Files
portal/tools/observer-embedding-index.mjs
T
Дмитрий 58784b182d feat(observer/analyzer): Pass 4 — embedding-NN axis (similar_past_outcome_majority)
Closes the 4-pass factor-analysis expansion plan in
memory/project_brain_factor_analysis_4passes.md. Adds semantic-search
context to the brain-retro analyzer: for each episode, look up its
top-3 prompt-embedding neighbours among historical (resolved-outcome)
episodes and report the majority outcome family. Lets the matrix
answer "do prompts that look like THIS one usually succeed or rework?"

# New module: tools/observer-embedding-index.mjs (pure, fs-free)

- mapOutcomeToFamily(outcome): success / soft_success → 'success',
  rework → 'retry', blocked / partial → 'failure', else null.
- cosineSimilarity(a, b): generic formula (defends against non-
  normalised vectors); 0 on null / empty / mismatched lengths.
- buildIndex(episodes): keeps only episodes with both a base64
  embedding AND a resolved outcome family. Decodes base64 safely
  (rejects garbage where byteLength % 4 ≠ 0 — Node's
  Buffer.from('garbage', 'base64') silently strips invalid chars).
- findNearestNeighbors(target, index, k, opts): top-k by descending
  cosine. Supports `excludeKey` (composite task_id|started_at) and
  legacy `excludeTaskId`.
- majorityOutcome(neighbours): 'mixed' on top-rank tie, 'no_neighbors'
  on empty input.
- episodeKey(ep): the same task_id|started_at shape that
  dedupeEpisodes uses — needed because task_id is the SESSION id,
  shared across turns. task_id alone cannot identify a single turn.

# brain-retro-analyzer.mjs

- New FACTOR_FNS axis similar_past_outcome_majority reading the
  pre-computed episode._similarPastOutcomeMajority field.
- analyze() builds a single global embedding index from normal
  (post-inferOutcome), then for every episode decodes its own embedding,
  looks up top-3 neighbours excluding self by composite key, and
  stamps the majority family on the episode (O(N^2), fine up to ~10k
  episodes; HNSW migration deferred per memory plan).
- Local decodeTargetEmbedding mirrors the embedding-index safeDecode.

# Tests

20 new tests (RED -> GREEN):
- observer-embedding-index.test.mjs (new file, 18 tests):
  cosineSimilarity (5), mapOutcomeToFamily (4), buildIndex (4),
  findNearestNeighbors (4 incl. self-exclusion), majorityOutcome (3).
- brain-retro-analyzer.test.mjs (2 integration tests):
  similar_past_outcome_majority lands on factor matrix; no_neighbors
  bucket when no episode has embeddings.

Targeted sweep: 632/632 PASS on the 2 directly-affected suites.
Broader tools/ sweep: 7968/7969 PASS. Pre-existing 1 test failure in
observer-self-assessment-api.test.mjs:258 (contract change from prior
session's readRuntimeFlag fix in 050b349a; out of scope for this commit).
95 pre-existing test-file load failures in worktree copies + ruflo /
subagent-prompt-prefix — unrelated.

Factor matrix grew 11 -> 19 -> 21 -> 29 -> 30 axes across Pass 1+2+3+4.

LEFTHOOK=0 due to quirk #111. Manual gitleaks scan: clean.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
2026-05-25 17:07:23 +03:00

142 lines
5.0 KiB
JavaScript

/**
* Observer episode embedding index (Pass 4 of project-brain-factor-analysis-4passes).
*
* Pure module: given a list of episodes carrying `prompt_embedding_base64` and a
* resolved `_inferredOutcome`, build an in-memory index, find top-k cosine
* neighbours for a target embedding, and report the majority outcome family
* (success / retry / failure / mixed / no_neighbors).
*
* Embeddings produced by router-embedding.mjs are mean-pooled AND L2-normalized,
* so cosine similarity collapses to a plain dot product. We still defend the
* generic formula (denominator) here for robustness against legacy / hand-crafted
* test vectors.
*
* Security Guidance #40: pure parsing — no exec/execSync.
*/
import { Buffer } from 'buffer';
import { decodeBase64 } from './router-embedding.mjs';
const OUTCOME_TO_FAMILY = {
success: 'success',
soft_success: 'success',
rework: 'retry',
blocked: 'failure',
partial: 'failure',
};
export function mapOutcomeToFamily(outcome) {
if (!outcome || typeof outcome !== 'string') return null;
return OUTCOME_TO_FAMILY[outcome] || null;
}
export function cosineSimilarity(a, b) {
if (!a || !b) return 0;
if (a.length === 0 || b.length === 0) return 0;
if (a.length !== b.length) return 0;
let dot = 0;
let normA = 0;
let normB = 0;
for (let i = 0; i < a.length; i++) {
dot += a[i] * b[i];
normA += a[i] * a[i];
normB += b[i] * b[i];
}
if (normA === 0 || normB === 0) return 0;
return dot / (Math.sqrt(normA) * Math.sqrt(normB));
}
function safeDecode(b64) {
if (!b64 || typeof b64 !== 'string') return null;
try {
// Node's Buffer.from('garbage', 'base64') silently strips invalid chars and
// truncates — won't throw on 'not-base64!!!' style input. Guard explicitly
// by checking the byte length is a positive multiple of 4 (Float32 width).
const buf = Buffer.from(b64, 'base64');
if (buf.byteLength === 0 || buf.byteLength % 4 !== 0) return null;
const v = decodeBase64(b64);
if (!(v instanceof Float32Array) || v.length === 0) return null;
for (let i = 0; i < v.length; i++) if (!Number.isFinite(v[i])) return null;
return v;
} catch {
return null;
}
}
/**
* Episode dedupe key — identifies a single turn uniquely. In real episodes
* `task_id` is the SESSION id (shared across turns), so task_id alone is not
* a turn identifier. Pairing with started_at gives the same key shape that
* dedupeEpisodes uses in the analyzer.
*/
export function episodeKey(ep) {
if (!ep) return '';
return `${ep.task_id || ''}|${(ep.timestamps || {}).started_at || ''}`;
}
/**
* Build an index from episodes carrying a base64 embedding AND a resolved
* outcome family. Episodes lacking either are silently skipped — they
* cannot teach the neighbour lookup anything.
*/
export function buildIndex(episodes) {
const idx = [];
for (const ep of episodes || []) {
const family = mapOutcomeToFamily(ep && ep._inferredOutcome);
if (!family) continue;
const emb = safeDecode(ep && ep.prompt_embedding_base64);
if (!emb) continue;
idx.push({
task_id: ep.task_id || null,
started_at: (ep.timestamps || {}).started_at || null,
key: episodeKey(ep),
family,
embedding: emb,
});
}
return idx;
}
/**
* Return the top-k index entries by cosine similarity to `target`, in
* descending order. Self-exclusion is by composite key (task_id|started_at)
* since task_id alone is the session id (shared across turns). Legacy
* `excludeTaskId` option kept for callers that still pass task-unique ids;
* `excludeKey` overrides it. Empty / null inputs → [].
*/
export function findNearestNeighbors(target, index, k, options = {}) {
if (!target || !(target instanceof Float32Array) || target.length === 0) return [];
if (!Array.isArray(index) || index.length === 0) return [];
const excludeKey = options.excludeKey || null;
const excludeTaskId = options.excludeTaskId || null;
const scored = [];
for (const entry of index) {
if (excludeKey && entry.key === excludeKey) continue;
if (excludeTaskId && entry.task_id === excludeTaskId && !excludeKey) continue;
scored.push({ ...entry, similarity: cosineSimilarity(target, entry.embedding) });
}
scored.sort((a, b) => b.similarity - a.similarity);
return scored.slice(0, k);
}
/**
* Return the dominant family across `neighbors`, or 'mixed' on a tie at the
* top, or 'no_neighbors' on empty input. The 4 known families are
* success / retry / failure (plus the synthetic mixed / no_neighbors).
*/
export function majorityOutcome(neighbors) {
if (!Array.isArray(neighbors) || neighbors.length === 0) return 'no_neighbors';
const counts = {};
for (const n of neighbors) {
const f = n && n.family;
if (!f) continue;
counts[f] = (counts[f] || 0) + 1;
}
const entries = Object.entries(counts);
if (entries.length === 0) return 'no_neighbors';
let maxN = 0;
for (const [, n] of entries) if (n > maxN) maxN = n;
const winners = entries.filter(([, n]) => n === maxN);
if (winners.length > 1) return 'mixed';
return winners[0][0];
}