diff --git a/app/agent/__init__.py b/app/agent/__init__.py index 1666a861..b815da90 100644 --- a/app/agent/__init__.py +++ b/app/agent/__init__.py @@ -373,7 +373,7 @@ class MoviePilotAgent: HumanMessage(content=str(message).strip()[:1000]), ] ) - content = LLMHelper._extract_text_content(getattr(response, "content", response)) + content = LLMHelper.extract_text_content(getattr(response, "content", response)) return self._sanitize_chat_title(content) async def prepare_chat_title(self, message: str) -> None: @@ -738,39 +738,6 @@ class MoviePilotAgent: runtime_config = await self._resolve_llm_runtime_config() return await LLMHelper.get_llm(streaming=streaming, **runtime_config) - @staticmethod - def _extract_text_content(content) -> str: - """ - 从消息内容中提取纯文本，过滤掉思考/推理类型的内容块。 - :param content: 消息内容，可能是字符串或内容块列表 - :return: 纯文本内容 - """ - if not content: - return "" - # 跳过思考/推理类型的内容块 - if isinstance(content, list): - text_parts = [] - for block in content: - if isinstance(block, str): - text_parts.append(block) - elif isinstance(block, dict): - # 优先检查 thought 标志（LangChain Google GenAI 方案） - if block.get("thought"): - continue - if block.get("type") in ( - "thinking", - "reasoning_content", - "reasoning", - "thought", - ): - continue - if block.get("type") == "text": - text_parts.append(block.get("text", "")) - else: - text_parts.append(str(block)) - return "".join(text_parts) - return str(content) - @classmethod def _has_image_input_content(cls, content: Any) -> bool: """ @@ -1252,7 +1219,7 @@ class MoviePilotAgent: if token.content: # content 可能是字符串或内容块列表，过滤掉思考类型的块 - content = self._extract_text_content(token.content) + content = LLMHelper.extract_text_content(token.content) if content: stripper.process(content, on_token) @@ -1355,7 +1322,7 @@ class MoviePilotAgent: for msg in reversed(final_messages): if hasattr(msg, "type") and msg.type == "ai" and msg.content: # 过滤掉思考/推理内容，只提取纯文本 - text = self._extract_text_content(msg.content) + text = LLMHelper.extract_text_content(msg.content) if text: # 过滤掉包含在 标签中的内容 text = re.sub( @@ -1388,7 +1355,7 @@ class MoviePilotAgent: ) for msg in reversed(final_messages): if hasattr(msg, "type") and msg.type == "ai" and msg.content: - display_text = self._extract_text_content(msg.content).strip() + display_text = LLMHelper.extract_text_content(msg.content).strip() break self._save_assistant_display_message_once(display_text) diff --git a/app/agent/llm/helper.py b/app/agent/llm/helper.py index 2b10b1fd..3e488893 100644 --- a/app/agent/llm/helper.py +++ b/app/agent/llm/helper.py @@ -1014,9 +1014,13 @@ class LLMHelper: return model @staticmethod - def _extract_text_content(content) -> str: + def extract_text_content(content: Any, fallback_to_string: bool = False) -> str: """ 从响应内容中提取纯文本，仅保留真实文本块。 + + :param content: 模型响应内容，可能是字符串、字典或内容块列表 + :param fallback_to_string: 未识别为文本内容时是否回退为字符串 + :return: 提取后的纯文本内容 """ if content is None: return "" @@ -1051,7 +1055,7 @@ class LLMHelper: return content.get("text", "") if not content.get("type") and isinstance(content.get("text"), str): return content.get("text", "") - return "" + return str(content) if fallback_to_string else "" @staticmethod async def test_current_settings( @@ -1092,7 +1096,7 @@ class LLMHelper: duration_ms = round((time.perf_counter() - start) * 1000) raise LLMTestError(str(err), duration_ms=duration_ms) from err - reply_text = LLMHelper._extract_text_content( + reply_text = LLMHelper.extract_text_content( getattr(response, "content", response) ).strip() duration_ms = round((time.perf_counter() - start) * 1000) diff --git a/app/agent/middleware/subagents.py b/app/agent/middleware/subagents.py index e84289a9..8aa11969 100644 --- a/app/agent/middleware/subagents.py +++ b/app/agent/middleware/subagents.py @@ -23,6 +23,7 @@ from langchain_core.messages import AIMessage, HumanMessage from langchain_core.tools import BaseTool, StructuredTool from pydantic import BaseModel, Field +from app.agent.llm import LLMHelper from app.agent.middleware.utils import append_to_system_message from app.agent.runtime import SubAgentDefinition, agent_runtime_manager from app.agent.tools.tags import ToolTag @@ -281,34 +282,6 @@ def _format_subagent_catalog(profiles: tuple[_SubAgentProfile, ...]) -> str: ) -def _extract_text_content(content: Any) -> str: - """从模型消息内容中提取可读文本。""" - if content is None: - return "" - if isinstance(content, str): - return content - if isinstance(content, list): - text_parts: list[str] = [] - for block in content: - if isinstance(block, str): - text_parts.append(block) - continue - if isinstance(block, dict): - if block.get("thought"): - continue - if block.get("type") in { - "thinking", - "reasoning_content", - "reasoning", - "thought", - }: - continue - if isinstance(block.get("text"), str): - text_parts.append(block["text"]) - return "".join(text_parts) - return str(content) - - def _extract_final_text(result: Any) -> str: """从子代理执行结果中提取最后一条 AI 文本。""" if isinstance(result, dict): @@ -318,11 +291,11 @@ def _extract_final_text(result: Any) -> str: for message in reversed(messages): if isinstance(message, AIMessage) and message.content: - text = _extract_text_content(message.content).strip() + text = LLMHelper.extract_text_content(message.content).strip() if text: return text - return _extract_text_content(result).strip() + return LLMHelper.extract_text_content(result, fallback_to_string=True).strip() def _clip_text(text: Any, max_chars: int) -> tuple[str, bool]: diff --git a/app/agent/middleware/tool_selection.py b/app/agent/middleware/tool_selection.py index 1a0082f3..fa305b4d 100644 --- a/app/agent/middleware/tool_selection.py +++ b/app/agent/middleware/tool_selection.py @@ -26,6 +26,7 @@ from langchain_core.tools import BaseTool from langgraph.runtime import Runtime from typing_extensions import TypedDict # noqa +from app.agent.llm import LLMHelper from app.agent.tools.tags import ToolTag from app.log import logger @@ -121,7 +122,7 @@ class ToolSelectorMiddleware(LLMToolSelectorMiddleware): else: continue - content = cls._extract_text_content(message.content).strip() + content = LLMHelper.extract_text_content(message.content).strip() if not content: continue rendered_messages.append(f"{role}: {content}") @@ -380,39 +381,6 @@ class ToolSelectorMiddleware(LLMToolSelectorMiddleware): or "api.deepseek.com" in base_url ) - @staticmethod - def _extract_text_content(content: Any) -> str: - """ - 从模型响应中提取纯文本。 - - 这里不依赖上层 LLMHelper，避免中间件与 LLM 构造逻辑互相耦合。 - """ - if content is None: - return "" - if isinstance(content, str): - return content - if isinstance(content, list): - text_parts: list[str] = [] - for block in content: - if isinstance(block, str): - text_parts.append(block) - continue - if isinstance(block, dict): - if block.get("type") == "text" and isinstance( - block.get("text"), str - ): - text_parts.append(block["text"]) - continue - if not block.get("type") and isinstance(block.get("text"), str): - text_parts.append(block["text"]) - return "".join(text_parts) - if isinstance(content, dict): - if content.get("type") == "text" and isinstance(content.get("text"), str): - return content["text"] - if not content.get("type") and isinstance(content.get("text"), str): - return content["text"] - return "" - @staticmethod def _parse_json_object(text: str) -> dict[str, Any]: """ @@ -504,7 +472,7 @@ class ToolSelectorMiddleware(LLMToolSelectorMiddleware): 解析并标准化 DeepSeek JSON 模式的工具筛选结果。 """ content = getattr(response, "content", response) - text = self._extract_text_content(content) + text = LLMHelper.extract_text_content(content) logger.debug(f"工具筛选原始响应: {text}") payload = self._parse_json_object(text) diff --git a/tests/test_llm_helper_testcall.py b/tests/test_llm_helper_testcall.py index cc16852f..ce2250c1 100644 --- a/tests/test_llm_helper_testcall.py +++ b/tests/test_llm_helper_testcall.py @@ -214,7 +214,7 @@ class LlmHelperTestCallTest(unittest.TestCase): {"type": "text", "text": "OK"}, ] - result = llm_module.LLMHelper._extract_text_content(content) + result = llm_module.LLMHelper.extract_text_content(content) self.assertEqual(result, "OK")