From e05eacfae3e1e439f22f276ce36ae4fe2c5be6d4 Mon Sep 17 00:00:00 2001 From: Wenjie Zhang Date: Thu, 2 Apr 2026 09:05:56 +0800 Subject: [PATCH] =?UTF-8?q?feat:=20=E4=BC=98=E5=8C=96=20summary=20?= =?UTF-8?q?=E7=BB=84=E4=BB=B6?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../agents/middlewares/summary_middleware.py | 19 +++++++------------ .../package/yuxi/services/subagent_service.py | 3 ++- 2 files changed, 9 insertions(+), 13 deletions(-) diff --git a/backend/package/yuxi/agents/middlewares/summary_middleware.py b/backend/package/yuxi/agents/middlewares/summary_middleware.py index 689a1d11..71323840 100644 --- a/backend/package/yuxi/agents/middlewares/summary_middleware.py +++ b/backend/package/yuxi/agents/middlewares/summary_middleware.py @@ -7,6 +7,7 @@ from __future__ import annotations +from pathlib import Path import uuid from collections.abc import Callable, Iterable, Mapping from functools import partial @@ -31,6 +32,8 @@ from langchain_core.messages.utils import ( from langgraph.graph.message import REMOVE_ALL_MESSAGES from langgraph.runtime import Runtime +from yuxi.utils.paths import OUTPUTS_DIR_NAME + TokenCounter = Callable[[Iterable[MessageLikeRepresentation]], int] DEFAULT_SUMMARY_PROMPT = """ @@ -74,9 +77,7 @@ Messages to summarize: """ _DEFAULT_MESSAGES_TO_KEEP = 20 -_DEFAULT_TRIM_TOKEN_LIMIT = 4000 _DEFAULT_FALLBACK_MESSAGE_COUNT = 15 -_DEFAULT_OFFLOAD_THRESHOLD = 1000 # Token 数阈值，超过此值则卸载到文件系统 _OFFLOAD_DIR = "/summary_offload" # 虚拟文件系统路径 ContextFraction = tuple[Literal["fraction"], float] @@ -143,7 +144,7 @@ def _offload_tool_result(msg: ToolMessage, threshold: int, token_counter: TokenC # 生成文件路径 (工具名称-xxx) message_id = msg.id or str(uuid.uuid4())[:8] safe_name = "".join(c if c.isalnum() or c in "-_" else "_" for c in tool_name) - file_path = f"{_OFFLOAD_DIR}/{safe_name}-{message_id}" + file_path = (Path(OUTPUTS_DIR_NAME) / f"{_OFFLOAD_DIR}/{safe_name}-{message_id}").as_posix() # 构建文件头部信息 header_lines = [ @@ -226,7 +227,7 @@ class SummaryOffloadMiddleware(AgentMiddleware): keep: ContextSize = ("messages", _DEFAULT_MESSAGES_TO_KEEP), token_counter: TokenCounter = count_tokens_approximately, summary_prompt: str = DEFAULT_SUMMARY_PROMPT, - trim_tokens_to_summarize: int | None = _DEFAULT_TRIM_TOKEN_LIMIT, + trim_tokens_to_summarize: int | None = 4000, # 工具结果卸载参数 summary_offload_threshold: int = 1000, max_retention_ratio: float = 0.6, @@ -240,18 +241,12 @@ class SummaryOffloadMiddleware(AgentMiddleware): keep: 摘要后保留的消息数量/ token 策略 (作为 fallback) token_counter: token 计数函数 summary_prompt: 生成摘要的提示词模板 - trim_tokens_to_summarize: 准备摘要消息时的最大 token 数 - summary_offload_threshold: Summary 时卸载阈值（token 数），默认 1000 + trim_tokens_to_summarize: Summary 时，无损保留的消息数 + summary_offload_threshold: Summary 时，工具调用结果超过此 token 数阈值则卸载到文件系统 max_retention_ratio: 触发 Summary 后，如果不超过此比例（相对于 trigger），则不删除消息。默认 0.6 """ super().__init__() - # Handle renamed argument for backward compatibility if needed, - # but since we are refactoring, we map deprecated 'result_offload_threshold' - # to 'summary_offload_threshold' if present in kwargs. - if "result_offload_threshold" in deprecated_kwargs: - summary_offload_threshold = deprecated_kwargs.pop("result_offload_threshold") - if isinstance(model, str): model = init_chat_model(model) diff --git a/backend/package/yuxi/services/subagent_service.py b/backend/package/yuxi/services/subagent_service.py index 8a6b3a88..c752dd1a 100644 --- a/backend/package/yuxi/services/subagent_service.py +++ b/backend/package/yuxi/services/subagent_service.py @@ -9,6 +9,7 @@ from sqlalchemy.ext.asyncio import AsyncSession from yuxi.repositories.subagent_repository import SubAgentRepository from yuxi.storage.postgres.manager import pg_manager from yuxi.utils import logger +from yuxi.utils.paths import OUTPUTS_DIR_NAME # SubAgent specs cache for get_subagent_specs _subagent_specs_cache: list[dict[str, Any]] | None = None @@ -34,7 +35,7 @@ _DEFAULT_SUBAGENTS = [ "你是一位专注的研究员。你的工作是根据用户的问题进行研究。" "进行彻底的研究，然后用详细的答案回复用户的问题，只有你的最终答案会被传递给用户。" "除了你的最终信息，他们不会知道任何其他事情，所以你的最终报告应该就是你的最终信息！" - "将调研结果保存到主题研究文件中 sub_research/xxx.md 中。" + f"将调研结果保存到主题研究文件中 {OUTPUTS_DIR_NAME}/sub_research/xxx.md 中。" ), "tools": ["tavily_search"], "is_builtin": True,