fix: xinference chat completion error (#952)

api/core/third_party/langchain/llms/xinference_llm.py

                 return combined_text_output
             else:
                 completion = model.chat(prompt=prompt, generate_config=generate_config)
-                return completion["choices"][0]["text"]
+                return completion["choices"][0]["message"]["content"]
         elif isinstance(model, RESTfulGenerateModelHandle):
             generate_config: "LlamaCppGenerateConfig" = kwargs.get("generate_config", {})
 
                 completion = combined_text_output
             else:
                 completion = model.chat(prompt=prompt, generate_config=generate_config)
-                completion = completion["choices"][0]["text"]
+                completion = completion["choices"][0]["message"]["content"]
 
             if stop is not None:
                 completion = enforce_stop_tokens(completion, stop)