k6 suites that assert the Sophia knowledge/RAG SLOs documented in
docs/performance/baselines.md §3.
| Script | Exercises | SLO gate |
|---|---|---|
search-load.js |
POST /v1/search, POST /v1/ask |
search p95 ≤ 500 ms, ask p95 ≤ 800 ms |
graph-load.js |
POST /v1/resolution/resolve, POST /v1/graph/subgraph |
resolve p95 ≤ 300 ms, subgraph p95 ≤ 600 ms |
Run locally#
bash
# Core services only — assumes search-api on :3500, knowledge-graph on the same host
docker compose -f docker/docker-compose.dev.yml up -d
# Read-heavy retrieval path
k6 run testing/performance/k6/sophia/search-load.js
# Entity resolution → 2-hop subgraph
k6 run testing/performance/k6/sophia/graph-load.js
Environment variables#
| Var | Default | Purpose |
|---|---|---|
SOPHIA_SEARCH_URL |
http://localhost:3500 |
search-api origin |
SOPHIA_GRAPH_URL |
$SOPHIA_SEARCH_URL |
knowledge-graph origin (set if deployed separately) |
JWT_SECRET |
sophia-load-test-secret |
HS256 secret used to mint per-iteration load-test tokens |
SOPHIA_ASK_RATIO |
0.3 |
Fraction of search iterations that follow up with /ask |
All stage durations and target RPS are overridable via the env vars declared at
the top of each script (SOPHIA_SEARCH_STAGE_1_RPS, etc.) so CI can dial the
load envelope per environment.
What counts as a regression#
A run regresses the Sophia SLO if any of the thresholds blocks in the scripts
fail — k6 exits non-zero, which is how CI gates on this. The budgets are the
same numbers encoded in docs/performance/baselines.md and in the cross-domain
latency Vitest benchmark; changing one requires updating both.