Performance and recall latency

Graphnosis recall is local and deterministic — no cloud round-trip — but latency still depends on cortex size, hardware, cache warmth, and whether you search one engram or many.

What the ~75 ms figure means

Marketing and the home-page stats strip cite ~75 ms median recall measured on a single engram with a warm embedding cache, using the local hybrid retrieval path (BGE-small embeddings + graph expansion) on developer hardware. The reference corpus is about 12k nodes / 68k edges — see the SDK benchmarks methodology on GitHub.

That number is not a guarantee for:

• Cold boot — first recall after unlock while embeddings are still loading
• Federated multi-engram recall across many .gai files
• Low-power or embedded hardware without tuning (GRAPHNOSIS_EMBED_WORKERS, etc.)
• Very large cortexes (tens of thousands of nodes per engram) without a fresh baseline

Use it as a representative warm, single-engram datapoint — not a SLA.

CI regression guard (every smoke test)

The desktop-sidecar smoke test ends with a recall-latency-regression phase:

1. Ingests the bundled docs offline into a smoke-latency-bench engram (deterministic corpus).
2. Runs one warm-up recall (not timed).
3. Times 5 recalls (override with GRAPHNOSIS_RECALL_LATENCY_RUNS) and computes P50.
4. Fails if P50 exceeds 200 ms (override with GRAPHNOSIS_RECALL_LATENCY_P50_MS).

The 200 ms bar is an internal regression guard for CI — intentionally generous so flaky runners do not fail PRs. It is separate from the ~75 ms marketing benchmark.

Skip the phase locally when iterating: GRAPHNOSIS_SKIP_RECALL_LATENCY=1.

pnpm --filter @graphnosis-app/desktop-sidecar smoke

Look for JSON lines recall-latency-regression.result and recall-latency-regression.ok in the output.

Manual large-cortex benchmark (pre-release)

Before tagging a release — especially after recall, embedding, or graph-index changes — run the manual script against your real cortex:

pnpm --filter @graphnosis-app/desktop-sidecar build

GRAPHNOSIS_CORTEX="$HOME/Documents/MyCortex" \
GRAPHNOSIS_PASSPHRASE="your-passphrase" \
GRAPHNOSIS_RECALL_BENCH_GRAPH=personal \
node apps/desktop-sidecar/scripts/recall-benchmark-manual.mjs

Checklist

1. Unlock the cortex in the app (or pass GRAPHNOSIS_PASSPHRASE / GRAPHNOSIS_RECOVERY_PHRASE).
2. Wait for background brain / embedding work to finish (status bar quiet, no ingest spinner).
3. Run the script — default 10 timed runs after warm-up; records P50/P95 to stdout.
4. Compare to your last saved baseline (spreadsheet or release notes). Investigate regressions > ~20% before shipping.
5. Optional: set GRAPHNOSIS_RECALL_LATENCY_P50_MS to fail the script on a hard ceiling.

Optional env vars are documented in Environment Variables.

Tuning for constrained hardware

See Enterprise FAQ — resource requirements for RAM/CPU guidance and GRAPHNOSIS_EMBED_WORKERS / GRAPHNOSIS_EMBED_DISABLE trade-offs.