implemenation on the bot

services/ai/collections.js

@@ -0,0 +1,77 @@
+import { ChromaClient } from 'chromadb';
+import { MODEL_ID, embed, embedMany } from './embed.js';
+
+// ==============================|| Chroma collections ||============================== //
+//
+// Two collections, deliberately separate:
+//
+//   intent_examples — ~20 phrasings per intent. Answers "which question is
+//                     this?", never "what is the number?".
+//   console_docs    — chunked markdown. Genuine document Q&A.
+//
+// NO OPERATIONAL DATA IS EMBEDDED. No bookings, riders or customers ever enter
+// the vector store. A vector store is a snapshot and this data changes by the
+// minute; every figure still comes from a live API call through the existing
+// typed functions. That is the whole reason this design is safe to build.
+
+export const INTENTS_COLLECTION = 'intent_examples';
+export const DOCS_COLLECTION = 'console_docs';
+
+const client = new ChromaClient({ path: process.env.CHROMA_URL || 'http://localhost:8000' });
+
+// Chroma would otherwise call its own default embedding function (which
+// downloads a different model). We embed ourselves so queries and documents
+// are guaranteed to come from the same model.
+const noopEmbeddingFunction = { generate: async (texts) => embedMany(texts) };
+
+export const getCollection = async (name) =>
+  client.getOrCreateCollection({
+    name,
+    embeddingFunction: noopEmbeddingFunction,
+    metadata: {
+      // Stamped so a model swap is detectable rather than silently degrading
+      // every score. Re-seed after changing it.
+      'hnsw:space': 'cosine',
+      embeddingModel: MODEL_ID
+    }
+  });
+
+export const resetCollection = async (name) => {
+  try {
+    await client.deleteCollection({ name });
+  } catch {
+    // Not present yet — nothing to delete.
+  }
+  return getCollection(name);
+};
+
+// Chroma returns cosine DISTANCE (0 = identical). Similarity is the useful
+// direction for a human-facing confidence, so convert once, here, rather than
+// leaving every call site to remember which way round it is.
+const toSimilarity = (distance) => 1 - Number(distance);
+
+export const query = async (name, text, topK = 5) => {
+  const collection = await getCollection(name);
+  const vector = await embed(text);
+  const res = await collection.query({ queryEmbeddings: [vector], nResults: topK });
+
+  const ids = res.ids?.[0] || [];
+  return ids.map((id, i) => ({
+    id,
+    document: res.documents?.[0]?.[i] || '',
+    metadata: res.metadatas?.[0]?.[i] || {},
+    score: toSimilarity(res.distances?.[0]?.[i] ?? 1)
+  }));
+};
+
+export const health = async () => {
+  await client.heartbeat();
+  const names = (await client.listCollections()).map((c) => c.name ?? c);
+  const counts = {};
+  for (const name of [INTENTS_COLLECTION, DOCS_COLLECTION]) {
+    if (!names.includes(name)) continue;
+    // eslint-disable-next-line no-await-in-loop
+    counts[name] = await (await getCollection(name)).count();
+  }
+  return { chroma: 'up', collections: counts };
+};