fix：解决qwen3.5空回复问题

2026-04-12 15:06:08 +08:00
parent 6db380b10d
commit ff75930466
1 changed files with 113 additions and 0 deletions
--- a/src/llm_models/model_client/openai_client.py
+++ b/src/llm_models/model_client/openai_client.py
@@ -79,6 +79,25 @@ THINK_CONTENT_PATTERN = re.compile(
 )
 """用于解析 `<think>` 推理块的正则表达式。"""

+XML_TOOL_CALL_PATTERN = re.compile(r"<tool_call>\s*(?P<body>.*?)\s*</tool_call>", re.DOTALL | re.IGNORECASE)
+"""用于兜底解析模型以 XML 文本返回的工具调用。
+
+这是一个暂时性兼容方案，专门处理“思维链内容里夹带工具调用”的情况；
+后续如果上游稳定返回标准 tool_calls 字段，这里可能会调整或移除。
+"""
+
+XML_FUNCTION_CALL_PATTERN = re.compile(
+    r"<function=(?P<name>[A-Za-z0-9_.-]+)>\s*(?P<arguments>.*?)\s*</function>",
+    re.DOTALL | re.IGNORECASE,
+)
+"""用于从 XML 风格工具调用块中提取函数名与参数。"""
+
+XML_PARAMETER_PATTERN = re.compile(
+    r"<parameter=(?P<name>[A-Za-z0-9_.-]+)>\s*(?P<value>.*?)\s*</parameter>",
+    re.DOTALL | re.IGNORECASE,
+)
+"""用于从 XML 风格工具调用块中提取参数列表。"""
+
 CHAT_COMPLETIONS_RESERVED_EXTRA_BODY_KEYS = {
    "max_tokens",
    "messages",
@@ -541,6 +560,66 @@ def _extract_reasoning_and_content(
    return None, match.group("content_only").strip() or None


+def _extract_xml_tool_calls(
+    raw_text: str | None,
+    parse_mode: ToolArgumentParseMode,
+    response: Any,
+) -> Tuple[str | None, List[ToolCall] | None]:
+    """从 XML 风格文本中兜底提取工具调用。"""
+    if not isinstance(raw_text, str) or not raw_text.strip():
+        return raw_text, None
+
+    tool_calls: List[ToolCall] = []
+
+    def _coerce_xml_parameter_value(raw_value: str) -> Any:
+        normalized_value = raw_value.strip()
+        if not normalized_value:
+            return ""
+        lowered_value = normalized_value.lower()
+        if lowered_value == "true":
+            return True
+        if lowered_value == "false":
+            return False
+        if lowered_value in {"null", "none"}:
+            return None
+        if normalized_value.startswith(("{", "[")):
+            try:
+                return repair_json(normalized_value, return_objects=True, logging=False)
+            except Exception:
+                return normalized_value
+        return normalized_value
+
+    def _parse_xml_parameters(raw_arguments: str) -> Dict[str, Any] | None:
+        parameters = {
+            match.group("name").strip(): _coerce_xml_parameter_value(match.group("value"))
+            for match in XML_PARAMETER_PATTERN.finditer(raw_arguments)
+        }
+        return parameters or None
+
+    def _replace_tool_call(match: re.Match[str]) -> str:
+        body = match.group("body")
+        function_match = XML_FUNCTION_CALL_PATTERN.search(body)
+        if function_match is None:
+            return match.group(0)
+
+        function_name = function_match.group("name").strip()
+        raw_arguments = function_match.group("arguments").strip()
+        arguments = _parse_xml_parameters(raw_arguments)
+        if arguments is None:
+            arguments = _parse_tool_arguments(raw_arguments, parse_mode, response) if raw_arguments else {}
+        tool_calls.append(
+            ToolCall(
+                call_id=f"xml_tool_call_{len(tool_calls) + 1}",
+                func_name=function_name,
+                args=arguments,
+            )
+        )
+        return ""
+
+    cleaned_text = XML_TOOL_CALL_PATTERN.sub(_replace_tool_call, raw_text).strip() or None
+    return cleaned_text, tool_calls or None
+
+
 def _log_length_truncation(finish_reason: str | None, model_name: str | None) -> None:
    """记录因长度截断导致的告警日志。

@@ -552,6 +631,38 @@ def _log_length_truncation(finish_reason: str | None, model_name: str | None) ->
        logger.info(f"模型{model_name or ''}因为超过最大 max_token 限制，可能仅输出部分内容，可视情况调整")


+def _apply_xml_tool_call_fallback(
+    response: APIResponse,
+    parse_mode: ToolArgumentParseMode,
+    raw_response: Any,
+) -> None:
+    """当上游未返回标准 tool_calls 时，尝试从 XML 文本兜底解析。
+
+    这是一个暂时性处理方法，用来兼容思维链中混入工具调用的返回格式，
+    后续可能随着模型或上游接口的规范化而变更。
+    """
+    if response.tool_calls:
+        return
+
+    reasoning_content, tool_calls = _extract_xml_tool_calls(response.reasoning_content, parse_mode, raw_response)
+    if reasoning_content != response.reasoning_content:
+        response.reasoning_content = reasoning_content
+    if tool_calls:
+        response.tool_calls = tool_calls
+        if not response.content and reasoning_content:
+            response.content = reasoning_content
+            response.reasoning_content = None
+        logger.warning("OpenAI 兼容响应未返回标准 tool_calls，已从 XML 文本兜底解析工具调用")
+        return
+
+    cleaned_content, tool_calls = _extract_xml_tool_calls(response.content, parse_mode, raw_response)
+    if cleaned_content != response.content:
+        response.content = cleaned_content
+    if tool_calls:
+        response.tool_calls = tool_calls
+        logger.warning("OpenAI 兼容响应未返回标准 tool_calls，已从 XML 文本兜底解析工具调用")
+
+
 def _coerce_openai_argument(value: Any) -> Any | Omit:
    """将可选参数转换为 OpenAI SDK 期望的值。

@@ -748,6 +859,7 @@ class _OpenAIStreamAccumulator:
                response.tool_calls.append(ToolCall(call_id=call_id, func_name=state.function_name, args=arguments))

        response.raw_data = {"model": self.model_name} if self.model_name else None
+        _apply_xml_tool_call_fallback(response, self.tool_argument_parse_mode, response.raw_data)

        if not response.content and not response.tool_calls:
            raise EmptyResponseException(response.raw_data)
@@ -873,6 +985,7 @@ def _default_normal_response_parser(

    finish_reason = getattr(resp.choices[0], "finish_reason", None)
    _log_length_truncation(finish_reason, getattr(resp, "model", None))
+    _apply_xml_tool_call_fallback(api_response, tool_argument_parse_mode, resp)

    if not api_response.content and not api_response.tool_calls:
        raise EmptyResponseException(resp)