fix(api): enforce global ollama endpoint order

apps/api/src/services/embedding_service.py

@@ -22,7 +22,7 @@ import httpx
 import structlog
 
 from src.services.model_registry import get_model as _get_model
-from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
+from src.services.ollama_endpoint_resolver import resolve_ollama_order
 
 logger = structlog.get_logger(__name__)
 
@@ -98,7 +98,15 @@ class OllamaEmbeddingService:
         P1 修復 (2026-03-29): 維度配置化，支援更多模型
         """
         self._model = model
-        self._ollama_url = ollama_url or resolve_ollama_endpoint("embedding")
+        if ollama_url:
+            self._ollama_endpoints = ((ollama_url, "custom"),)
+        else:
+            self._ollama_endpoints = tuple(
+                (endpoint.url, endpoint.provider_name)
+                for endpoint in resolve_ollama_order("embedding")
+                if endpoint.url
+            )
+        self._ollama_url = self._ollama_endpoints[0][0] if self._ollama_endpoints else ""
         self._timeout = timeout
         self._default_dimension = default_dimension or self.MODEL_DIMENSIONS.get(
             model, self.DEFAULT_DIMENSION
@@ -142,43 +150,65 @@ class OllamaEmbeddingService:
         """
         client = await self._get_client()
 
-        try:
-            response = await client.post(
-                f"{self._ollama_url}/api/embeddings",
-                json={
-                    "model": self._model,
-                    "prompt": text,
-                },
-            )
-            response.raise_for_status()
+        last_error: Exception | None = None
+        for endpoint_url, provider_name in self._ollama_endpoints:
+            try:
+                response = await client.post(
+                    f"{endpoint_url}/api/embeddings",
+                    json={
+                        "model": self._model,
+                        "prompt": text,
+                    },
+                )
+                response.raise_for_status()
 
-            data = response.json()
-            embedding = data.get("embedding", [])
+                data = response.json()
+                embedding = data.get("embedding", [])
 
-            # 更新維度快取
-            if self._dimension is None and embedding:
-                self._dimension = len(embedding)
-                logger.info(
-                    "embedding_dimension_detected",
+                # 更新維度快取
+                if self._dimension is None and embedding:
+                    self._dimension = len(embedding)
+                    logger.info(
+                        "embedding_dimension_detected",
+                        model=self._model,
+                        dimension=self._dimension,
+                        provider=provider_name,
+                    )
+
+                return embedding
+
+            except httpx.TimeoutException as e:
+                last_error = e
+                logger.error(
+                    "embedding_timeout",
                     model=self._model,
-                    dimension=self._dimension,
+                    text_len=len(text),
+                    provider=provider_name,
+                )
+            except httpx.HTTPStatusError as e:
+                last_error = e
+                logger.error(
+                    "embedding_http_error",
+                    status=e.response.status_code,
+                    model=self._model,
+                    provider=provider_name,
+                )
+            except Exception as e:
+                last_error = e
+                logger.error(
+                    "embedding_error",
+                    error=str(e),
+                    model=self._model,
+                    provider=provider_name,
                 )
 
-            return embedding
-
-        except httpx.TimeoutException as e:
-            logger.error("embedding_timeout", model=self._model, text_len=len(text))
-            raise EmbeddingError(f"Embedding timeout after {self._timeout}s") from e
-        except httpx.HTTPStatusError as e:
-            logger.error(
-                "embedding_http_error",
-                status=e.response.status_code,
-                model=self._model,
-            )
-            raise EmbeddingError(f"Ollama API error: {e.response.status_code}") from e
-        except Exception as e:
-            logger.error("embedding_error", error=str(e), model=self._model)
-            raise EmbeddingError(f"Embedding failed: {e}") from e
+        if isinstance(last_error, httpx.TimeoutException):
+            raise EmbeddingError(f"Embedding timeout after {self._timeout}s") from last_error
+        if isinstance(last_error, httpx.HTTPStatusError):
+            raise EmbeddingError(
+                f"Ollama API error: {last_error.response.status_code}"
+            ) from last_error
+        raise EmbeddingError("Embedding failed on all Ollama endpoints") from last_error
 
     async def embed_batch(
         self,