| import { describe, expect, it } from "vitest"; |
| import { dequantizeEmbedding, quantizeNormalizedEmbedding } from "./embedding-quantization.js"; |
| import { cosineSimilarity, createMemoryEmbedding } from "./memory-vectors.js"; |
|
|
| describe("embedding quantization", () => { |
| it("round-trips normalized embeddings with bounded cosine loss", () => { |
| const embedding = createMemoryEmbedding({ |
| content: "preserve semantic similarity for holographic memory resonance", |
| metadata: { area: "memory", mode: "experimental" }, |
| }); |
| const quantized = quantizeNormalizedEmbedding({ embedding, bits: 6 }); |
|
|
| expect(quantized).not.toBeNull(); |
| const restored = dequantizeEmbedding(quantized!); |
| expect(restored).toHaveLength(embedding.length); |
| expect(cosineSimilarity(embedding, restored)).toBeGreaterThan(0.92); |
| }); |
|
|
| it("preserves top-1 semantic retrieval for most benchmark queries", () => { |
| const corpus = [ |
| "stabilize telegram topic routing for forum threads", |
| "compress semantic fossils with online vector quantization", |
| "repair gateway session subscriptions for live transcript updates", |
| "measure continuity drift across autonomous study cycles", |
| "tighten ACP tool approval against spoofed hints", |
| "propagate tool events from agent runtime into gateway observers", |
| "restore robust config path resolution with legacy fallback", |
| "track benchmark agreement between omega policy and kernel", |
| "recover preview-finalized telegram hooks for observability", |
| "preserve internal project focus across awakenings", |
| "reduce world model blast radius by splitting synchronization phases", |
| "record research hypotheses only when correlation crosses threshold", |
| ].map((content, index) => ({ |
| id: `doc-${index}`, |
| content, |
| embedding: createMemoryEmbedding({ content, metadata: { area: "benchmark", index } }), |
| })); |
|
|
| const restoredCorpus = corpus.map((entry) => ({ |
| ...entry, |
| restored: |
| dequantizeEmbedding( |
| quantizeNormalizedEmbedding({ |
| embedding: entry.embedding, |
| bits: 6, |
| })!, |
| ) ?? entry.embedding, |
| })); |
|
|
| const queries = [ |
| "forum thread routing for telegram topics", |
| "online vector quantization for memory fossils", |
| "live transcript session subscription in gateway", |
| "continuity drift in autonomous cycles", |
| "tool approval spoofing in ACP", |
| "gateway tool event observers", |
| "legacy config path fallback", |
| "omega versus kernel benchmark agreement", |
| "telegram preview finalization observability", |
| "focus continuity across awakenings", |
| "split world model synchronization phases", |
| "research correlation threshold for hypotheses", |
| ]; |
|
|
| let matches = 0; |
| for (let index = 0; index < queries.length; index += 1) { |
| const queryEmbedding = createMemoryEmbedding({ |
| content: queries[index]!, |
| metadata: { area: "benchmark-query", index }, |
| }); |
| const denseTop = corpus |
| .map((entry) => ({ |
| id: entry.id, |
| score: cosineSimilarity(queryEmbedding, entry.embedding), |
| })) |
| .sort((left, right) => right.score - left.score)[0]; |
| const quantizedTop = restoredCorpus |
| .map((entry) => ({ |
| id: entry.id, |
| score: cosineSimilarity(queryEmbedding, entry.restored), |
| })) |
| .sort((left, right) => right.score - left.score)[0]; |
|
|
| if (denseTop?.id === quantizedTop?.id) { |
| matches += 1; |
| } |
| } |
|
|
| expect(matches / queries.length).toBeGreaterThanOrEqual(0.83); |
| }); |
| }); |
|
|
