Skip to content

Benchmark Workflows ​

SecureVault includes two benchmark workflows under secure-vault/benchmark and secure-vault/scripts/benchmark.

Use them for different questions:

  • Use the retrieval benchmark when you want to show search responsiveness after documents are already indexed.
  • Use the pipeline benchmark when you want to show end-to-end semantic indexing and ranking quality with the live Google embedding path.

What is in the repository ​

  • Evaluator-facing reports and methodology notes live in secure-vault/benchmark.
  • Executable benchmark scripts live in secure-vault/scripts/benchmark.
  • retrieval/ contains the latency benchmark entrypoint and report builders.
  • pipeline/ contains the end-to-end accuracy benchmark entrypoint and report builders.
  • shared/runtime.ts loads .env.local or .env from secure-vault/ and checks MariaDB vector support before a run starts.

Which benchmark to run ​

BenchmarkCommandBest forIncludes live embeddings
Retrieval latencynpm run benchmark:semanticShowing semantic and hybrid search response time after indexingNo
Pipeline accuracynpm run benchmark:semantic:pipelineShowing indexing behavior and retrieval quality end to endYes

Latest benchmark results ​

The latest checked-in reports were generated on 2026-04-20.

Retrieval latency report:

BenchmarkSamplesAvgP50P95P99MaxAvg results
Semantic retrieval only151015.31 ms1005.73 ms1104.04 ms1104.04 ms1104.04 ms10.00
Hybrid retrieval151013.42 ms1016.98 ms1063.83 ms1063.83 ms1063.83 ms10.00

Pipeline accuracy report:

SuiteBenchmarkTop-1 AccuracyTop-3 RecallMRRAvg query timeAvg indexing time
ControlledSemantic100.0%100.0%1.000456.50 ms5691.43 ms
ControlledHybrid100.0%100.0%1.000781.83 ms5691.43 ms
StressSemantic50.0%100.0%0.750749.99 ms6800.25 ms
StressHybrid50.0%100.0%0.750562.53 ms6800.25 ms

Use these numbers as the current evaluator-facing snapshot. Re-run the benchmark scripts before demos when hardware, MariaDB version, embedding settings, or dataset size changes.

Prerequisites ​

Run both commands from secure-vault/.

Both benchmarks need:

  • MariaDB running and reachable through DATABASE_HOST, DATABASE_PORT, DATABASE_USER, DATABASE_PASSWORD, and DATABASE_NAME
  • MariaDB vector support available for VEC_FromText(...) and vec_distance_cosine(...)
  • a local env file in secure-vault/.env.local or secure-vault/.env if the values are not already exported in your shell

The pipeline benchmark also needs:

  • SEMANTIC_INDEXING_ENABLED=true
  • SEMANTIC_INDEXING_PROVIDER=google
  • GEMINI_API_KEY set

Recommended local path for the pipeline benchmark:

env
SEMANTIC_INDEXING_ENABLED=true
SEMANTIC_INDEXING_EXECUTION_MODE=inline
SEMANTIC_INDEXING_PROVIDER=google
GEMINI_API_KEY=<your-key>

inline is the simpler local mode. If you intentionally switch to queued, the semantic config also requires Redis to be configured and available.

Retrieval benchmark ​

Use this benchmark for evaluator-facing latency numbers after indexing is already complete.

Command:

powershell
cd secure-vault
npm run benchmark:semantic

What it does:

  • checks that MariaDB vector functions are available
  • seeds a synthetic benchmark user, files, embedding jobs, and embedding chunks
  • runs semantic-only and hybrid retrieval through the real application search path
  • writes markdown and JSON reports
  • deletes the seeded benchmark data in a cleanup step

Useful flags:

powershell
npm run benchmark:semantic -- --themes 6 --files-per-theme 500 --chunks-per-file 3 --queries-per-theme 5

Available options:

  • --themes
  • --files-per-theme
  • --chunks-per-file
  • --queries-per-theme
  • --warmup-runs
  • --query-top-k
  • --file-batch-size
  • --chunk-batch-size
  • --output-dir

Default output files:

Read the generated report as a responsiveness benchmark, not an accuracy benchmark. It excludes live embedding latency by design.

Pipeline benchmark ​

Use this benchmark when you need evidence that SecureVault can index benchmark documents end to end and still retrieve the correct file.

Command:

powershell
cd secure-vault
npm run benchmark:semantic:pipeline

What it does:

  • checks MariaDB vector support
  • validates that semantic indexing is enabled with the Google provider
  • generates temporary benchmark PDFs
  • chunks and embeds them through the real semantic pipeline
  • stores vectors in MariaDB
  • runs semantic-only and hybrid retrieval for benchmark queries
  • writes markdown and JSON reports

Useful flags:

powershell
npm run benchmark:semantic:pipeline -- --suite stress --themes 4 --files-per-theme 3

Available options:

  • --themes
  • --files-per-theme
  • --suite controlled|stress|both
  • --output-dir

Default output files:

The benchmark runs two suites by default:

  • controlled for clean, direct phrasing and easier topical separation
  • stress for paraphrased queries and confusable same-theme documents

Use the pipeline report for ranking quality discussions. Pair it with the retrieval benchmark if you also want to show search speed.

Reading the reports ​

Use the retrieval report to answer "How fast is search after indexing?"

  • focus on average latency and P95
  • compare semantic-only and hybrid retrieval at the same dataset size

Use the pipeline report to answer "How well does the system retrieve the right file?"

  • focus on Top-1 Accuracy, Top-3 Recall, and MRR together
  • treat controlled as the cleaner correctness benchmark
  • treat stress as the more realistic retrieval benchmark

Built with VitePress and deployed through GitHub Pages.