One database to replace your Neo4j+Qdrant+embeddings stack

What it does

NornicDB is a graph database that speaks Neo4j’s Bolt/Cypher and Qdrant’s gRPC, aiming to collapse the typical three-system stack (graph DB + vector store + embedding pipeline) into one deployment. It adds temporal MVCC for historical reads, managed embeddings, HNSW vector search, and hardware acceleration via CUDA, Metal, or Vulkan.

The interesting bit

The project doesn’t just bolt a vector index onto a graph engine. It uses a hybrid execution model with shape-specialized streaming fast paths for common queries, while keeping a general Cypher engine for everything else. A neat optimization: BM25-seeded insertion order cut HNSW build times by ~2.7× in published tests by reducing traversal waste during construction.

Key highlights

• Sub-millisecond local latency for vector search plus one-hop graph traversal (network dominates once you go remote)
• Snapshot isolation with explicit MVCC retention floors; historical reads fail safely below the retained threshold
• Parser mode switching: a fast custom nornic path for production, antlr mode for strict validation
• UCLouvain benchmark: 2.2× faster than Neo4j on automata learning workloads (1,443 state-transition queries)
• LDBC Social Network Benchmark claims up to 52× speedup on specific query shapes vs Neo4j (M3 Max, 64GB)
• MCP server and agent skill files for Claude/workflows in docs/skills/

Caveats

• Docker on macOS does not expose Metal acceleration; GPU on Apple Silicon requires native install
• Search paths are intentionally current-state only; MVCC historical reads are separate
• Benchmarks are self-published; methodology docs exist but independent verification is unclear

Verdict

Worth evaluating if you’re running a Neo4j+Qdrant+embeddings stack for Graph-RAG or agent memory and want to halve your moving parts. Skip if you’re happy with separate systems or need deep Neo4j ecosystem compatibility beyond wire protocol.