@rizom/ops 0.2.0-alpha.679 → 0.2.0-alpha.680

package/dist/load-registry.d.ts

@@ -87,6 +87,7 @@ export interface ResolvedUserIdentity {
     profileKind?: string | undefined;
     anchorKind?: AnchorKind | undefined;
     embeddingEnabled?: boolean | undefined;
+    embeddingModel?: string | undefined;
     plugins?: UserConfig["plugins"];
     spaces?: string[] | undefined;
     topicExtractionEnabled?: boolean | undefined;

package/dist/schema.d.ts

@@ -141,6 +141,7 @@ export declare const userSchema: z.ZodObject<{
     plugins: z.ZodOptional<typeof userPluginsSchema>;
     spaces: z.ZodOptional<z.ZodArray<z.ZodString>>;
     embeddingEnabled: z.ZodOptional<z.ZodBoolean>;
+    embeddingModel: z.ZodOptional<z.ZodString>;
     topicExtractionEnabled: z.ZodOptional<z.ZodBoolean>;
     skillDerivationEnabled: z.ZodOptional<z.ZodBoolean>;
     swotDerivationEnabled: z.ZodOptional<z.ZodBoolean>;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@rizom/ops",
-  "version": "0.2.0-alpha.679",
+  "version": "0.2.0-alpha.680",
   "description": "Operator CLI for managing private brain fleet registry repos",
   "keywords": [
     "brains",

package/templates/rover-pilot/.github/workflows/deploy.yml

@@ -207,6 +207,14 @@ jobs:
           SERVICE_NAME: rover
         run: bun deploy/scripts/create-predeploy-backup.ts
 
+      - name: Pull local models
+        env:
+          SERVER_IP: ${{ steps.provision.outputs.server_ip }}
+          LOCAL_MODELS: ${{ steps.user_config.outputs.local_models }}
+        run: |
+          SSH_USER="$(ruby -e 'require "yaml"; config = YAML.load_file("deploy/kamal/deploy.yml") || {}; puts(config.dig("ssh", "user") || "root")')" \
+            bun deploy/scripts/pull-local-models.ts
+
       - name: Release operator-confirmed stale deploy lock
         if: ${{ inputs.release_stale_lock }}
         env:
@@ -230,6 +238,7 @@ jobs:
           WWW_DOMAIN: ${{ steps.user_config.outputs.www_domain }}
           PREVIEW_DOMAIN: ${{ steps.user_config.outputs.preview_domain }}
           BRAIN_YAML_PATH: ${{ steps.user_config.outputs.brain_yaml_path }}
+          LOCAL_MODELS: ${{ steps.user_config.outputs.local_models }}
         run: kamal setup --skip-push -c deploy/kamal/deploy.yml
 
       - name: Install container health watchdog

package/templates/rover-pilot/deploy/scripts/pull-local-models.ts

@@ -0,0 +1,69 @@
+import { requireEnv } from "./helpers";
+
+/** The model runtime the `models` accessory runs; kamal-deploy.yml pins the same image. */
+export const MODEL_RUNTIME_IMAGE = "ollama/ollama:0.40.11";
+const MODELS_VOLUME = "/opt/brain-models:/root/.ollama";
+const MODEL_TAG = /^[A-Za-z0-9][A-Za-z0-9._:/-]*$/;
+
+/**
+ * The local models the deploy names (`LOCAL_MODELS`, the `local:` models of
+ * brain.yaml without their prefix), once each.
+ */
+export function parseLocalModels(value: string | undefined): string[] {
+  const models = (value ?? "").split(/[\s,]+/).filter(Boolean);
+  const invalid = models.find((model) => !MODEL_TAG.test(model));
+  if (invalid !== undefined) {
+    throw new Error(`Invalid local model "${invalid}"`);
+  }
+  return [...new Set(models)];
+}
+
+/**
+ * Pulls each model into the models volume with a one-off runtime container,
+ * so the models are on disk before the brain starts, whether or not the
+ * accessory runs yet.
+ */
+export function buildLocalModelPullScript(models: readonly string[]): string {
+  const pulls = models.map((model) => `ollama pull ${model}`).join(" && ");
+  return `#!/usr/bin/env bash
+set -euo pipefail
+SUDO=""
+if [ "$(id -u)" -ne 0 ]; then
+  SUDO=sudo
+fi
+
+$SUDO mkdir -p /opt/brain-models
+$SUDO docker run --rm --entrypoint /bin/sh -v ${MODELS_VOLUME} ${MODEL_RUNTIME_IMAGE} -c '
+ollama serve >/dev/null 2>&1 &
+for attempt in $(seq 1 30); do
+  ollama list >/dev/null 2>&1 && break
+  sleep 1
+done
+${pulls}
+'
+`;
+}
+
+export async function pullLocalModels(): Promise<void> {
+  const models = parseLocalModels(process.env["LOCAL_MODELS"]);
+  if (models.length === 0) {
+    console.log("No local models in brain.yaml; nothing to pull.");
+    return;
+  }
+  const serverIp = requireEnv("SERVER_IP");
+  const sshUser = requireEnv("SSH_USER").trim();
+  if (!sshUser) throw new Error("Missing SSH_USER");
+  const child = Bun.spawn(["ssh", `${sshUser}@${serverIp}`, "bash -s"], {
+    stdin: new Blob([buildLocalModelPullScript(models)]),
+    stdout: "inherit",
+    stderr: "inherit",
+  });
+  const exitCode = await child.exited;
+  if (exitCode !== 0) {
+    throw new Error(`Pulling local models failed with exit ${exitCode}`);
+  }
+}
+
+if (import.meta.main) {
+  await pullLocalModels();
+}

package/templates/rover-pilot/deploy/scripts/resolve-user-config.ts

@@ -25,6 +25,20 @@ if (!brainDomain) {
   throw new Error(`Missing domain in ${brainYamlPath}`);
 }
 
+// The models the runtime beside the brain serves: brain.yaml's local:
+// models, unprefixed. None means the deploy runs no runtime.
+const localModels = [
+  ...new Set(
+    [
+      ...brainYaml.matchAll(
+        /^(?:embeddingModel|answerModel|visionModel|model):\s*['"]?local:([^'"\s]+)/gm,
+      ),
+    ]
+      .map((match) => match[1] ?? "")
+      .filter(Boolean),
+  ),
+].join(",");
+
 const registry = await loadPilotRegistry(process.cwd());
 const user = registry.users.find((entry) => entry.handle === handle);
 if (!user) {
@@ -69,6 +83,7 @@ const outputs: Record<string, string> = {
   image_repository: `ghcr.io/${repository}`,
   image_tag: imageTag,
   registry_username: repositoryOwner,
+  local_models: localModels,
 };
 
 const required = ["brain_version", "registry_username"];