@adia-ai/a2ui-retrieval 0.4.5 → 0.4.7

package/embedding/embedding-retriever.js

@@ -1,152 +0,0 @@
-/**
- * Embedding retriever — loads the build-time pattern embedding index and
- * scores a query against every pattern via cosine similarity.
- *
- * Index: packages/a2ui/corpus/pattern-embeddings.json (built by
- * scripts/build-embeddings.mjs). When missing or empty, the retriever is
- * effectively a no-op — callers see a scoreFor(name) of 0 and should fall
- * back to keyword-only ranking.
- *
- * Query embedding uses the same provider as the index (baked in at build),
- * so the query path also requires the provider's API key at runtime. If
- * the key is absent, the retriever exposes a `available()` probe returning
- * false and callers degrade gracefully.
- */
-
-import { detectProvider, cosine, voyage, openai } from './embedding-provider.js';
-
-const IS_NODE = typeof process !== 'undefined' && !!process.versions?.node;
-
-let _index = null;
-let _indexByName = null;   // Map<string, Float32Array>
-let _loadPromise = null;
-let _embedFn = null;
-let _available = null;      // lazy, once the first probe runs
-
-async function _loadIndex() {
-  if (_index) return _index;
-  if (_loadPromise) return _loadPromise;
-  _loadPromise = (async () => {
-    try {
-      if (IS_NODE) {
-        const fs = await import(/* @vite-ignore */ 'node:fs/promises');
-        const path = await import(/* @vite-ignore */ 'node:path');
-        const url = await import(/* @vite-ignore */ 'node:url');
-        // Try the package-import path first (works under node_modules
-        // install layout — `@adia-ai/a2ui-corpus` exports
-        // `./pattern-embeddings`). Fall back to the relative source-tree
-        // path so a source-checkout monorepo without symlinked deps
-        // still resolves.
-        let p = null;
-        try {
-          const { createRequire } = await import(/* @vite-ignore */ 'node:module');
-          const require = createRequire(import.meta.url);
-          p = require.resolve('@adia-ai/a2ui-corpus/pattern-embeddings');
-        } catch {
-          const here = path.dirname(url.fileURLToPath(import.meta.url));
-          p = path.resolve(here, '../../corpus/pattern-embeddings.json');
-        }
-        const raw = await fs.readFile(p, 'utf8');
-        _index = JSON.parse(raw);
-      } else {
-        const url = new URL('../../corpus/pattern-embeddings.json', import.meta.url);
-        const res = await fetch(url).catch(() => null);
-        _index = res?.ok ? await res.json().catch(() => null) : null;
-      }
-    } catch {
-      _index = null;
-    }
-    if (_index?.patterns?.length) {
-      _indexByName = new Map();
-      for (const p of _index.patterns) {
-        if (p?.name && Array.isArray(p.vector)) {
-          _indexByName.set(p.name, Float32Array.from(p.vector));
-        }
-      }
-    }
-    return _index;
-  })();
-  return _loadPromise;
-}
-
-/** Resolve the embed function matching the index's provider. */
-function _resolveEmbed(providerName, model) {
-  // The index's recorded (provider, model) is the source of truth — query
-  // embeddings MUST be generated by the same model the corpus was indexed
-  // with. Cross-provider mixing produces meaningless cosine scores; even
-  // same-provider/different-model emits different-dim vectors which cosine()
-  // short-circuits to 0 — silent retrieval failure. Fail loud (warn + null)
-  // when the recorded provider's key is unset, instead of auto-picking a
-  // different provider that would silently corrupt similarity rankings.
-  // (See chunk-embedding-retriever.js for the parallel implementation.)
-  let fn = null;
-  if (providerName === 'voyage') fn = voyage({ model });
-  else if (providerName === 'openai') fn = openai({ model });
-  else {
-    // No provider recorded in index header (legacy index) — fall through.
-    const auto = detectProvider();
-    return auto?.embed || null;
-  }
-  if (!fn && typeof console !== 'undefined') {
-    console.warn(
-      `[embedding-retriever] index was built with provider=${providerName} model=${model}, ` +
-      `but the corresponding API key is not set. Embeddings will be unavailable; falling back ` +
-      `to keyword-only retrieval. Set the matching API key, or rebuild the index with the ` +
-      `available provider via \`npm run build:embeddings\`.`
-    );
-  }
-  return fn;
-}
-
-/**
- * True when both the index AND the matching provider's API key are available.
- * Callers use this to decide whether to include embedding scores in the blend.
- */
-export async function available() {
-  if (_available !== null) return _available;
-  const idx = await _loadIndex();
-  if (!idx || !idx.patterns?.length) { _available = false; return false; }
-  _embedFn = _resolveEmbed(idx.provider, idx.model);
-  _available = !!_embedFn;
-  return _available;
-}
-
-/**
- * Embed a query string and return a { patternName → cosineScore } map.
- * Returns an empty Map when unavailable (no index or no API key).
- *
- * @param {string} query — the user's intent (steelman or raw)
- * @returns {Promise<Map<string, number>>}
- */
-export async function scoreAll(query) {
-  if (!query || typeof query !== 'string') return new Map();
-  if (!(await available())) return new Map();
-
-  let qVec;
-  try {
-    const [v] = await _embedFn([query]);
-    qVec = v;
-  } catch (e) {
-    // Don't let a runtime embedding failure nuke retrieval — fall back cleanly.
-    if (typeof console !== 'undefined') console.warn('[embedding-retriever]', e.message);
-    return new Map();
-  }
-
-  const out = new Map();
-  for (const [name, vec] of _indexByName) {
-    out.set(name, cosine(qVec, vec));
-  }
-  return out;
-}
-
-/** Number of patterns in the index. Useful for logging/diagnostics. */
-export async function size() {
-  const idx = await _loadIndex();
-  return idx?.patterns?.length || 0;
-}
-
-/** Diagnostics: the provider/model the index was built with. */
-export async function providerInfo() {
-  const idx = await _loadIndex();
-  return idx ? { provider: idx.provider, model: idx.model, dims: idx.dims } : null;
-}