fix: prevent list index out of range in chat streaming (#10238)

### What problem does this PR solve?
issue:
[Bug]: ERROR: list index out of range #10188
change:
fix a potential list index out of range error in chat response parsing
by adding explicit checks for empty choices.

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)

agent/component/string_transform.py

@@ -90,9 +90,7 @@ class StringTransform(Message, ABC):
         for k,v in kwargs.items():
             if not v:
                 v = ""
-            k = re.sub(r'\\m', 'm', k)
-            v = re.sub(r'\\m', 'm', v)
-            script = re.sub(k, v, script)
+            script = re.sub(k, lambda match: v, script)
 
         self.set_output("result", script)
 

api/apps/sdk/session.py

@@ -182,7 +182,7 @@ def chat_completion_openai_like(tenant_id, chat_id):
     stream = True
     reference = True
 
-    completion = client.responses.create(
+    completion = client.chat.completions.create(
         model=model,
         messages=[
             {"role": "system", "content": "You are a helpful assistant."},

rag/llm/chat_model.py

@@ -144,9 +144,9 @@ class Base(ABC):
         if self.model_name.lower().find("qwen3") >= 0:
             kwargs["extra_body"] = {"enable_thinking": False}
         
-        response = self.client.responses.create(model=self.model_name, messages=history, **gen_conf, **kwargs)
+        response = self.client.chat.completions.create(model=self.model_name, messages=history, **gen_conf, **kwargs)
 
-        if any([not response.choices, not response.choices[0].message, not response.choices[0].message.content]):
+        if (not response.choices or not response.choices[0].message or not response.choices[0].message.content):
             return "", 0
         ans = response.choices[0].message.content.strip()
         if response.choices[0].finish_reason == "length":
@@ -158,9 +158,9 @@ class Base(ABC):
         reasoning_start = False
         
         if kwargs.get("stop") or "stop" in gen_conf:
-            response = self.client.responses.create(model=self.model_name, messages=history, stream=True, **gen_conf, stop=kwargs.get("stop"))
+            response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, **gen_conf, stop=kwargs.get("stop"))
         else:
-            response = self.client.responses.create(model=self.model_name, messages=history, stream=True, **gen_conf)
+            response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, **gen_conf)
         
         for resp in response:
             if not resp.choices:
@@ -193,21 +193,30 @@ class Base(ABC):
             return ans + LENGTH_NOTIFICATION_CN
         return ans + LENGTH_NOTIFICATION_EN
 
-    def _exceptions(self, e, attempt):
+    @property
+    def _retryable_errors(self) -> set[str]:
+        return {
+            LLMErrorCode.ERROR_RATE_LIMIT,
+            LLMErrorCode.ERROR_SERVER,
+        }
+
+    def _should_retry(self, error_code: str) -> bool:
+        return error_code in self._retryable_errors
+
+    def _exceptions(self, e, attempt) -> str | None:
         logging.exception("OpenAI chat_with_tools")
         # Classify the error
         error_code = self._classify_error(e)
         if attempt == self.max_retries:
             error_code = LLMErrorCode.ERROR_MAX_RETRIES
 
-        # Check if it's a rate limit error or server error and not the last attempt
-        should_retry = error_code == LLMErrorCode.ERROR_RATE_LIMIT or error_code == LLMErrorCode.ERROR_SERVER
-        if not should_retry:
-            return f"{ERROR_PREFIX}: {error_code} - {str(e)}"
+        if self._should_retry(error_code):
+            delay = self._get_delay()
+            logging.warning(f"Error: {error_code}. Retrying in {delay:.2f} seconds... (Attempt {attempt + 1}/{self.max_retries})")
+            time.sleep(delay)
+            return None
 
-        delay = self._get_delay()
-        logging.warning(f"Error: {error_code}. Retrying in {delay:.2f} seconds... (Attempt {attempt + 1}/{self.max_retries})")
-        time.sleep(delay)
+        return f"{ERROR_PREFIX}: {error_code} - {str(e)}"
 
     def _verbose_tool_use(self, name, args, res):
         return "<tool_call>" + json.dumps({"name": name, "args": args, "result": res}, ensure_ascii=False, indent=2) + "</tool_call>"
@@ -257,7 +266,7 @@ class Base(ABC):
             try:
                 for _ in range(self.max_rounds + 1):
                     logging.info(f"{self.tools=}")
-                    response = self.client.responses.create(model=self.model_name, messages=history, tools=self.tools, tool_choice="auto", **gen_conf)
+                    response = self.client.chat.completions.create(model=self.model_name, messages=history, tools=self.tools, tool_choice="auto", **gen_conf)
                     tk_count += self.total_token_count(response)
                     if any([not response.choices, not response.choices[0].message]):
                         raise Exception(f"500 response structure error. Response: {response}")
@@ -342,7 +351,7 @@ class Base(ABC):
                 for _ in range(self.max_rounds + 1):
                     reasoning_start = False
                     logging.info(f"{tools=}")
-                    response = self.client.responses.create(model=self.model_name, messages=history, stream=True, tools=tools, tool_choice="auto", **gen_conf)
+                    response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, tools=tools, tool_choice="auto", **gen_conf)
                     final_tool_calls = {}
                     answer = ""
                     for resp in response:
@@ -405,7 +414,7 @@ class Base(ABC):
 
                 logging.warning(f"Exceed max rounds: {self.max_rounds}")
                 history.append({"role": "user", "content": f"Exceed max rounds: {self.max_rounds}"})
-                response = self.client.responses.create(model=self.model_name, messages=history, stream=True, **gen_conf)
+                response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, **gen_conf)
                 for resp in response:
                     if any([not resp.choices, not resp.choices[0].delta, not hasattr(resp.choices[0].delta, "content")]):
                         raise Exception("500 response structure error.")
@@ -536,6 +545,14 @@ class AzureChat(Base):
         self.client = AzureOpenAI(api_key=api_key, azure_endpoint=base_url, api_version=api_version)
         self.model_name = model_name
 
+    @property
+    def _retryable_errors(self) -> set[str]:
+        return {
+            LLMErrorCode.ERROR_RATE_LIMIT,
+            LLMErrorCode.ERROR_SERVER,
+            LLMErrorCode.ERROR_QUOTA,
+        }
+
 
 class BaiChuanChat(Base):
     _FACTORY_NAME = "BaiChuan"
@@ -559,7 +576,7 @@ class BaiChuanChat(Base):
         }
 
     def _chat(self, history, gen_conf={}, **kwargs):
-        response = self.client.responses.create(
+        response = self.client.chat.completions.create(
             model=self.model_name,
             messages=history,
             extra_body={"tools": [{"type": "web_search", "web_search": {"enable": True, "search_mode": "performance_first"}}]},
@@ -581,7 +598,7 @@ class BaiChuanChat(Base):
         ans = ""
         total_tokens = 0
         try:
-            response = self.client.responses.create(
+            response = self.client.chat.completions.create(
                 model=self.model_name,
                 messages=history,
                 extra_body={"tools": [{"type": "web_search", "web_search": {"enable": True, "search_mode": "performance_first"}}]},
@@ -651,7 +668,7 @@ class ZhipuChat(Base):
         tk_count = 0
         try:
             logging.info(json.dumps(history, ensure_ascii=False, indent=2))
-            response = self.client.responses.create(model=self.model_name, messages=history, stream=True, **gen_conf)
+            response = self.client.chat.completions.create(model=self.model_name, messages=history, stream=True, **gen_conf)
             for resp in response:
                 if not resp.choices[0].delta.content:
                     continue
@@ -1364,7 +1381,7 @@ class LiteLLMBase(ABC):
             drop_params=True,
             timeout=self.timeout,
         )
-        # response = self.client.responses.create(model=self.model_name, messages=history, **gen_conf, **kwargs)
+        # response = self.client.chat.completions.create(model=self.model_name, messages=history, **gen_conf, **kwargs)
 
         if any([not response.choices, not response.choices[0].message, not response.choices[0].message.content]):
             return "", 0
@@ -1424,21 +1441,30 @@ class LiteLLMBase(ABC):
             return ans + LENGTH_NOTIFICATION_CN
         return ans + LENGTH_NOTIFICATION_EN
 
-    def _exceptions(self, e, attempt):
+    @property
+    def _retryable_errors(self) -> set[str]:
+        return {
+            LLMErrorCode.ERROR_RATE_LIMIT,
+            LLMErrorCode.ERROR_SERVER,
+        }
+
+    def _should_retry(self, error_code: str) -> bool:
+        return error_code in self._retryable_errors
+
+    def _exceptions(self, e, attempt) -> str | None:
         logging.exception("OpenAI chat_with_tools")
         # Classify the error
         error_code = self._classify_error(e)
         if attempt == self.max_retries:
             error_code = LLMErrorCode.ERROR_MAX_RETRIES
 
-        # Check if it's a rate limit error or server error and not the last attempt
-        should_retry = error_code == LLMErrorCode.ERROR_RATE_LIMIT or error_code == LLMErrorCode.ERROR_SERVER
-        if not should_retry:
-            return f"{ERROR_PREFIX}: {error_code} - {str(e)}"
+        if self._should_retry(error_code):
+            delay = self._get_delay()
+            logging.warning(f"Error: {error_code}. Retrying in {delay:.2f} seconds... (Attempt {attempt + 1}/{self.max_retries})")
+            time.sleep(delay)
+            return None
 
-        delay = self._get_delay()
-        logging.warning(f"Error: {error_code}. Retrying in {delay:.2f} seconds... (Attempt {attempt + 1}/{self.max_retries})")
-        time.sleep(delay)
+        return f"{ERROR_PREFIX}: {error_code} - {str(e)}"
 
     def _verbose_tool_use(self, name, args, res):
         return "<tool_call>" + json.dumps({"name": name, "args": args, "result": res}, ensure_ascii=False, indent=2) + "</tool_call>"

rag/llm/cv_model.py

@@ -75,7 +75,7 @@ class Base(ABC):
 
     def chat(self, system, history, gen_conf, images=[], **kwargs):
         try:
-            response = self.client.responses.create(
+            response = self.client.chat.completions.create(
                 model=self.model_name,
                 messages=self._form_history(system, history, images)
             )
@@ -87,7 +87,7 @@ class Base(ABC):
         ans = ""
         tk_count = 0
         try:
-            response = self.client.responses.create(
+            response = self.client.chat.completions.create(
                 model=self.model_name,
                 messages=self._form_history(system, history, images),
                 stream=True
@@ -174,8 +174,7 @@ class GptV4(Base):
 
     def describe(self, image):
         b64 = self.image2base64(image)
-        # Check if this is a GPT-5 model and use responses.create API
-        res = self.client.responses.create(
+        res = self.client.chat.completions.create(
             model=self.model_name,
             messages=self.prompt(b64),
         )
@@ -183,7 +182,7 @@ class GptV4(Base):
 
     def describe_with_prompt(self, image, prompt=None):
         b64 = self.image2base64(image)
-        res = self.client.responses.create(
+        res = self.client.chat.completions.create(
             model=self.model_name,
             messages=self.vision_llm_prompt(b64, prompt),
         )