brainy/tests/benchmarks/brainy-scale.js
David Snelling c605b34f98 test(8.0): A/B benchmark harness (open leg) — generic corpus+metrics lib, brainy-alone scaling bench, boundary guard, real-embedding recall guard
The open-core, Cortex-free half of the library A/B (handoff AJ/AK), authored once in
brainy so the proprietary A/B comparison can import it for both legs:

- tests/benchmarks/lib/corpus.js — deterministic clustered-mixture corpus generator
  (recompute-on-demand, O(clusters·dim) memory) for latency/ingest/memory at scale.
- tests/benchmarks/lib/metrics.js — percentiles, brute-force recall@k, RSS snapshot.
- tests/benchmarks/brainy-scale.js — brainy-alone scaling leg (ingest, find p50/p99 for
  vector/metadata/graph/triple, RSS). Recall is intentionally NOT measured on synthetic
  data — see below.
- tests/unit/boundary-no-cortex.test.ts — CI guard: fails if @soulcraft/cortex ever
  appears in a src/ or tests/ import or in any package.json dependency field.
- tests/integration/vector-recall.test.ts — semantic-search correctness on REAL
  embeddings (19-20/20 exact-text top-1).

Methodology note: synthetic vectors (random/one-hot/clustered/latent) are near-orthogonal
under cosine, so HNSW (any graph ANN, incl. DiskANN) cannot navigate them and recall
collapses regardless of engine — a property of the data, not the index. Brainy vector
search is verified correct on real embeddings. The A/B recall@10 column is therefore
measured on SIFT/BIGANN, identically for both legs.
2026-06-15 15:51:17 -07:00

124 lines
6 KiB
JavaScript

#!/usr/bin/env node
/**
* @module tests/benchmarks/brainy-scale
* @description Brainy-alone scaling benchmark — the OPEN-CORE leg of the
* library A/B (the proprietary side runs this same workload with the native
* provider registered and computes the delta). Pure TypeScript backends only;
* no Cortex, no native code, no model load (vectors are precomputed).
*
* Measures the AI.2 metric set through Brainy's PUBLIC API (`add`/`find`):
* ingest throughput, `find()` p50/p99 (vector and the triple-intelligence path),
* recall@10 vs brute-force ground truth, and peak RSS. Doubles as Brainy's own
* perf-regression guard.
*
* Usage:
* node --max-old-space-size=8192 tests/benchmarks/brainy-scale.js [N]
* DIM=128 node ... brainy-scale.js 1000000 # match a 128-dim baseline
*
* Emits a `JSON {…}` line for machine capture. Every query type asserts a
* non-empty result so an empty (misleadingly fast) path fails loudly.
*/
import { Brainy } from '../../dist/index.js'
import { NounType, VerbType } from '../../dist/types/graphTypes.js'
import { makeCorpus } from './lib/corpus.js'
import { summarize, memSnapshot } from './lib/metrics.js'
const N = Number(process.argv[2] ?? 100_000)
const DIM = Number(process.env.DIM ?? 384)
const QUERIES = 200
async function bench(label, iters, fn, expectNonEmpty = true) {
const lat = []
let empty = 0
for (let i = 0; i < iters; i++) {
const t = process.hrtime.bigint()
const r = await fn(i)
lat.push(Number(process.hrtime.bigint() - t) / 1e6)
if (!r || r.length === 0) empty++
}
const s = summarize(lat)
const flag = expectNonEmpty && empty === iters ? ' ⚠️ ALL EMPTY' : ''
console.log(`${label.padEnd(30)} p50 ${s.p50.toFixed(2).padStart(8)}ms p95 ${s.p95.toFixed(2).padStart(8)}ms p99 ${s.p99.toFixed(2).padStart(8)}ms (empty ${empty}/${iters})${flag}`)
return { ...s, empty }
}
async function main() {
console.log('='.repeat(96))
console.log(`Brainy 8.0 OPEN-CORE scaling leg @ N=${N.toLocaleString()} dim=${DIM} (memory storage, JS backends)`)
console.log('='.repeat(96))
const corpus = makeCorpus({ n: N, dim: DIM })
const brain = new Brainy({ storage: { type: 'memory' }, eagerEmbeddings: false, requireSubtype: false, silent: true })
await brain.init()
// ---- Ingest ---------------------------------------------------------------
// Explicit deterministic ids: id(i) holds vector(i). This is independent of
// addMany completion order (parallel batches return ids out of submission
// order), which both recall ground-truth and the edge graph depend on.
const CHUNK = 5000
const t0 = Date.now()
let batch = [], written = 0
for (let i = 0; i < N; i++) {
batch.push({ id: corpus.id(i), vector: corpus.vector(i), type: NounType.Document, metadata: corpus.metadata(i) })
if (batch.length === CHUNK) {
const r = await brain.addMany({ items: batch, parallel: true, chunkSize: 500 })
written += r.successful.length
batch = []
if (written % 100_000 === 0) console.log(` ingested ${written.toLocaleString()} (${Math.round(written / ((Date.now() - t0) / 1000)).toLocaleString()}/s)`)
}
}
if (batch.length) {
const r = await brain.addMany({ items: batch, parallel: true, chunkSize: 500 })
written += r.successful.length
}
const buildMs = Date.now() - t0
const ingestRate = Math.round(written / (buildMs / 1000))
console.log(`ingest: ${written.toLocaleString()} in ${(buildMs / 1000).toFixed(1)}s (${ingestRate.toLocaleString()}/s)`)
// ---- Edges (for the graph + triple path) ----------------------------------
let edges = 0
for (let h = 0; h < corpus.hubs; h++) {
for (let f = 0; f < corpus.fanout; f++) {
await brain.relate({ from: corpus.id(h), to: corpus.id(corpus.neighborIndex(h, f)), type: VerbType.References, weight: 0.8 })
edges++
}
}
console.log(`edges: ${edges.toLocaleString()}`)
// ---- Warmup ---------------------------------------------------------------
for (let i = 0; i < 20; i++) await brain.find({ vector: corpus.queryVector(i), limit: 10 })
// ---- Latency --------------------------------------------------------------
console.log('-'.repeat(96))
const vector = await bench('vector knn (k=10)', QUERIES, (i) => brain.find({ vector: corpus.queryVector(i), limit: 10 }))
const metadata = await bench('metadata filter (category)', QUERIES, (i) => brain.find({ where: { category: i % corpus.categories }, limit: 10 }))
const graph = await bench('graph 1-hop out', QUERIES, (i) => brain.find({ connected: { from: corpus.id(i % corpus.hubs), via: VerbType.References, direction: 'out', depth: 1 }, limit: 10 }))
const triple = await bench('triple vec+meta+graph', QUERIES, (i) => {
const hub = i % corpus.hubs
const nIdx = corpus.neighborIndex(hub, 0)
return brain.find({ vector: corpus.vector(nIdx), where: { category: nIdx % corpus.categories }, connected: { from: corpus.id(hub), via: VerbType.References, direction: 'out', depth: 1 }, limit: 10 })
})
// NOTE: recall@10 is intentionally NOT measured here. This synthetic
// clustered corpus is built for latency/ingest/memory at scale; its tight
// clusters make intra-cluster points near-cosine-identical, so the "exact
// top-10" is ill-defined and recall vs brute-force is dominated by tie-breaks,
// not index quality. The A/B recall@10 column is measured on a REAL dataset
// (SIFT/BIGANN, which has genuine neighbour structure + canonical ground
// truth), identically for both legs. Index correctness is guarded separately
// by tests/integration/vector-recall.test.ts.
// ---- Memory ---------------------------------------------------------------
const mem = memSnapshot(written)
console.log(`memory: RSS ${mem.rssGB} GB (${mem.rssBytesPerEntity.toLocaleString()} B/entity)`)
console.log('='.repeat(96))
await brain.close()
console.log('JSON ' + JSON.stringify({
leg: 'brainy-open-core', n: written, dim: DIM, edges, ingestPerSec: ingestRate,
rssGB: mem.rssGB, bytesPerEntityRss: mem.rssBytesPerEntity,
vector, metadata, graph, triple
}))
}
main().catch((e) => { console.error(e); process.exit(1) })