Replace local SLM embeddings with scope.models.embed()

.gitignore

@@ -1,5 +1,4 @@
 node_modules/
-models/
 .env
 CONFIG.env
 *.env

config.yaml

@@ -6,6 +6,10 @@ loadEnv:
 
 rest: true
 
+# Captures the Scope object so resources can call scope.models.embed().
+# See lib/modelCapture.js.
+extensionModule: 'lib/modelCapture.js'
+
 graphqlSchema:
   files: 'schemas/*.graphql'
 

lib/embeddings.js

@@ -1,14 +1,14 @@
-import { init, embed as llamaEmbed } from 'harper-fabric-embeddings'
-import { resolve } from 'path'
-import { fileURLToPath } from 'url'
-
-const __dirname = fileURLToPath(new URL('.', import.meta.url))
-const modelPath = resolve(__dirname, '../models/bge-small-en-v1.5-q4_k_m.gguf')
-
-// Model is pre-downloaded by the predev/prestart npm hook (scripts/download-model.js)
-const initPromise = init({ modelPath })
+// Thin wrapper around `scope.models.embed()` (harper#510). The host's
+// configured backend (Ollama on Fabric GPU hosts, or any backend configured
+// via the `models:` block in harperdb-config.yaml / env vars) handles the
+// actual inference. Returns a plain Array<number> for compatibility with
+// Harper's HNSW vector index storage.
 
 export async function embed(text) {
-  await initPromise
-  return llamaEmbed(text)
+	const scope = globalThis.harperScope;
+	if (!scope) {
+		throw new Error('Harper scope not yet captured — modelCapture plugin must run before first embed call');
+	}
+	const [vector] = await scope.models.embed(text);
+	return Array.from(vector);
 }

lib/modelCapture.js

@@ -0,0 +1,11 @@
+// Plugin entry — captures the Scope object so `resources/*.js` can call
+// `scope.models.embed()` against the host's configured embedding backend.
+//
+// Harper's `scope` is passed to plugins via `handleApplication(scope)` but
+// isn't exposed as a global to Resource classes. This tiny plugin stashes the
+// Scope on `globalThis.harperScope` at app boot, then `lib/embeddings.js`
+// reads it from there.
+
+export function handleApplication(scope) {
+	globalThis.harperScope = scope;
+}