perf: optimize system concurrency, I/O stability and fix WebSocket disconnects

backend/llm/models.py

@@ -444,6 +444,16 @@ def create_model(
     """
     provider = canonicalize_model_provider(provider)
 
+    # If provider is default OPENAI but model name looks like deepseek, 
+    # check if we should switch to DASHSCOPE.
+    if provider == "OPENAI" and "deepseek" in model_name.lower() and os.getenv("DASHSCOPE_API_KEY"):
+        provider = "DASHSCOPE"
+
+    # Intelligent routing: if it's a DeepSeek model and we have DashScope credentials,
+    # prefer using DashScopeChatModel over OpenAIChatModel.
+    if provider == "DEEPSEEK" and os.getenv("DASHSCOPE_API_KEY"):
+        provider = "DASHSCOPE"
+
     model_class = PROVIDER_MODEL_MAP.get(provider)
     if model_class is None:
         raise ValueError(f"Unsupported provider: {provider}")