Exact search
Every embedding is scored. There is no approximate index and no recall setting. 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.
10.8×
Faster than Neo4j on the query an LLM actually runs.
LightningDB appears as your standard Cypher graph network, built on a lightning-fast C++ underlying architecture. On a 28,000-node corpus, it finds the nearest documents, walks three hops, and returns a finished context block in 0.86 ms. Neo4j takes 9.27 ms.
Improving AI-RAG inferencing by factors
LightningDB GPU0.86 ms
Neo4j9.27 ms
10.8×
2.5 million nodes
LightningDB GPU6.48 ms
LightningDB CPU9.18 ms
Neo4j12.12 ms
1.9×
One engine. Three jobs.
Most GraphRAG stacks are a vector store, a graph database, and the glue between them. LightningDB does the work in one process, on one contiguous layout.
Contiguous memory
Cypher compiles to scans over a type-sliced CSR. A multi-hop walk reads ranges of memory. It does not chase pointers through a record store. Result counts match Neo4j on every published row.
One request
POST /v1/rag returns the hits, the documents, and the clock. LangChain and LlamaIndex retrievers are in the repository. High-throughput non-blocking architecture serving 629 requests per second.
Where Neo4j falls behind.
Exact cosine, 250,000 rows, ten results. Scored on CPU and GPU against Neo4j 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 same ladder — 46× to 630× ahead of Neo4j. The GPU widens the lead to 1,830×. These clocks stop when the context is packed. Generating tokens takes two to three seconds with TinyLlama on every database, including ours.
Priced to replace it.
The engine is the same in every edition. What changes is cores, GPU retrieval, and who runs the machine.
Free
Private. You run it, on two CPUs, from the community source.
$35 / GB / month
Professional. A production license, 32 CPUs, business-hours support. At 8 GB that is $280, against $520 for Aura Professional.
Enterprise and Cloud
Enterprise adds GPU retrieval, replicas, and a named engineer. Cloud, run by us, starts at $299 a month.
Run the benchmarks yourself.
The engine, the benchmarks, and the automated proof script are on GitHub. Try here on GitHub to clone the repository and run the clocks on your own hardware.
Every published benchmark against Neo4j is open source and reproducible with our public Docker scripts.