fix(ai): 遷移 Ollama embedding 至 api embed

services/ollama_service.py

@@ -21,6 +21,7 @@ OLLAMA_HOST = os.getenv('OLLAMA_HOST', 'http://192.168.0.111:11434')
 DEFAULT_MODEL = os.getenv('OLLAMA_MODEL', 'llama3.1:8b')  # 較快速的模型
 TIMEOUT = int(os.getenv('OLLAMA_TIMEOUT', '120'))  # 秒 - 2 分鐘
 COPY_TIMEOUT = int(os.getenv('OLLAMA_COPY_TIMEOUT', '180'))  # 文案生成專用超時 - 3 分鐘
+EMBED_TIMEOUT = int(os.getenv('OLLAMA_EMBED_TIMEOUT', os.getenv('EMBEDDING_TIMEOUT', '45')))
 
 
 @dataclass
@@ -505,8 +506,25 @@ class OllamaService:
 
         return self.generate(prompt, system_prompt=system_prompt, temperature=0.5, timeout=120)
 
+    @staticmethod
+    def _extract_embedding(payload: Dict[str, Any]) -> List[float]:
+        """Normalize Ollama /api/embed and legacy /api/embeddings response shapes."""
+        embeddings = payload.get("embeddings")
+        if isinstance(embeddings, list) and embeddings:
+            first = embeddings[0]
+            if isinstance(first, list):
+                return first
+            if all(isinstance(value, (int, float)) for value in embeddings):
+                return embeddings
+
+        embedding = payload.get("embedding")
+        if isinstance(embedding, list):
+            return embedding
+
+        return []
+
     def generate_embedding(self, text: str, model: str = "bge-m3:latest",
-                           host: str = None) -> List[float]:
+                           host: str = None, timeout: int = None) -> List[float]:
         """
         [ADR-007, Step 3] 呼叫 Ollama API 將文字轉換為向量 Embedding
 
@@ -516,22 +534,38 @@ class OllamaService:
         避免 self.host 若指向公開 ollama.wooo.work 時回 401。
         可透過 host 參數 override。
         """
-        target_host = host or os.getenv("EMBEDDING_HOST", "http://192.168.0.111:11434")
+        target_host = (host or os.getenv("EMBEDDING_HOST", "http://192.168.0.111:11434")).rstrip("/")
+        request_timeout = timeout or EMBED_TIMEOUT
         try:
-            payload = {"model": model, "prompt": text}
+            payload = {"model": model, "input": text}
             response = requests.post(
-                f"{target_host}/api/embeddings",
+                f"{target_host}/api/embed",
                 json=payload,
-                timeout=60,
+                timeout=request_timeout,
             )
             if response.status_code == 200:
-                data = response.json()
-                return data.get("embedding", [])
-            else:
+                vec = self._extract_embedding(response.json())
+                if vec:
+                    return vec
+                logger.warning(f"Ollama Embed Empty Response @ {target_host}/api/embed")
+            elif response.status_code not in (404, 405):
                 logger.error(
-                    f"Ollama Embed Error HTTP {response.status_code} @ {target_host}: {response.text[:200]}"
+                    f"Ollama Embed Error HTTP {response.status_code} @ {target_host}/api/embed: {response.text[:200]}"
                 )
                 return []
+
+            # V-Fix: 舊 Ollama 相容；/api/embeddings 已 deprecated，但仍是部分舊節點唯一可用路徑。
+            legacy_response = requests.post(
+                f"{target_host}/api/embeddings",
+                json={"model": model, "prompt": text},
+                timeout=request_timeout,
+            )
+            if legacy_response.status_code == 200:
+                return self._extract_embedding(legacy_response.json())
+            logger.error(
+                f"Ollama Embed Error HTTP {legacy_response.status_code} @ {target_host}/api/embeddings: {legacy_response.text[:200]}"
+            )
+            return []
         except Exception as e:
             logger.error(f"Ollama Embed Exception @ {target_host}: {e}")
             return []