backends/advanced/src/advanced_omi_backend/memory/providers/llm_providers.py (2)

242-250: Ollama connection check should use explicit timeout and avoid env-coupling inside provider

• Pass a finite timeout to httpx to avoid hangs.
• Avoid relying on global LLM_PROVIDER; base behavior should be driven by instance config.

Apply:

-            if os.getenv("LLM_PROVIDER", "openai").lower() == "ollama":
-                import httpx
-                async with httpx.AsyncClient() as client:
-                    # For Ollama, test connection by hitting the /v1/models endpoint
-                    response = await client.get(f"{self.base_url}/models")
+            if "ollama" in (os.getenv("LLM_PROVIDER", "openai").lower()):
+                import httpx
+                async with httpx.AsyncClient(timeout=httpx.Timeout(5.0)) as client:
+                    # For Ollama's OpenAI‑compatible API:
+                    response = await client.get(f"{self.base_url.rstrip('/')}/models")
                     response.raise_for_status()
                 return True

Please confirm your Ollama is using the OpenAI‑compatible base_url (ending with /v1). If native API is desired here, switch to GET {ollama_root}/api/tags.

---

186-201: response_format='json_object' may not be supported by all providers/models

Some OpenAI‑compatible backends (incl. Ollama) ignore or reject response_format. Consider gating this by provider/model or falling back gracefully on error.

Add a quick run against your configured Ollama model to verify JSON‑object support in chat.completions.

Also applies to: 193-195

backends/advanced/src/advanced_omi_backend/memory/config.py (1)

151-154: Provider allowlist duplicated as literals

Use the Enum values to avoid drift.

-        if llm_provider not in ["openai", "ollama"]:
+        if llm_provider not in {p.value for p in LLMProvider}:

backends/advanced/src/advanced_omi_backend/memory/memory_service.py (1)

72-75: Provider class naming vs behavior

You instantiate OpenAIProvider for both OPENAI and OLLAMA. Works with OpenAI‑compatible endpoints, but is confusing.

Rename to OpenAICompatibleProvider or route to a thin OllamaProvider alias for clarity.

backends/advanced/webui/src/pages/System.tsx (1)

7-18: Remove unnecessary any casts by typing service entry

You already declared optional fields; casting to any is avoidable and hides errors.

Define a ServiceStatus interface and use it:

-interface HealthData {
+interface ServiceStatus {
+  healthy: boolean
+  message?: string
+  status?: string
+  base_url?: string
+  model?: string
+  embedder_model?: string
+  embedder_status?: string
+  provider?: string
+  worker_count?: number
+  active_workers?: number
+  idle_workers?: number
+}
+interface HealthData {
   status: 'healthy' | 'partial' | 'unhealthy'
-  services: Record<string, {
-    healthy: boolean
-    message?: string
-    status?: string;
-    base_url?: string;
-    model?: string;
-    embedder_model?: string;
-    embedder_status?: string;
-    provider?: string;
-  }>
+  services: Record<string, ServiceStatus>
 }

Then replace (status as any) usages with typed status.

Also applies to: 275-290

backends/advanced/src/advanced_omi_backend/routers/modules/health_routes.py (2)

195-206: AudioAI health: consider 'partial' when embedder missing

Treating “Embedder Model Not Configured” as healthy can mask degraded capability.

• Keep healthy boolean, but set status message to reflect partial state when embedder is missing/unhealthy.
• Optionally map overall system status to 'degraded' if LLM is OK but embedder is not, when MEMORY_PROVIDER=friend_lite.

Confirm whether embedder is required for your current flows; adjust semantics accordingly.

Also applies to: 207-216

---

90-115: Health response exposes internal URIs

mongodb_uri and llm_base_url leak infra details. If /health is ever unauthenticated, redact.

Mask hosts or omit credentials/components in the JSON payload for non‑admin callers.

backends/advanced/init.py (1)

371-399: .env CORS origins differ from docker-compose CORS_ORIGINS

Compose sets a broader CORS list (3010, https). The generated .env shows 3000/5173 only and may confuse admins.

Align defaults or add a note that docker-compose overrides CORS via environment.

backends/advanced/src/advanced_omi_backend/llm_client.py (2)

110-156: Also verify Ollama LLM model availability

You check server and embedder; also check the default LLM model exists to avoid false “Connected”.

Apply:

                 # Check embedder model availability
                 embedder_model_name = os.getenv("OLLAMA_EMBEDDER_MODEL")
                 if embedder_model_name:
                     try:
                         # Use /api/show to check if model exists
                         async with session.post(f"{ollama_health_url}/api/show", json={"name": embedder_model_name}, timeout=aiohttp.ClientTimeout(total=5)) as embedder_response:
                             if embedder_response.status == 200:
                                 response_data["embedder_status"] = "✅ Available"
                             else:
                                 response_data["embedder_status"] = "⚠️ Embedder Model Unhealthy"
                     except aiohttp.ClientError:
                         response_data["embedder_status"] = "❌ Embedder Model Connection Failed"
                     except asyncio.TimeoutError:
                         response_data["embedder_status"] = "❌ Embedder Model Timeout"
                 else:
                     response_data["embedder_status"] = "⚠️ Embedder Model Not Configured"
+                # Verify chat model presence too
+                try:
+                    async with session.post(f"{ollama_health_url}/api/show", json={"name": self.model}, timeout=aiohttp.ClientTimeout(total=5)) as model_resp:
+                        if response_data["status"].startswith("✅") and model_resp.status != 200:
+                            response_data["status"] = "⚠️ Connected, model not pulled"
+                except (aiohttp.ClientError, asyncio.TimeoutError):
+                    response_data["status"] = "❌ Ollama Model Check Failed"

---

157-172: OpenAI path returns “Connected” without a network check

This can mask invalid keys/base URLs. Consider a lightweight call (e.g., list models) with short timeout, or coordinate with MemoryService connection test for consistency.

If you prefer to keep this cheap, reflect config-only status explicitly (e.g., “Configured (not verified)”).

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro