@
feat(圆桌): Step3 三产物并行 + 大屏多页模式 + 后台挂起查看恢复流程 后台挂起 Step3: - engine 总结报告/结果绘制/大屏数据 asyncio.gather 三流并行,故障隔离 - request.state 回归修复(进程内虚拟 request 补 state),恢复后台报告 - ROUNDTABLE_JOB_GATEWAY 默认 inprocess;产物自愈重试(截断/缺失纠偏) - 大屏后端只产 report-json(dashboardJson),前端按当前主题渲染 结果绘制大屏多页模式(模式二): - report-template 模板渲染(report-json→大屏HTML),零截断、主题跟随 - 与模式一同款交互(流式+沙箱+可修改);沙箱新窗口打开按钮 最终方案展示:左侧席位栏+总控头像、交付/文件合并、收展动画、内联沙箱 后台挂起查看=恢复实时流程: - handleLoadDraft 返回 Promise,await hydrate 后订阅 streamJob SSE 跟随 - saveDraftNow 强制存(发起前+查看前各一次)防丢 第二步优化:总控判断串/并行(并行单列可写)、席位上下文+交付模板、 peerDeliveries、失败重试、共识标记、流程图小窗(缩放/拖动/并行批) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> @
This commit is contained in:
1 parent
a7596001c5
commit
d5f319d9cb
51 files changed
+5016
-1186
No files matched your search
+295
-220
@@ -1,4 +1,17 @@
|
||||
"""素材收集专家节点:关键词提取 → 并发搜索(含笔记) → 相关度排序。"""
|
||||
"""素材收集专家节点:关键词提取 → 配置技能编排检索(含笔记) → 相关度排序。
|
||||
|
||||
检索类型(material_source)只有两种:
|
||||
- ``general``(通用检索,默认):按素材收集专家 agent(ai-writing-researcher)
|
||||
config.yaml 配置的技能顺序编排检索。内置「知识库检索」技能
|
||||
(knowledge-base-search)走原生 web/intranet 关键词并发检索;其它技能用
|
||||
其 description 生成定向 query 后调真实 provider 检索。
|
||||
- ``notebook``(我的空间):仅用笔记 API 素材。
|
||||
旧值 ``knowledge_base`` / ``skill`` 一律按 ``general`` 兼容处理。
|
||||
|
||||
PAUSE-1 素材确认卡的单输入框:用户用自然语言描述想补充检索的内容
|
||||
(如「检索 A 的数据」)时,让模型从已配置技能中按相关度匹配,逐个检索,
|
||||
把新素材**追加**到现有素材列表(URL 去重、保持原有顺序、不重排不丢弃)。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
@@ -6,8 +19,7 @@ import logging
|
||||
import os
|
||||
import re
|
||||
import time
|
||||
import uuid
|
||||
from typing import List
|
||||
from types import SimpleNamespace
|
||||
|
||||
from langchain_core.runnables import RunnableConfig
|
||||
|
||||
@@ -71,11 +83,63 @@ _MAX_MAX_MATERIALS = 100
|
||||
_DEFAULT_KEYWORD_COUNT = 6
|
||||
_MAX_KEYWORD_COUNT = 10
|
||||
|
||||
# 单个技能的 web 检索超时(秒)。技能检索是顺序循环、最多 5 个技能,
|
||||
# 单个技能的 web 检索超时(秒)。技能检索是顺序循环,
|
||||
# DuckDuckGo 等 provider 偶发限速/挂起会把整段"技能检索过程"拖到几十秒看起来卡死;
|
||||
# 超时即放弃该技能、emit 0 条结果继续下一个,保证时间线持续推进。
|
||||
_SKILL_SEARCH_TIMEOUT_SECONDS = 25
|
||||
|
||||
# 内置「知识库检索」技能名:素材收集专家 agent 的默认技能。researcher 对它
|
||||
# 特殊处理——不用 description 拼 query,而是对 Step 1 生成的全部检索词做
|
||||
# 原生 web/intranet 并发检索(即旧版 knowledge_base 模式的行为)。
|
||||
KNOWLEDGE_SEARCH_SKILL_NAME = "knowledge-base-search"
|
||||
_KNOWLEDGE_SKILL_DISPLAY = "知识库检索"
|
||||
_KNOWLEDGE_SKILL_FALLBACK_DESC = "对写作意图提炼的检索词执行 web/intranet 知识库检索,获取相关素材"
|
||||
|
||||
|
||||
def _fallback_knowledge_skill() -> SimpleNamespace:
|
||||
"""技能目录里找不到 knowledge-base-search 时的内存兜底(原生路径不依赖 SKILL.md)。"""
|
||||
return SimpleNamespace(name=KNOWLEDGE_SEARCH_SKILL_NAME, description=_KNOWLEDGE_SKILL_FALLBACK_DESC)
|
||||
|
||||
|
||||
def _load_configured_skills() -> list:
|
||||
"""解析素材收集专家 agent(ai-writing-researcher)配置的技能列表。
|
||||
|
||||
config.yaml 的 ``skills`` 字段是权威来源(admin/用户在智能体管理页可改);
|
||||
未配置 / 解析失败 / 技能文件全部缺失时,一律回退为内置「知识库检索」,
|
||||
保证通用检索永远可用。返回的元素带 ``name`` / ``description`` 属性。
|
||||
"""
|
||||
names: list[str] = []
|
||||
try:
|
||||
from deerflow.config.agents_config import load_agent_config
|
||||
|
||||
cfg = load_agent_config(RESEARCHER_AGENT_ID)
|
||||
if cfg is not None:
|
||||
names = [n.strip() for n in (cfg.skills or []) if isinstance(n, str) and n.strip()]
|
||||
except Exception:
|
||||
logger.warning("researcher: 读取 agent 配置技能失败,回退知识库检索", exc_info=True)
|
||||
|
||||
if not names:
|
||||
return [_fallback_knowledge_skill()]
|
||||
|
||||
by_name: dict[str, object] = {}
|
||||
try:
|
||||
from deerflow.skills.storage import get_or_new_skill_storage
|
||||
|
||||
by_name = {s.name: s for s in get_or_new_skill_storage().load_skills(enabled_only=False)}
|
||||
except Exception:
|
||||
logger.warning("researcher: 加载技能目录失败", exc_info=True)
|
||||
|
||||
resolved: list = []
|
||||
for n in names:
|
||||
sk = by_name.get(n)
|
||||
if sk is None and n == KNOWLEDGE_SEARCH_SKILL_NAME:
|
||||
sk = _fallback_knowledge_skill()
|
||||
if sk is None:
|
||||
logger.warning("researcher: 配置的技能 '%s' 不存在于技能目录,已跳过", n)
|
||||
continue
|
||||
resolved.append(sk)
|
||||
return resolved or [_fallback_knowledge_skill()]
|
||||
|
||||
|
||||
def _resolve_keyword_count(raw: object) -> int:
|
||||
"""把用户传入的检索关键词数量规整到 [1, _MAX_KEYWORD_COUNT],非法值回退默认 6。"""
|
||||
@@ -247,32 +311,24 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
# 检查干预模式(PAUSE-1 素材确认 / 素材不足求助)
|
||||
intervention = state.get("last_user_intervention") or {}
|
||||
extra_keywords: list[str] = intervention.get("_extra_keywords") or intervention.get("extra_keywords") or []
|
||||
# PAUSE-1 单输入框:用户写的自然语言(技能名 / 来源 / 检索意图),优先级最高
|
||||
# PAUSE-1 单输入框:用户写的自然语言(想补充检索的内容),优先级最高
|
||||
user_query: str = (intervention.get("user_query") or "").strip()
|
||||
action = intervention.get("action")
|
||||
|
||||
# 素材不足求助点触发的「补充检索」:保留原大纲,检索完直接回写作、不走素材确认/大纲
|
||||
is_supplement = action == "supplement_search"
|
||||
# PAUSE-1 素材确认卡用户点「按输入重新搜索」并填了 user_query:
|
||||
# PAUSE-1 素材确认卡用户在输入框填了自然语言(如「检索 A 的数据」):
|
||||
# - 跳过 LLM 重新生成检索词(节省一次调用 + 尊重用户意图)
|
||||
# - 直接用 user_query 作为唯一检索词
|
||||
# - 保留 existing material_package,按 URL 去重合并
|
||||
is_user_query_search = action == "re_search" and bool(user_query)
|
||||
# - 从 agent 配置的技能中匹配最相关的技能,逐个检索
|
||||
# - **追加**:保留 existing material_package,新素材按 URL 去重后追加在末尾
|
||||
# 旧版「启用技能检索」按钮(action=enable_skill_search)按同一语义兼容处理。
|
||||
is_user_query_search = action in ("re_search", "enable_skill_search") and bool(user_query)
|
||||
# 兼容旧版"追加关键词"调用方(如果还有 client 传 extra_keywords 数组而非 user_query)
|
||||
is_append_only = action == "re_search" and not user_query and bool(extra_keywords)
|
||||
# 用户在写作表单选择的素材来源:knowledge_base(默认 web/intranet 检索)/ skill(技能检索)。
|
||||
# 首次检索(没有用户干预)时按它决定检索方式;缺省 knowledge_base。
|
||||
material_source = (state.get("material_source") or "knowledge_base").strip()
|
||||
is_first_run = not intervention
|
||||
is_initial_skill_source = is_first_run and material_source == "skill"
|
||||
# 技能检索触发条件(任一即可):
|
||||
# - 首次检索且表单选了「技能检索」来源
|
||||
# - PAUSE-1 素材确认卡用户点「启用技能检索」按钮
|
||||
# 技能检索流程:
|
||||
# - 不再走 web/intranet 检索,改用 SkillSearchProvider
|
||||
# - 顺序调用候选技能 + 早停(第一个技能够就停,不够再加下一个)
|
||||
# - 旧 material_package 完全丢弃(覆盖式替换,让用户对比 web vs skill 结果)
|
||||
is_skill_search = action == "enable_skill_search" or is_initial_skill_source
|
||||
# 用户在写作表单选择的检索类型:general(通用检索,按 agent 配置技能编排)/
|
||||
# notebook(我的空间)。旧值 knowledge_base / skill 一律按 general 兼容。
|
||||
raw_source = (state.get("material_source") or "general").strip()
|
||||
material_source = "notebook" if raw_source == "notebook" else "general"
|
||||
|
||||
# 用户指定的检索关键词数量(1~10),决定生成多少个检索词、检索多少个方向
|
||||
keyword_count = _resolve_keyword_count(state.get("keyword_count"))
|
||||
@@ -291,7 +347,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
max_materials = _resolve_max_materials(ai_writing_config.get("max_materials"))
|
||||
|
||||
# 内置 agent 模式(config.yaml ai_writing.use_builtin_agents.researcher):
|
||||
# 检索编排(并发搜索 / 技能编排 / 笔记检索)仍由节点控制,只把两个 LLM 子任务
|
||||
# 检索编排(技能编排 / 笔记检索)仍由节点控制,只把两个 LLM 子任务
|
||||
# (提取检索词 / 素材排序)换成跑 ai-writing-researcher agent(SOUL 可被 admin
|
||||
# 编辑;模型优先级 agent.model > keyword_model/rank_model > 表单主模型)。
|
||||
use_agent = use_builtin_agent("researcher")
|
||||
@@ -308,14 +364,11 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
writer({"type": "research_progress", "message": message})
|
||||
|
||||
# Step 1: 提取关键词
|
||||
kw_result: dict = {} # 走"用户输入 / 追加 / 技能"分支时跳过 LLM 调用,notice 拼接需要兜底值
|
||||
if is_skill_search:
|
||||
# 技能模式:每个 skill 会自己拼 query(user_query + skill.description),这里不用 LLM 生成关键词
|
||||
keywords: list[str] = [user_query or intent]
|
||||
elif is_user_query_search:
|
||||
kw_result: dict = {} # 走"用户输入 / 追加"分支时跳过 LLM 调用,notice 拼接需要兜底值
|
||||
if is_user_query_search:
|
||||
# 用户输入分支:把整段自然语言作为唯一检索词,尊重原意,不再做关键词拆分
|
||||
keywords = [user_query]
|
||||
_emit(f"按你的输入「{user_query[:30]}」检索…")
|
||||
keywords: list[str] = [user_query]
|
||||
_emit(f"按你的输入「{user_query[:30]}」匹配技能检索…")
|
||||
elif is_append_only:
|
||||
# 兼容旧版追加模式:直接用用户填的关键词,不再调 LLM
|
||||
keywords = _clean_keywords(extra_keywords, intent)
|
||||
@@ -359,7 +412,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
# Step 1.5: 笔记检索(如果传入了 notebook_search_space_id)
|
||||
notebook_search_space_id: int | None = state.get("notebook_search_space_id")
|
||||
notebook_materials: list[Material] = []
|
||||
|
||||
|
||||
if notebook_search_space_id:
|
||||
notebook_source_ids: list[int] = state.get("notebook_source_ids") or []
|
||||
notebook_token: str = cfg.get("notebook_token") or ""
|
||||
@@ -373,12 +426,13 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
if notebook_materials:
|
||||
_emit(f"从笔记中获取到 {len(notebook_materials)} 条素材")
|
||||
|
||||
# Step 2: 搜索
|
||||
# 技能检索:技能不是素材,是"检索策略 / 视角"。流程:
|
||||
# a. SkillPicker 选出按相关度排序的 skills(推 list_skills + select_skills 事件)
|
||||
# b. 按顺序遍历 skills:每个 skill 用其 description 让 LLM 生成针对性关键词
|
||||
# c. 调真实 web/intranet provider 执行检索,推 invoke_skill + skill_result 事件
|
||||
# d. 累积素材到 max_materials 即 break,推 stop_early
|
||||
# Step 2: 检索
|
||||
# 通用检索 = 配置技能编排。技能不是素材,是"检索策略 / 视角":
|
||||
# a. 从 ai-writing-researcher agent config.yaml 读取配置技能(默认知识库检索)
|
||||
# b. 「知识库检索」技能 → 对 Step 1 检索词做原生并发 web/intranet 检索
|
||||
# c. 其它技能 → 用 skill.description 拼定向 query 调真实 provider
|
||||
# d. user_query 模式(PAUSE-1 输入框)→ 先让模型按用户输入匹配技能再检索
|
||||
# e. 累积素材到 max_materials 即 break,推 stop_early
|
||||
skill_steps: list[dict] = []
|
||||
search_errors: list[str] = []
|
||||
|
||||
@@ -390,188 +444,208 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
time.monotonic(), step.get("type"), step.get("skill_name") or "")
|
||||
writer({"type": "skill_search_step", "agent_name": researcher_label, **step})
|
||||
|
||||
if is_skill_search:
|
||||
# ── 技能编排检索 ──
|
||||
# 技能选择走流式 LLM(推 skill_select_chunk)+ <think> 分离,前端打字机展示
|
||||
picker = SkillPicker(
|
||||
config=config,
|
||||
model_name=keyword_model,
|
||||
emit_step=_emit_skill_step,
|
||||
max_select=5,
|
||||
user_query=user_query,
|
||||
)
|
||||
selected = await picker.pick(user_intent=intent)
|
||||
search_errors.extend(picker.errors)
|
||||
# 本轮新检索到的原始结果(统一在下方合并段转 Material、与旧素材去重)
|
||||
new_results: list[SearchResult] = []
|
||||
provider = None
|
||||
|
||||
# 真实 web/intranet provider(按 config.yaml 配置)
|
||||
web_provider = get_search_provider(ai_writing_config)
|
||||
|
||||
# 每个 skill 尽量多取(不分摊),第一个就达到 max_materials 即可早停;
|
||||
# 这才贴近"第一个技能够了就停,否则下一个补充"的语义。
|
||||
per_skill = max(1, max_materials)
|
||||
raw_materials_pre: list[Material] = []
|
||||
seen_urls_pre: set[str] = set()
|
||||
stopped_early = False
|
||||
|
||||
for entry in selected:
|
||||
sk = entry["skill"]
|
||||
skill_query = f"{user_query} {sk.description or ''}".strip() or intent
|
||||
# invoke_skill 步骤带上实际检索词(截断防过长),让用户在每次长 await 之前
|
||||
# 就看到"这一步在用技能 X 检索什么",配合前端 spinner 体现进行中。
|
||||
_emit_skill_step({
|
||||
"type": "invoke_skill",
|
||||
"message": f"使用技能「{sk.name}」检索:{skill_query[:60]}",
|
||||
"skill_name": sk.name,
|
||||
})
|
||||
try:
|
||||
# 加超时兜底:单个技能检索最多等 _SKILL_SEARCH_TIMEOUT_SECONDS 秒,
|
||||
# 超时不再阻塞后续技能,避免整段过程看起来"冻结"。
|
||||
res = await asyncio.wait_for(
|
||||
web_provider.search(skill_query, max_results=per_skill),
|
||||
timeout=_SKILL_SEARCH_TIMEOUT_SECONDS,
|
||||
)
|
||||
except asyncio.TimeoutError:
|
||||
logger.warning("skill[%s] search timed out after %ss", sk.name, _SKILL_SEARCH_TIMEOUT_SECONDS)
|
||||
search_errors.append(f"skill[{sk.name}]: 检索超时(>{_SKILL_SEARCH_TIMEOUT_SECONDS}s)")
|
||||
res = []
|
||||
except Exception as e:
|
||||
logger.warning("skill[%s] search failed: %s", sk.name, e)
|
||||
search_errors.append(f"skill[{sk.name}]: {e}")
|
||||
res = []
|
||||
|
||||
# 去重累积
|
||||
new_count = 0
|
||||
for r in res or []:
|
||||
if r.url and r.url in seen_urls_pre:
|
||||
continue
|
||||
if r.url:
|
||||
seen_urls_pre.add(r.url)
|
||||
raw_materials_pre.append(Material(
|
||||
id=f"m{len(raw_materials_pre)+1:03d}",
|
||||
title=r.title,
|
||||
content=r.content[:500],
|
||||
source=f"{r.source}(via 技能 {sk.name})",
|
||||
published_at=r.published_at,
|
||||
relevance_score=r.relevance_score,
|
||||
source_type=r.source_type,
|
||||
url=r.url,
|
||||
recUuid=r.recUuid,
|
||||
))
|
||||
new_count += 1
|
||||
|
||||
_emit_skill_step({
|
||||
"type": "skill_result",
|
||||
"message": f"技能「{sk.name}」检索到 {new_count} 条新素材",
|
||||
"skill_name": sk.name,
|
||||
"matched_count": new_count,
|
||||
})
|
||||
|
||||
# 早停:累积素材已达上限
|
||||
if len(raw_materials_pre) >= max_materials:
|
||||
stopped_early = True
|
||||
_emit_skill_step({
|
||||
"type": "stop_early",
|
||||
"message": f"已获取 {len(raw_materials_pre)} 条素材,达到上限,停止调用后续技能",
|
||||
"matched_count": len(raw_materials_pre),
|
||||
})
|
||||
break
|
||||
|
||||
# 技能模式不再走下面的并发/合并逻辑,all_results_nested 留空、跳过 web 通用搜索
|
||||
all_results_nested: list = []
|
||||
# 把已收集到的素材挂在 raw_materials_pre,下方合并阶段直接挪过去
|
||||
elif material_source == "notebook" and notebook_search_space_id:
|
||||
if material_source == "notebook" and notebook_search_space_id:
|
||||
# ── 纯笔记模式 ──
|
||||
# 用户在表单选了「我的空间」作为素材来源 → 跳过 web/intranet 检索,
|
||||
# 用户在表单选了「我的空间」作为检索类型 → 跳过技能编排,
|
||||
# 仅使用 Step 1.5 已拉取的 notebook_materials。
|
||||
# 与 UI 文案"从空间中检索素材"对齐。
|
||||
_emit(f"仅使用笔记素材,跳过 web/intranet 检索({len(notebook_materials)} 条)")
|
||||
all_results_nested = []
|
||||
else:
|
||||
# ── 常规 web/intranet 模式 ──
|
||||
_emit(f"仅使用笔记素材,跳过技能检索({len(notebook_materials)} 条)")
|
||||
elif is_append_only:
|
||||
# ── 兼容旧版追加关键词模式 ──
|
||||
# 直接对用户给的关键词做原生并发检索(不走技能编排),结果追加到旧素材。
|
||||
provider = get_search_provider(ai_writing_config)
|
||||
|
||||
# 按用户指定的关键词数量截断。
|
||||
# 用户输入 / 追加模式:keywords 本身就是用户填的,全部检索不截断。
|
||||
if is_user_query_search or is_append_only:
|
||||
search_keywords = keywords
|
||||
else:
|
||||
search_limit = keyword_count
|
||||
if extra_keywords:
|
||||
search_limit = max(search_limit, len(keywords))
|
||||
search_keywords = keywords[:search_limit]
|
||||
total = len(search_keywords)
|
||||
per_keyword = _per_keyword_limit(max_materials, total)
|
||||
_emit(f"检索词已生成,开始检索 {total} 个方向…")
|
||||
|
||||
completed = 0
|
||||
|
||||
async def _search_one(kw: str) -> list[SearchResult]:
|
||||
nonlocal completed
|
||||
try:
|
||||
return await provider.search(kw, max_results=per_keyword)
|
||||
finally:
|
||||
completed += 1
|
||||
if completed < total:
|
||||
_emit(f"已完成 {completed}/{total} 次检索,正在继续检索相关内容…")
|
||||
else:
|
||||
_emit(f"已完成全部 {total} 次检索")
|
||||
|
||||
search_tasks = [_search_one(kw) for kw in search_keywords]
|
||||
all_results_nested = await asyncio.gather(*search_tasks, return_exceptions=True)
|
||||
# provider 内部吞掉的检索异常(如内网 ES 连接失败),一并取出用于提示用户
|
||||
search_errors.extend(getattr(provider, "errors", []))
|
||||
|
||||
# 合并去重
|
||||
# - 首次检索 / 普通 re_search(不带 extra_keywords,用户只想 LLM 重出):旧素材丢弃
|
||||
# - 素材不足求助 supplement_search:保留旧素材 + 追加新检索结果
|
||||
# - PAUSE-1 追加检索 is_append_only:保留旧素材 + 仅检索新关键词
|
||||
# - 技能检索 is_skill_search:覆盖式替换,旧素材完全丢弃(编排循环已经把素材累积到 raw_materials_pre)
|
||||
existing_pkg = state.get("material_package")
|
||||
if is_skill_search:
|
||||
# 技能编排循环已经完成素材累积,直接用 raw_materials_pre 当 raw_materials
|
||||
# 笔记素材追加在技能素材之后
|
||||
raw_materials: list[Material] = raw_materials_pre # type: ignore[name-defined]
|
||||
raw_materials.extend(notebook_materials)
|
||||
else:
|
||||
seen_urls: set[str] = set()
|
||||
# 笔记素材优先排在最前面,赋予高相关度
|
||||
raw_materials = list(notebook_materials)
|
||||
for m in notebook_materials:
|
||||
if m.get("url"):
|
||||
seen_urls.add(m["url"])
|
||||
|
||||
preserve_existing = existing_pkg and (is_supplement or is_append_only)
|
||||
if preserve_existing:
|
||||
for m in existing_pkg.get("materials", []):
|
||||
seen_urls.add(m.get("url", ""))
|
||||
raw_materials.append(m)
|
||||
|
||||
for results in all_results_nested:
|
||||
per_keyword = _per_keyword_limit(max_materials, len(keywords))
|
||||
_emit(f"开始追加检索 {len(keywords)} 个方向…")
|
||||
gathered = await asyncio.gather(
|
||||
*[provider.search(kw, max_results=per_keyword) for kw in keywords],
|
||||
return_exceptions=True,
|
||||
)
|
||||
for results in gathered:
|
||||
if isinstance(results, Exception):
|
||||
logger.warning("search failed: %s", results)
|
||||
search_errors.append(f"{type(results).__name__}: {results}")
|
||||
continue
|
||||
for r in results:
|
||||
if r.url and r.url in seen_urls:
|
||||
new_results.extend(results or [])
|
||||
search_errors.extend(getattr(provider, "errors", []))
|
||||
else:
|
||||
# ── 通用检索:配置技能编排 ──
|
||||
configured = _load_configured_skills()
|
||||
provider = get_search_provider(ai_writing_config)
|
||||
per_skill = max(1, max_materials)
|
||||
|
||||
ordered = configured
|
||||
if is_user_query_search and len(configured) > 1:
|
||||
# 用户输入了自然语言:让模型在配置技能中按相关度匹配排序
|
||||
# (技能选择走流式 LLM,推 skill_select_chunk,前端打字机展示)
|
||||
picker = SkillPicker(
|
||||
config=config,
|
||||
model_name=keyword_model,
|
||||
emit_step=_emit_skill_step,
|
||||
max_select=5,
|
||||
user_query=user_query,
|
||||
candidates=configured,
|
||||
)
|
||||
ranked = await picker.pick(user_intent=intent)
|
||||
search_errors.extend(picker.errors)
|
||||
ordered = [entry["skill"] for entry in ranked] or configured
|
||||
else:
|
||||
_emit_skill_step({
|
||||
"type": "list_skills",
|
||||
"message": f"素材收集专家已配置 {len(configured)} 个检索技能",
|
||||
"skill_names": [s.name for s in configured],
|
||||
})
|
||||
|
||||
seen_urls_new: set[str] = set()
|
||||
|
||||
def _collect(results: list[SearchResult]) -> int:
|
||||
added = 0
|
||||
for r in results or []:
|
||||
if r.url and r.url in seen_urls_new:
|
||||
continue
|
||||
if r.url:
|
||||
seen_urls.add(r.url)
|
||||
raw_materials.append(Material(
|
||||
id=f"m{len(raw_materials)+1:03d}",
|
||||
title=r.title,
|
||||
content=r.content[:500],
|
||||
source=r.source,
|
||||
published_at=r.published_at,
|
||||
relevance_score=r.relevance_score,
|
||||
source_type=r.source_type,
|
||||
url=r.url,
|
||||
recUuid=r.recUuid,
|
||||
))
|
||||
seen_urls_new.add(r.url)
|
||||
new_results.append(r)
|
||||
added += 1
|
||||
return added
|
||||
|
||||
# Step 3: 相关度排序(如素材较多,让 LLM 排序)
|
||||
_emit("检索完成,正在筛选排序素材…")
|
||||
async def _timed_search(skill_name: str, query: str, limit: int) -> list[SearchResult]:
|
||||
# 加超时兜底:单个技能检索最多等 _SKILL_SEARCH_TIMEOUT_SECONDS 秒,
|
||||
# 超时不再阻塞后续技能,避免整段过程看起来"冻结"。
|
||||
try:
|
||||
return await asyncio.wait_for(
|
||||
provider.search(query, max_results=limit),
|
||||
timeout=_SKILL_SEARCH_TIMEOUT_SECONDS,
|
||||
) or []
|
||||
except asyncio.TimeoutError:
|
||||
logger.warning("skill[%s] search timed out after %ss", skill_name, _SKILL_SEARCH_TIMEOUT_SECONDS)
|
||||
search_errors.append(f"skill[{skill_name}]: 检索超时(>{_SKILL_SEARCH_TIMEOUT_SECONDS}s)")
|
||||
return []
|
||||
except Exception as e:
|
||||
logger.warning("skill[%s] search failed: %s", skill_name, e)
|
||||
search_errors.append(f"skill[{skill_name}]: {e}")
|
||||
return []
|
||||
|
||||
for sk in ordered:
|
||||
is_knowledge = sk.name == KNOWLEDGE_SEARCH_SKILL_NAME
|
||||
if is_knowledge and not is_user_query_search:
|
||||
# 原生知识库检索:对全部检索词并发检索(旧版默认路径)。
|
||||
# 用户带 extra_keywords 时全部检索,否则按 keyword_count 截断。
|
||||
search_keywords = keywords if extra_keywords else keywords[:keyword_count]
|
||||
total = len(search_keywords)
|
||||
per_keyword = _per_keyword_limit(max_materials, total)
|
||||
_emit(f"检索词已生成,开始检索 {total} 个方向…")
|
||||
_emit_skill_step({
|
||||
"type": "invoke_skill",
|
||||
"message": f"使用技能「{_KNOWLEDGE_SKILL_DISPLAY}」检索 {total} 个方向:{('、'.join(search_keywords[:3]))[:60]}",
|
||||
"skill_name": sk.name,
|
||||
})
|
||||
completed = 0
|
||||
|
||||
async def _search_one(kw: str) -> list[SearchResult]:
|
||||
nonlocal completed
|
||||
try:
|
||||
return await provider.search(kw, max_results=per_keyword)
|
||||
finally:
|
||||
completed += 1
|
||||
if completed < total:
|
||||
_emit(f"已完成 {completed}/{total} 次检索,正在继续检索相关内容…")
|
||||
else:
|
||||
_emit(f"已完成全部 {total} 次检索")
|
||||
|
||||
gathered = await asyncio.gather(*[_search_one(kw) for kw in search_keywords], return_exceptions=True)
|
||||
added = 0
|
||||
for results in gathered:
|
||||
if isinstance(results, Exception):
|
||||
logger.warning("search failed: %s", results)
|
||||
search_errors.append(f"{type(results).__name__}: {results}")
|
||||
continue
|
||||
added += _collect(results)
|
||||
search_errors.extend(getattr(provider, "errors", []))
|
||||
else:
|
||||
# 定向技能检索:user_query 模式优先用用户原话;非知识库技能再
|
||||
# 附上 description 让 query 贴合该技能的领域。
|
||||
if is_user_query_search:
|
||||
skill_query = user_query if is_knowledge else f"{user_query} {sk.description or ''}".strip()
|
||||
else:
|
||||
skill_query = f"{intent} {sk.description or ''}".strip() or intent
|
||||
_emit_skill_step({
|
||||
"type": "invoke_skill",
|
||||
"message": f"使用技能「{sk.name}」检索:{skill_query[:60]}",
|
||||
"skill_name": sk.name,
|
||||
})
|
||||
added = _collect(await _timed_search(sk.name, skill_query, per_skill))
|
||||
|
||||
_emit_skill_step({
|
||||
"type": "skill_result",
|
||||
"message": f"技能「{_KNOWLEDGE_SKILL_DISPLAY if is_knowledge else sk.name}」检索到 {added} 条新素材",
|
||||
"skill_name": sk.name,
|
||||
"matched_count": added,
|
||||
})
|
||||
|
||||
# 早停:累积素材已达上限
|
||||
if len(new_results) >= max_materials:
|
||||
_emit_skill_step({
|
||||
"type": "stop_early",
|
||||
"message": f"已获取 {len(new_results)} 条素材,达到上限,停止调用后续技能",
|
||||
"matched_count": len(new_results),
|
||||
})
|
||||
break
|
||||
|
||||
# 合并去重
|
||||
# - 首次检索 / 普通 re_search(无输入,用户只想重出):旧素材丢弃,覆盖式替换
|
||||
# - 素材不足求助 supplement_search:保留旧素材 + 追加新检索结果
|
||||
# - PAUSE-1 输入框 is_user_query_search:保留旧素材 + **追加**新素材(核心交互)
|
||||
# - 旧版追加关键词 is_append_only:保留旧素材 + 仅检索新关键词
|
||||
existing_pkg = state.get("material_package")
|
||||
preserve_existing = bool(existing_pkg) and (is_supplement or is_append_only or is_user_query_search)
|
||||
|
||||
seen_urls: set[str] = set()
|
||||
# 笔记素材优先排在最前面,赋予高相关度
|
||||
raw_materials: list[Material] = list(notebook_materials)
|
||||
for m in notebook_materials:
|
||||
if m.get("url"):
|
||||
seen_urls.add(m["url"])
|
||||
|
||||
if preserve_existing:
|
||||
for m in existing_pkg.get("materials", []):
|
||||
seen_urls.add(m.get("url", ""))
|
||||
raw_materials.append(m)
|
||||
|
||||
appended_count = 0
|
||||
for r in new_results:
|
||||
if r.url and r.url in seen_urls:
|
||||
continue
|
||||
if r.url:
|
||||
seen_urls.add(r.url)
|
||||
raw_materials.append(Material(
|
||||
id=f"m{len(raw_materials)+1:03d}",
|
||||
title=r.title,
|
||||
content=r.content[:500],
|
||||
source=r.source,
|
||||
published_at=r.published_at,
|
||||
relevance_score=r.relevance_score,
|
||||
source_type=r.source_type,
|
||||
url=r.url,
|
||||
recUuid=r.recUuid,
|
||||
))
|
||||
appended_count += 1
|
||||
|
||||
# Step 3: 相关度排序
|
||||
# 用户主动追加(输入框 / 追加关键词):保持旧素材原有顺序,新素材追加在
|
||||
# 末尾,不调 LLM 重排、不按上限丢弃旧素材(只做 100 条硬上限防爆)。
|
||||
rank_error = ""
|
||||
if len(raw_materials) > max_materials:
|
||||
is_user_append = is_user_query_search or is_append_only
|
||||
if is_user_append:
|
||||
materials = raw_materials[:_MAX_MAX_MATERIALS]
|
||||
dropped = len(raw_materials) - len(materials)
|
||||
summary = f"追加检索到 {appended_count} 条新素材,当前共 {len(materials)} 条"
|
||||
if dropped > 0:
|
||||
summary += f"(超出 {_MAX_MAX_MATERIALS} 条硬上限,丢弃末尾 {dropped} 条)"
|
||||
elif len(raw_materials) > max_materials:
|
||||
_emit("检索完成,正在筛选排序素材…")
|
||||
materials_list = "\n".join(
|
||||
f"- ID: {m['id']} 标题: {m['title']} 摘要: {m['content'][:100]}"
|
||||
for m in raw_materials
|
||||
@@ -608,26 +682,29 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
materials = sorted_materials[:max_materials]
|
||||
summary = rank_result.get("summary", f"共搜索到{len(raw_materials)}条素材,精选{len(materials)}条")
|
||||
else:
|
||||
_emit("检索完成,正在筛选排序素材…")
|
||||
materials = sorted(raw_materials, key=lambda m: m.get("relevance_score", 0), reverse=True)[:max_materials]
|
||||
summary = f"共搜索到{len(raw_materials)}条素材,精选{len(materials)}条"
|
||||
|
||||
# 空结果兜底:未检索到任何素材时给出明确文案,前端据此弹出「无数据继续」选项
|
||||
if not materials:
|
||||
summary = "未检索到相关素材,可重新搜索或在无素材的情况下继续写作"
|
||||
elif is_user_append and appended_count == 0:
|
||||
summary = f"本次未检索到新素材(已有 {len(materials)} 条保持不变),可换个说法再试"
|
||||
|
||||
# 追加模式下,pkg.keywords 体现到目前为止累积的检索词(旧 + 新追加),
|
||||
# 让前端列表 / 历史回显能看到完整的搜索路径。
|
||||
pkg_keywords = keywords
|
||||
if is_append_only and existing_pkg:
|
||||
if is_user_append and existing_pkg:
|
||||
old_kws = list(existing_pkg.get("keywords") or [])
|
||||
pkg_keywords = list(dict.fromkeys(old_kws + keywords))
|
||||
pkg = MaterialPackage(keywords=pkg_keywords, materials=materials, summary=summary)
|
||||
# 技能检索时把过程步骤一并放进 pkg,前端在素材确认卡里能展示"使用了哪些技能"
|
||||
if is_skill_search:
|
||||
# 技能编排过程步骤放进 pkg,前端在素材确认卡里能展示"使用了哪些技能"
|
||||
if skill_steps:
|
||||
pkg["skill_search_used"] = True
|
||||
pkg["skill_search_steps"] = skill_steps
|
||||
elif existing_pkg and existing_pkg.get("skill_search_used") and (is_supplement or is_append_only):
|
||||
# 保留之前一次技能检索的痕迹(不污染纯 web 检索)
|
||||
elif existing_pkg and existing_pkg.get("skill_search_used") and preserve_existing:
|
||||
# 保留之前一次技能检索的痕迹
|
||||
pkg["skill_search_used"] = existing_pkg.get("skill_search_used", False)
|
||||
pkg["skill_search_steps"] = existing_pkg.get("skill_search_steps", [])
|
||||
|
||||
@@ -640,9 +717,7 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
if search_errors:
|
||||
uniq = list(dict.fromkeys(search_errors))
|
||||
# 解析失败时收集到的接口原始响应体,前端折叠展示便于定位
|
||||
# (技能模式下 provider 局部变量没暴露,所以 raw_dumps 兜底为空列表)
|
||||
provider_obj = locals().get("provider") or locals().get("web_provider")
|
||||
raw_dumps = list(dict.fromkeys(getattr(provider_obj, "raw_responses", []) or []))
|
||||
raw_dumps = list(dict.fromkeys(getattr(provider, "raw_responses", []) or [])) if provider is not None else []
|
||||
notices.append(_notice(
|
||||
"ES 波动导致检索异常",
|
||||
";".join(uniq[:3]),
|
||||
@@ -650,13 +725,13 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
|
||||
agent_name=researcher_label,
|
||||
))
|
||||
|
||||
# 进度事件文案:用户输入 / 追加模式 / 技能模式下让用户看到差异化提示
|
||||
if is_skill_search:
|
||||
step_start_msg = f"启用技能检索({len(skill_steps)} 步)"
|
||||
elif is_user_query_search:
|
||||
step_start_msg = f"按你的输入检索:{user_query[:40]}"
|
||||
# 进度事件文案:用户输入 / 追加模式下让用户看到差异化提示
|
||||
if is_user_query_search:
|
||||
step_start_msg = f"按你的输入匹配技能检索并追加:{user_query[:40]}"
|
||||
elif is_append_only:
|
||||
step_start_msg = f"追加检索关键词:{', '.join(keywords[:3])}"
|
||||
elif skill_steps:
|
||||
step_start_msg = f"通用检索(技能编排 {len(skill_steps)} 步):{', '.join(keywords[:3])}"
|
||||
else:
|
||||
step_start_msg = f"搜索关键词:{', '.join(keywords[:3])}"
|
||||
return {
|
||||
|
||||
Reference in new issue
Block a user