10.8×

The query that feeds the model.

Single-request GraphRAG: exact vector search followed immediately by a multi-hop CSR graph walk. LightningDB measured directly against Neo4j on the identical query workload and corpus.

20,000 people and 8,000 papers (28,000 nodes)

LightningDB GPU 0.86 ms

Neo4j Community 9.27 ms

10.8× faster

2 million people and 500,000 papers (2.5 million nodes)

LightningDB GPU 6.48 ms

LightningDB CPU (no GPU) 9.18 ms

Neo4j Community 12.12 ms

1.9× faster

What LightningDB runs.

Exact cosine on AVX2 or CUDA followed by an undirected three-hop CSR traversal over 4,200 context nodes in a single POST /v1/rag request. No driver roundtrips and zero intermediate deserialization.

What Neo4j runs.

An approximate vector index lookup followed by a multi-hop Bolt Cypher traversal on localhost using the identical seed, corpus, and hop depth. Result node counts match on every row.

Exact cosine, 32 to 1,536 dimensions.

250,000 rows, top ten kept. LightningDB scores every vector with zero approximation.

Vector width LightningDB CPU LightningDB GPU Neo4j Community Advantage
32MiniLM class 12.03 ms 2.65 ms 559 ms 211×
256 10.73 ms 3.74 ms 3.89 s 1,040×
384MiniLM-L12 11.48 ms 4.25 ms 4.25 s 1,000×
768BERT / E5 14.40 ms 6.02 ms 8.65 s 1,440×
1,024 15.36 ms 7.37 ms 14.6 s 1,980×
1,536OpenAI / Cohere 27.67 ms 9.48 ms 17.3 s 1,830×

High-throughput parallel execution.

Production demands non-blocking parallel execution under load. LightningDB sustains 629 completed queries per second across 128 parallel calls with zero queue delay and zero drops.

MetricResult
Parallel callers128
Completed requests2,048 / 2,048
Dropped or delayed0
Throughput629 / s

Rerun it on Wikipedia.

Run the reproducible harness on the public SNAP Wikipedia corpus with CPU and GPU lanes.

# Run the community proof script
bash scripts/proof_community.sh