LightningDB C++ Graph Database

Faster than Neo4j on the query an LLM actually runs.

Exact vector search, multi-hop CSR traversal, and finished context packing in one contiguous C++ process.

ADDR 0x0000 // INCOMING VECTOR ADDR 0x1A00 // SIMD & CUDA CORES ADDR 0x4800 // TYPE-SLICED CSR MEMORY ADDR 0x7F00 // PACKED CONTEXT [0.86 ms] Query Vector 1,536-d Exact Cosine SIMD 250,000 rows · 0 approx CSR Hop 1 CSR Hop 2 3-Hop Subgraph 4,200 context nodes 0.86 ms Finished Context Pack

28,000-node corpus (20k people, 8k papers)

LightningDB GPU 0.86 ms

Neo4j Community 9.27 ms

10.8× faster

2.5 million nodes (2M people, 500k papers)

LightningDB GPU 6.48 ms

LightningDB CPU (no GPU) 9.18 ms

Neo4j Community 12.12 ms

1.9× faster

One engine. Three jobs.

Most GraphRAG stacks stitch a standalone vector store to a graph database with network drivers. LightningDB executes all three operations in one contiguous C++ process.

01 // VECTORS Exact search

Every embedding is scored. There is no approximate index and no recall trade-off. At the 1,536-wide vectors used by OpenAI and Cohere, 250,000 rows return in 9.48 ms on a GPU. Neo4j takes 17.3 seconds.

02 // TOPOLOGY Contiguous memory

Cypher queries compile to scans over a type-sliced CSR. A multi-hop walk reads sequential ranges of memory rather than chasing record pointers across disk. Result counts match Neo4j on every published benchmark.

03 // INFERENCE One request

POST /v1/rag returns ranked hits, document bodies, and the nanosecond execution clock. Native LangChain and LlamaIndex retrievers live directly in the repository, serving 629 requests per second across 128 concurrent clients.

Where Neo4j falls behind.

Exact cosine similarity, 250,000 rows, top ten results kept. Scored on CPU and GPU against Neo4j Community on CPU.

Vector width LightningDB CPU LightningDB GPU Neo4j Advantage
32MiniLM class 12.03 ms 2.65 ms 559 ms 211×
768BERT / E5 14.40 ms 6.02 ms 8.65 s 1,440×
1,536OpenAI / Cohere 27.67 ms 9.48 ms 17.3 s 1,830×

Even without a GPU, LightningDB CPU is 12.03 ms to 27.67 ms across the ladder — 46× to 630× ahead of Neo4j. The GPU widens the lead to 1,830×. These clocks stop when the context block is packed. Generating tokens takes two to three seconds with TinyLlama on every database, including ours.

Priced to replace it.

The engine is identical across every tier. What changes is core caps, GPU retrieval, and who runs the hardware.

SELF-HOSTED Free → Pro → Pro+

Free Edition runs on 2 CPUs with community source. Pro adds 1 CUDA GPU and 16 cores for $25 / mo flat. Pro+ scales to 32 cores, 2 CUDA GPUs, and entropy worker for $50 / mo flat. Enterprise unlocks unlimited cores and multi-server clusters.

MANAGED CLOUD From $55 / GB / mo

Fully managed Pro and Pro+, billed per GB of working set — cheaper than Neo4j Aura at $65/GB. At 8 GB Cloud Pro is $440/mo, against $520/mo for Neo4j Aura Professional.

ENTERPRISE Dedicated infrastructure

Unlimited cores, multi-server parallel replication, custom topologies, and a named support engineer. Available both self-hosted and cloud via custom quote.

Run the benchmarks yourself.

The C++ database engine, the benchmark suites, and the automated proof scripts are open on GitHub. Clone the repository and measure the clocks on your own hardware.

# Clone and run the automated proof script
git clone https://github.com/conwayek/GraphDB.git
cd GraphDB
bash scripts/proof_community.sh