Experimental · performance contract
Benchmark numbers
need their evidence.
ORAG runs a fixed no-key SDK workload, then accepts its baseline only when workload, data, runtime, build and load are recorded and verifiable.
orag.performance-baseline.v1
strict JSONObserved workflow
Run the real path.
The runner uses the public Go SDK to ingest three documents, warm up, measure twenty queries, and run an evaluation. It derives throughput, p50/p95 and cache hits from observed results.
What is rejected
No apples-to-oranges.
Unknown fields, real-provider ambiguity, weak sample counts, invalid percentiles and non-recomputable throughput fail verification.
No-key path
Generate, then verify.
make benchmark-report-run BENCHMARK_REPORT=.tmp/baseline.json
make benchmark-report-verify BENCHMARK_REPORT=.tmp/baseline.json
make console-real-tutorial-benchmark-e2eThe report is regression evidence for the disclosed machine and build, never a production or cross-hardware claim.
Public baseline evidence
Inspect one real local run.
The 2026-07-17-darwin-arm64-main-75eda8f evidence records the exact mock report, safe machine disclosure, manifest and SHA-256 checksums for main commit 75eda8f. It is local regression evidence, not a production or cross-hardware benchmark.
Open checksums →