代码修改

This commit is contained in:
fengxiangdebaozi committed 2026-05-29 11:19:58 +08:00
1 parent de597b602a
commit 1a122eb7fd
20 files changed
+812 -140

No files matched your search

@@ -9,7 +9,7 @@ from typing import List
from langchain_core.runnables import RunnableConfig
from ..search import get_search_provider, SearchResult
from ..search import get_search_provider, SearchResult, SkillSearchProvider
from ..state import AIWritingState, Material, MaterialPackage
from ..prompts.researcher_prompts import (
EXTRACT_KEYWORDS_SYSTEM, EXTRACT_KEYWORDS_USER,
@@ -96,15 +96,24 @@ 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
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「追加检索」并填了新关键词:
# PAUSE-1 素材确认卡用户点「按输入重新搜索」并填了 user_query:
# - 跳过 LLM 重新生成检索词(节省一次调用 + 尊重用户意图)
# - 只检索 extra_keywords 这几个新词,不重检索之前 LLM 出过的词
# - 直接用 user_query 作为唯一检索词
# - 保留 existing material_package,按 URL 去重合并
is_append_only = action == "re_search" and bool(extra_keywords)
is_user_query_search = action == "re_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)
# PAUSE-1 素材确认卡用户点「启用技能检索」:
# - 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"
# 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"))
@@ -126,10 +135,14 @@ 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 = {} # 走追加分支时跳过 LLM 调用,但下方 notice 拼接需要兜底值
if is_append_only:
# 追加模式:直接用用户填的关键词,不再调 LLM
keywords: list[str] = _clean_keywords(extra_keywords, intent)
kw_result: dict = {} # 走"用户输入 / 追加关键词"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: list[str] = [user_query]
_emit(f"按你的输入「{user_query[:30]}」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)
_emit(f"追加检索 {len(keywords)} 个关键词…")
else:
_emit("正在生成检索词…")
@@ -150,13 +163,31 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
if extra_keywords:
keywords = list(dict.fromkeys(keywords + extra_keywords))
# Step 2: 并发搜索
provider = get_search_provider(ai_writing_config)
# Step 2: 搜索
# 技能检索:用 SkillSearchProvider 替代 web/intranet;步骤事件单独走 skill_search_step 通道
skill_steps: list[dict] = []
if is_skill_search:
def _emit_skill_step(step: dict) -> None:
skill_steps.append(step)
writer({"type": "skill_search_step", "agent_name": "素材收集专家", **step})
async def _llm_select(system: str, user_prompt: str, fallback: dict) -> dict:
return await llm_json(system, user_prompt, fallback=fallback, config=config, model_name=keyword_model)
provider = SkillSearchProvider(
llm_select=_llm_select,
emit_step=_emit_skill_step,
max_select=max(1, min(int(max_materials), 10)),
user_query=user_query,
)
else:
provider = get_search_provider(ai_writing_config)
# 按用户指定的关键词数量截断。
# 追加模式:keywords 本身就是用户填的,全部检索不截断(用户填 N 个就搜 N 个)。
# 用户输入 / 追加模式:keywords 本身就是用户填的,全部检索不截断。
# 其他模式:受 keyword_count 限制,但 extra_keywords 在内时上限放宽到 keywords 总数。
if is_append_only:
if is_user_query_search or is_append_only:
search_keywords = keywords
else:
search_limit = keyword_count
@@ -168,31 +199,48 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
per_keyword = _per_keyword_limit(max_materials, total)
_emit(f"检索词已生成,开始检索 {total} 个方向…")
# 并发搜索,但每完成一个就推送进度(计数在单线程 asyncio 下无需加锁)
completed = 0
# 技能模式:串行调用 provider(早停语义靠 provider 内部累积计数)
# 其它模式:并发调用,每完成一个就推送进度
if is_skill_search:
all_results_nested: list[list[SearchResult]] = []
for kw in search_keywords:
try:
res = await provider.search(kw, max_results=per_keyword, user_intent=intent)
except Exception as e:
logger.warning("skill search failed: %s", e)
all_results_nested.append(e) # type: ignore[arg-type]
continue
all_results_nested.append(res)
# provider 内部 early_stop_at 已触发时不会再返回新 skill,提前 break 避免空轮询
if not res:
break
else:
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")
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: list[list[SearchResult]] = await asyncio.gather(*search_tasks, return_exceptions=True)
search_tasks = [_search_one(kw) for kw in search_keywords]
all_results_nested = await asyncio.gather(*search_tasks, return_exceptions=True)
# 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(用户期望对比 web vs skill 结果)
seen_urls: set[str] = set()
seen_skill_ids: set[str] = set() # skill 素材无 URL,按 skill_name 去重
raw_materials: list[Material] = []
existing_pkg = state.get("material_package")
preserve_existing = existing_pkg and (is_supplement or is_append_only)
preserve_existing = existing_pkg and (is_supplement or is_append_only) and not is_skill_search
if preserve_existing:
for m in existing_pkg.get("materials", []):
seen_urls.add(m.get("url", ""))
@@ -205,10 +253,19 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
search_errors.append(f"{type(results).__name__}: {results}")
continue
for r in results:
if r.url and r.url in seen_urls:
# skill 素材按 skill_name 去重(同一个 skill 被多个关键词命中只保留一次)
if r.source_type == "skill":
meta = getattr(r, "extra_metadata", {}) or {}
sk_name = meta.get("skill_name", "")
if sk_name and sk_name in seen_skill_ids:
continue
if sk_name:
seen_skill_ids.add(sk_name)
elif r.url and r.url in seen_urls:
continue
seen_urls.add(r.url)
raw_materials.append(Material(
else:
seen_urls.add(r.url)
mat: Material = Material(
id=f"m{len(raw_materials)+1:03d}",
title=r.title,
content=r.content[:500],
@@ -218,7 +275,11 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
source_type=r.source_type,
url=r.url,
recUuid=r.recUuid,
))
)
extra = getattr(r, "extra_metadata", None)
if extra:
mat["extra_metadata"] = extra
raw_materials.append(mat)
# 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", []))
@@ -264,6 +325,14 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
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:
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)
pkg["skill_search_used"] = existing_pkg.get("skill_search_used", False)
pkg["skill_search_steps"] = existing_pkg.get("skill_search_steps", [])
# 容错告警:检索接口 / 模型调用出错时以非致命 notice 提示用户,不打断写作流程
notices: list[dict] = []
@@ -281,12 +350,15 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
raw="\n\n----------\n\n".join(raw_dumps[:3]),
))
# 进度事件文案:追加模式下让用户看到只搜了哪几个新词
step_start_msg = (
f"追加检索关键词:{', '.join(keywords[:3])}"
if is_append_only
else f"搜索关键词:{', '.join(keywords[:3])}"
)
# 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]}"
elif is_append_only:
step_start_msg = f"追加检索关键词:{', '.join(keywords[:3])}"
else:
step_start_msg = f"搜索关键词:{', '.join(keywords[:3])}"
return {
"material_package": pkg,
# 补充检索:直接回写作(沿用原大纲);常规检索:进入素材确认暂停点