Skip to content

Not MCP proof

This page measures the old TypeScript indexer (MemoryStorage, profile coderag-public-proof) on a fixed fixture. It is not a measurement of the Rust MCP server in @sylphx/locus.

Benchmark Proof ​

Locus public performance claims are backed by a single checked-in script and a fixed fixture corpus — not hand-tuned marketing numbers.

What Gets Measured ​

PhaseWhat it proves
IndexingFiles/sec on fixtures/benchmark-corpus/ (30 TypeScript files, AST chunking)
SearchTF-IDF search p50 / min / max latency over 20 iterations (3 warmup)

The benchmark uses in-memory storage (MemoryStorage) and lowMemoryMode: false to measure best-case local search latency on the fixture corpus. Results are machine- and fixture-specific.

Reproduce ​

bash
bun install
bun run build
bun run benchmark:public-proof

The script prints JSON to stdout and a human-readable summary to stderr.

Example output shape:

json
{
  "profile": "coderag-public-proof",
  "corpus": {
    "fixtureTypeScriptFiles": 30,
    "indexedFiles": "<from script>",
    "indexedChunks": "<from script>"
  },
  "indexing": {
    "durationMs": "<from script>",
    "filesPerSecond": "<from script>"
  },
  "search": {
    "iterations": 20,
    "warmupIterations": 3,
    "p50Ms": "<from script>",
    "minMs": "<from script>",
    "maxMs": "<from script>",
    "avgMs": "<from script>"
  }
}

Note: Run the script on your machine and paste the output here when updating release notes. Do not edit the numbers in README or this page without running bun run benchmark:public-proof.

Latest Run ​

Run bun run benchmark:public-proof locally to populate this table for your environment. Commit updated numbers only from actual script output.

MetricResult
Fixture corpus30 TypeScript files → 31 indexed files, 71 chunks
Indexing throughput186.1 files/sec (166.6 ms total; run 2 of 2)
Search p50 (TF-IDF)0.0 ms (20 iterations, 3 warmup)
Search min / max0.0 ms / 0.5 ms (avg 0.0 ms)

What This Proves ​

  • Indexing is bounded on a fixed corpus — throughput is measurable and reproducible.
  • Search latency is bounded on a warm index — p50/min/max come from 20 timed runs.
  • Claims stay honest — README performance section links here; no invented MacBook tables.

Fixture Corpus ​

The corpus lives at fixtures/benchmark-corpus/src/ and covers auth, database, API, middleware, utilities, and services modules. Do not edit fixtures without re-running the benchmark.

Released under the MIT License.