commands
Run Hypertensor GraphRAG as an embedded .NET daemon, CLI query engine, or Python FastMCP bridge.
dotnet run --project src/HFLabs/Products/HypertensorGraphRAG/Server/Server.csproj -c Release Spins up the in-memory graph store and binds zero-copy IPC socket interfaces.
dotnet run --project src/HFLabs/Products/HypertensorGraphRAG/App/App.csproj -c Release -- --headless --query "Director" Executes 2-hop graph reasoning with SIMD cosine scoring and outputs JSON results.
python -m mcp_graphrag.server --transport sse --port 8000 Exposes HT GraphRAG traversal primitives as Model Context Protocol tools for AI agents.