implemenation on the bot

services/ai/embed.js

@@ -0,0 +1,66 @@
+import { pipeline, env } from '@xenova/transformers';
+
+// ==============================|| Embeddings ||============================== //
+//
+// all-MiniLM-L6-v2 running IN PROCESS. No API key, no network call per query,
+// no per-call cost.
+//
+// That is not an incidental choice. assistant/CLAUDE.md §2 blocked LLM work on
+// exactly one ground: a hosted model needs a secret key and a static CRA build
+// has nowhere to put one. A local embedding model has no key, so this plan is
+// outside that blocker. Swapping to a hosted embedding model (text-embedding-3
+// -small et al) re-opens §2 and needs its own decision — do not do it quietly.
+
+export const MODEL_ID = 'Xenova/all-MiniLM-L6-v2';
+export const DIMENSIONS = 384;
+
+// Weights are cached on disk (mounted in docker-compose) so a container
+// restart doesn't re-download 25MB.
+env.cacheDir = process.env.TRANSFORMERS_CACHE || './.cache';
+// Nothing here should reach the network except the one-time model fetch.
+env.allowRemoteModels = true;
+
+let extractor = null;
+let loading = null;
+
+// Loaded once, lazily, and shared. Concurrent callers await the same promise
+// rather than each triggering their own model load.
+const getExtractor = async () => {
+  if (extractor) return extractor;
+  if (!loading) {
+    loading = pipeline('feature-extraction', MODEL_ID).then((p) => {
+      extractor = p;
+      return p;
+    });
+  }
+  return loading;
+};
+
+export const warmUp = async () => {
+  await getExtractor();
+  return { model: MODEL_ID, dimensions: DIMENSIONS };
+};
+
+// Mean-pooled + L2-normalised sentence embedding.
+//
+// Normalisation matters: Chroma's cosine space assumes unit vectors, and
+// queries must be embedded exactly the same way documents were. A mismatch
+// doesn't error — it silently degrades every score, which is the worst kind of
+// bug to chase. `seed` writes the model id into collection metadata so a
+// mismatch is at least detectable.
+export const embed = async (text) => {
+  const pipe = await getExtractor();
+  const output = await pipe(String(text || '').trim(), { pooling: 'mean', normalize: true });
+  return Array.from(output.data);
+};
+
+export const embedMany = async (texts) => {
+  const out = [];
+  for (const t of texts) {
+    // Sequential on purpose: batching MiniLM in-process gives no meaningful
+    // speed-up at this corpus size (~700 vectors) and makes memory spikier.
+    // eslint-disable-next-line no-await-in-loop
+    out.push(await embed(t));
+  }
+  return out;
+};