Benchmarks
Benchmarks are only useful when you can see the methodology. This page reports the measurements CGraph publishes, states how they're produced, and is explicit about what is not yet measured — no invented numbers.
Adaptive gather: recall vs. token cost
The question adaptive gather answers is whether targeting the third hop buys back most of the recall of a full 3-hop gather at a fraction of the token cost.
| Strategy | Grade-2 recall | Candidate tokens |
|---|---|---|
| Adaptive third hop | +0.057 | +13% |
| Full 3-hop gather | (baseline recall) | +96% |
The takeaway: adaptive gather recovers most of the recall of a full third hop for roughly one-seventh of the extra tokens.
Reproducing a query benchmark
graphd includes a one-off benchmark mode to time a query against a built graph:
graphd --benchmark-query --graph cgraph-out --query "PaymentProcessor"This is the supported path for measuring query behavior on your own graph rather than relying on a headline number.
What is not yet published
Beyond the adaptive-gather result above, CGraph does not publish a full benchmark suite (extraction throughput, query latency distributions, memory footprint, scaling curves) in a form we can cite. Those tables are intentionally left pending — this page will grow only with real, reproducible measurements and their methodology, never with placeholder figures.
Where to go next
- Performance — tuning guidance grounded in the same mechanics.
- Context Packing — the gather strategy the benchmark evaluates.