mirror of
https://github.com/jxxghp/MoviePilot.git
synced 2026-03-31 09:21:04 +08:00
refactor: 活动日志摘要改用 LLM 总结替代文本截取
This commit is contained in:
@@ -2,7 +2,7 @@
|
||||
活动日志中间件 - 自动记录 Agent 每次交互的操作摘要。
|
||||
|
||||
按日期存储在 CONFIG_PATH/agent/activity/YYYY-MM-DD.md 中,
|
||||
每次 Agent 执行完毕后自动追加一条活动记录,
|
||||
每次 Agent 执行完毕后自动调用 LLM 对本轮对话生成简洁的活动摘要,
|
||||
并在每次 Agent 启动时加载近几天的活动日志注入系统提示词。
|
||||
"""
|
||||
|
||||
@@ -21,7 +21,7 @@ from langchain.agents.middleware.types import (
|
||||
PrivateStateAttr, # noqa
|
||||
ResponseT,
|
||||
)
|
||||
from langchain_core.messages import AIMessage, HumanMessage
|
||||
from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
|
||||
from langgraph.runtime import Runtime
|
||||
|
||||
from app.agent.middleware.utils import append_to_system_message
|
||||
@@ -30,15 +30,23 @@ from app.log import logger
|
||||
# 活动日志保留天数
|
||||
DEFAULT_RETENTION_DAYS = 7
|
||||
|
||||
# 单次活动记录的最大长度
|
||||
MAX_ENTRY_LENGTH = 300
|
||||
|
||||
# 注入系统提示词时加载的天数
|
||||
PROMPT_LOAD_DAYS = 3
|
||||
|
||||
# 每日日志文件最大大小 (256KB)
|
||||
MAX_LOG_FILE_SIZE = 256 * 1024
|
||||
|
||||
# 提取本轮对话上下文的最大字符数(避免过长的对话消耗太多 token)
|
||||
MAX_CONTEXT_FOR_SUMMARY = 4000
|
||||
|
||||
# LLM 总结的提示词
|
||||
SUMMARY_PROMPT = """请根据以下 AI 助手与用户的对话记录,生成一条简洁的活动摘要(中文,一句话,不超过80字)。
|
||||
摘要应包含:用户的需求是什么、助手做了什么、结果如何。
|
||||
只输出摘要内容,不要加任何前缀、标点序号或解释。
|
||||
|
||||
对话记录:
|
||||
{conversation}"""
|
||||
|
||||
|
||||
class ActivityLogState(AgentState):
|
||||
"""ActivityLogMiddleware 的状态模型。"""
|
||||
@@ -53,22 +61,21 @@ class ActivityLogStateUpdate(TypedDict):
|
||||
activity_log_contents: dict[str, str]
|
||||
|
||||
|
||||
def _extract_activity_summary(messages: list) -> str | None:
|
||||
"""从本次对话的消息列表中提取活动摘要。
|
||||
def _extract_last_round(messages: list) -> list | None:
|
||||
"""从完整消息列表中提取最后一轮交互。
|
||||
|
||||
只关注最后一轮交互(从最后一条用户消息到结尾),
|
||||
分析用户问题、Agent 使用的工具、最终回复,生成一行简洁的活动描述。
|
||||
从最后一条 HumanMessage 到消息末尾即为本轮交互。
|
||||
|
||||
参数:
|
||||
messages: Agent 执行后的完整消息列表。
|
||||
|
||||
返回:
|
||||
活动摘要字符串,如果没有有意义的活动则返回 None。
|
||||
本轮交互的消息子列表,如果无有效交互则返回 None。
|
||||
"""
|
||||
if not messages:
|
||||
return None
|
||||
|
||||
# 找到最后一条用户消息的索引,从此处开始截取本轮交互
|
||||
# 找到最后一条用户消息的索引
|
||||
last_human_idx = None
|
||||
for i in range(len(messages) - 1, -1, -1):
|
||||
if isinstance(messages[i], HumanMessage) and messages[i].content:
|
||||
@@ -78,52 +85,91 @@ def _extract_activity_summary(messages: list) -> str | None:
|
||||
if last_human_idx is None:
|
||||
return None
|
||||
|
||||
# 本轮交互的消息
|
||||
round_messages = messages[last_human_idx:]
|
||||
|
||||
# 提取用户问题
|
||||
# 检查是否为系统心跳消息
|
||||
user_msg = round_messages[0]
|
||||
user_content = (
|
||||
user_msg.content if isinstance(user_msg.content, str) else str(user_msg.content)
|
||||
)
|
||||
|
||||
# 跳过系统心跳消息
|
||||
if user_content.strip().startswith("[System Heartbeat]"):
|
||||
return None
|
||||
|
||||
user_query = user_content.strip()[:100]
|
||||
return round_messages
|
||||
|
||||
|
||||
def _format_conversation_for_summary(round_messages: list) -> str:
|
||||
"""将本轮对话消息格式化为文本,供 LLM 总结。
|
||||
|
||||
参数:
|
||||
round_messages: 本轮交互的消息列表。
|
||||
|
||||
返回:
|
||||
格式化后的对话文本。
|
||||
"""
|
||||
lines = []
|
||||
total_len = 0
|
||||
|
||||
# 收集本轮交互中使用的工具(仅限本轮)
|
||||
tool_names = set()
|
||||
for msg in round_messages:
|
||||
if isinstance(msg, AIMessage) and hasattr(msg, "tool_calls") and msg.tool_calls:
|
||||
for tc in msg.tool_calls:
|
||||
if isinstance(tc, dict) and "name" in tc:
|
||||
tool_names.add(tc["name"])
|
||||
|
||||
# 提取本轮最后一条 AI 回复的摘要
|
||||
ai_reply = None
|
||||
for msg in reversed(round_messages):
|
||||
if (
|
||||
isinstance(msg, AIMessage)
|
||||
and msg.content
|
||||
and not (hasattr(msg, "tool_calls") and msg.tool_calls)
|
||||
):
|
||||
if isinstance(msg, HumanMessage):
|
||||
content = msg.content if isinstance(msg.content, str) else str(msg.content)
|
||||
ai_reply = content.strip()[:120]
|
||||
line = f"用户: {content}"
|
||||
elif isinstance(msg, AIMessage):
|
||||
if hasattr(msg, "tool_calls") and msg.tool_calls:
|
||||
tool_names = [
|
||||
tc["name"]
|
||||
for tc in msg.tool_calls
|
||||
if isinstance(tc, dict) and "name" in tc
|
||||
]
|
||||
line = f"助手调用工具: {', '.join(tool_names)}"
|
||||
elif msg.content:
|
||||
content = (
|
||||
msg.content if isinstance(msg.content, str) else str(msg.content)
|
||||
)
|
||||
line = f"助手: {content}"
|
||||
else:
|
||||
continue
|
||||
elif isinstance(msg, ToolMessage):
|
||||
content = msg.content if isinstance(msg.content, str) else str(msg.content)
|
||||
# 工具返回可能很长,截断
|
||||
if len(content) > 200:
|
||||
content = content[:200] + "..."
|
||||
line = f"工具返回: {content}"
|
||||
else:
|
||||
continue
|
||||
|
||||
# 控制总长度
|
||||
if total_len + len(line) > MAX_CONTEXT_FOR_SUMMARY:
|
||||
lines.append("...(后续对话省略)")
|
||||
break
|
||||
lines.append(line)
|
||||
total_len += len(line)
|
||||
|
||||
# 组装摘要
|
||||
parts = [f"用户: {user_query}"]
|
||||
if tool_names:
|
||||
parts.append(f"工具: {', '.join(sorted(tool_names))}")
|
||||
if ai_reply:
|
||||
parts.append(f"结果: {ai_reply}")
|
||||
return "\n".join(lines)
|
||||
|
||||
summary = " | ".join(parts)
|
||||
if len(summary) > MAX_ENTRY_LENGTH:
|
||||
summary = summary[: MAX_ENTRY_LENGTH - 3] + "..."
|
||||
return summary
|
||||
|
||||
async def _summarize_with_llm(conversation_text: str) -> str | None:
|
||||
"""调用 LLM 对对话文本生成活动摘要。
|
||||
|
||||
参数:
|
||||
conversation_text: 格式化后的对话文本。
|
||||
|
||||
返回:
|
||||
LLM 生成的摘要字符串,失败时返回 None。
|
||||
"""
|
||||
try:
|
||||
from app.helper.llm import LLMHelper
|
||||
|
||||
llm = LLMHelper.get_llm(streaming=False)
|
||||
prompt = SUMMARY_PROMPT.format(conversation=conversation_text)
|
||||
response = await llm.ainvoke(prompt)
|
||||
summary = response.content.strip()
|
||||
# 清理模型可能输出的前缀(如 "摘要:" "总结:")
|
||||
summary = re.sub(r"^(摘要|总结|活动记录)[::]\s*", "", summary)
|
||||
return summary if summary else None
|
||||
except Exception as e:
|
||||
logger.debug("LLM summarization failed: %s", e)
|
||||
return None
|
||||
|
||||
|
||||
ACTIVITY_LOG_SYSTEM_PROMPT = """<activity_log>
|
||||
@@ -331,14 +377,24 @@ class ActivityLogMiddleware(AgentMiddleware[ActivityLogState, ContextT, Response
|
||||
async def aafter_agent(
|
||||
self, state: ActivityLogState, runtime: Runtime
|
||||
) -> dict[str, Any] | None:
|
||||
"""Agent 执行完毕后,从对话中提取摘要并追加到当日活动日志。"""
|
||||
"""Agent 执行完毕后,调用 LLM 对本轮对话生成摘要并追加到当日活动日志。"""
|
||||
try:
|
||||
messages = state.get("messages", [])
|
||||
if not messages:
|
||||
return None
|
||||
|
||||
# 提取活动摘要
|
||||
summary = _extract_activity_summary(messages)
|
||||
# 提取本轮交互
|
||||
round_messages = _extract_last_round(messages)
|
||||
if not round_messages:
|
||||
return None
|
||||
|
||||
# 格式化对话文本
|
||||
conversation_text = _format_conversation_for_summary(round_messages)
|
||||
if not conversation_text:
|
||||
return None
|
||||
|
||||
# 调用 LLM 生成摘要
|
||||
summary = await _summarize_with_llm(conversation_text)
|
||||
if summary:
|
||||
await self._append_activity(summary)
|
||||
except Exception as e:
|
||||
|
||||
Reference in New Issue
Block a user