作家写作

This commit is contained in:
fengxiangdebaozi committed 2026-05-29 11:42:30 +08:00
1 parent 1a122eb7fd
commit 07ad87365f
7 files changed
+380 -228

No files matched your search

@@ -9,7 +9,8 @@ from typing import List
from langchain_core.runnables import RunnableConfig
from ..search import get_search_provider, SearchResult, SkillSearchProvider
from ..search import get_search_provider, SearchResult
from ..search.skill_search import SkillPicker
from ..state import AIWritingState, Material, MaterialPackage
from ..prompts.researcher_prompts import (
EXTRACT_KEYWORDS_SYSTEM, EXTRACT_KEYWORDS_USER,
@@ -135,10 +136,13 @@ 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 / 追加关键词"分支时跳过 LLM 调用,notice 拼接需要兜底值
if is_user_query_search:
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:
# 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]
keywords = [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
@@ -164,57 +168,115 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
keywords = list(dict.fromkeys(keywords + extra_keywords))
# 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
# 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
skill_steps: list[dict] = []
search_errors: list[str] = []
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)
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(
# ── 技能编排检索 ──
picker = SkillPicker(
llm_select=_llm_select,
emit_step=_emit_skill_step,
max_select=max(1, min(int(max_materials), 10)),
max_select=5,
user_query=user_query,
)
selected = await picker.pick(user_intent=intent)
search_errors.extend(picker.errors)
# 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"]
# 用 LLM 把 user_query + skill.description 转成针对性关键词
_emit_skill_step({
"type": "invoke_skill",
"message": f"使用技能「{sk.name}」开始检索…",
"skill_name": sk.name,
})
skill_query = f"{user_query} {sk.description or ''}".strip() or intent
try:
res = await web_provider.search(skill_query, max_results=per_skill)
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
else:
# ── 常规 web/intranet 模式 ──
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。
# 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_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)
# 每个检索词分到的检索条数 = 总量上限 / 检索词数(向上取整)
per_keyword = _per_keyword_limit(max_materials, total)
_emit(f"检索词已生成,开始检索 {total} 个方向…")
# 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…")
# 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]:
@@ -230,59 +292,48 @@ async def researcher_node(state: AIWritingState, config: RunnableConfig) -> dict
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(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] = []
# - 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")
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", ""))
raw_materials.append(m)
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
raw_materials: list[Material] = raw_materials_pre # type: ignore[name-defined]
else:
seen_urls: set[str] = set()
raw_materials = []
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)
search_errors: list[str] = []
for results in all_results_nested:
if isinstance(results, Exception):
logger.warning("search failed: %s", results)
search_errors.append(f"{type(results).__name__}: {results}")
continue
for r in results:
# 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:
for results in all_results_nested:
if isinstance(results, Exception):
logger.warning("search failed: %s", results)
search_errors.append(f"{type(results).__name__}: {results}")
continue
else:
seen_urls.add(r.url)
mat: Material = 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,
)
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", []))
for r in 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,
))
# 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…")
@@ -343,7 +394,9 @@ 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
raw_dumps = list(dict.fromkeys(getattr(provider, "raw_responses", []) or []))
# (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 []))
notices.append(_notice(
"ES 波动导致检索异常",
";".join(uniq[:3]),