Performance / evidence
The database is not the bottleneck.
Measured Amana latency shows that the long tail comes primarily from model-provider calls, while deterministic Java/PostgreSQL commerce paths remain comparatively small.
A / Provider-backed Buyer path
The model call owns the tail.
Multilingual utterance to pre-cart discovery, with real Gemini intent and embedding providers.
Provider GEMINIIntent gemini-3.1-flash-liteEmbedding gemini-embedding-2Ranker hybrid-v3Sample n=12
PROVIDER / PIPELINE
Gemini intent
- p50
- 3.24 s
- p95
- 13.30 s
PROVIDER / PIPELINE
Query embedding
- p50
- 542.2 ms
- p95
- 690.8 ms
LOCAL DETERMINISTIC
Java + PostgreSQL retrieval
- p50
- 32.8 ms
- p95
- 45.3 ms
Estimated as hybrid-v3 wall time minus the separately timed embedding call.
PROVIDER / PIPELINE
Total pre-cart discovery
- p50
- 3.84 s
- p95
- 13.84 s
Percentiles describe the same small recall cohort but are not additive; no percentage-of-total claim is made. Gemini intent p95 is 13.30 s, versus 45.3 msfor estimated local deterministic retrieval.
Retrieval performance
Retrieval is two very different costs.
External embedding latency is materially larger than local scoring and database work in this measured run.
- p95
- 45.3 ms
- n
- 12
- p95
- 690.8 ms
- n
- 12
- p95
- 720.9 ms
- n
- 12
- 01FTSPostgreSQL full-text candidates
- 02TrigramTypo and near-string evidence
- 03pgvectorOne query vector candidate pass
- 04QualificationHybrid score, then local semantic reuse
- 05Exact gatesIdentity, budget, safety, and authority remain deterministic
B / Local deterministic + stub measurements
Commerce authority stays comparatively small.
PostgreSQL 17 Testcontainers with deterministic merchant and payment adapters; one warm-up journey excluded.
Local catalogue retrieval p5010.3 ms
Local catalogue retrieval p9510.9 ms
n=10 · testcontainers postgresql 17 local stub providers
- p50
- 20.2 ms
- p95
- 21.5 ms
- p50
- 10.4 ms
- p95
- 10.9 ms
- p50
- 15.4 ms
- p95
- 17.0 ms
- p50
- 14.2 ms
- p95
- 25.9 ms
- p50
- 12.8 ms
- p95
- 17.2 ms
- p50
- 8.5 ms
- p95
- 10.1 ms
The same artifact also measures the local intent stub at 7.9 ms p50 / 8.4 ms p95. Gemini and live Razorpay latency are excluded.
Concurrent correctness
Convergence, not a throughput claim.
Critical operations were called concurrently at N=8 and N=32 against PostgreSQL/Testcontainers.
| Operation | Callers | Rows expected / observed | Provider calls expected / observed | Result |
|---|---|---|---|---|
| Execution Reservation | N=8 | 1 / 1 | 0 / 0 | Pass |
| Provider Order Creation | N=8 | 1 / 1 | 1 / 1 | Pass |
| Webhook Ingestion | N=8 | 1 / 1 | 0 / 0 | Pass |
| Transactional Outbox Claim | N=8 | 1 / 1 | 1 / 1 | Pass |
| Refund Ledger Reservation | N=8 | 1 / 1 | 0 / 0 | Pass |
| Execution Reservation | N=32 | 1 / 1 | 0 / 0 | Pass |
| Provider Order Creation | N=32 | 1 / 1 | 1 / 1 | Pass |
| Webhook Ingestion | N=32 | 1 / 1 | 0 / 0 | Pass |
| Transactional Outbox Claim | N=32 | 1 / 1 | 1 / 1 | Pass |
| Refund Ledger Reservation | N=32 | 1 / 1 | 0 / 0 | Pass |
What this measuresCorrectness and convergence under concurrent callers.What it does not measureMaximum production throughput or production-scale load.
Evidence boundaries
What these numbers do not prove.
- These measurements are not production SLAs.
- Production-scale throughput/load testing has not been performed.
- No maximum throughput is claimed.
- Provider and model revisions, quota, and network conditions can change latency.
- Render cold starts are not represented in local deterministic metrics.
- The provider-backed pipeline stops at pre-cart discovery; it does not execute authorization or payment.
- No deployed smoke latency was measured for this page.
Generated at via pnpm proof:evaluate:multilingual.
Continue reviewing