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

@@ -3,6 +3,7 @@ from .internet_search import InternetSearchProvider
from .intranet_stub import IntranetSearchProvider
from .demo_search import DemoSearchProvider
from .duckduckgo_search import DuckDuckGoSearchProvider
from .skill_search import SkillSearchProvider
def get_search_provider(ai_writing_config: dict) -> BaseSearchProvider:
@@ -60,5 +61,6 @@ __all__ = [
"BaseSearchProvider", "SearchResult",
"InternetSearchProvider", "IntranetSearchProvider",
"DemoSearchProvider", "DuckDuckGoSearchProvider",
"SkillSearchProvider",
"get_search_provider",
]
@@ -0,0 +1,248 @@
"""技能检索 Provider:把已启用的 Skill 当作可检索的素材源。
流程(保持简单,对应素材确认卡上的"技能检索"按钮):
1) list_skills 事件 —— 列出当前启用的全部 skill
2) select_skills 事件 —— LLM 根据用户输入 + 写作意图按相关度选出若干个
3) invoke_skill 事件 —— 按相关度从高到低顺序产出 SearchResult
输出格式严格对齐 BaseSearchProvider 的 SearchResult,让 researcher 节点
不需要分支处理,下游 Material 字段同构。
"""
from __future__ import annotations
import logging
import re
from typing import Callable, List, Optional
from .base import BaseSearchProvider, SearchResult
logger = logging.getLogger(__name__)
_SELECT_SKILLS_SYSTEM = """你是一名情报研究员,需要从可用「技能」中挑出对写作任务有帮助的若干个,并按相关度从高到低排序。
技能是一段封装好的指令 / 知识 / 工作流,每个技能有 name 和 description。
【严禁思考】
- 这是一个简单的匹配任务,禁止任何思考、分析、推理过程,禁止输出 <think></think> 标签。
- 收到输入后立即直接输出最终 JSON,不得有任何前置文字。
【输出要求】
- 只输出一个 JSON 对象,不要 markdown 代码块。
- 必须使用英文双引号。
- selected_skills 中的 name 必须严格来自下方"可用技能"列表,不要编造,按相关度从高到低排序。
- 如果用户输入里明确点名了某个技能(按 name 完全匹配),把它排在第一位。
【完整示例】
{"selected_skills": ["web-search", "knowledge-lookup"], "reason": "用户想检索互联网信息,web-search 最相关,知识库查询作为补充"}"""
_SELECT_SKILLS_USER = """写作意图:{user_intent}
用户补充输入:{user_query}
检索关键词:{keywords}
可用技能(共 {skill_count} 个):
{skill_list}
请从中按相关度从高到低选出最多 {max_select} 个技能。如果没有合适的技能,selected_skills 返回空数组。
只输出 JSON:
{{"selected_skills": ["name1", "name2"], "reason": "为什么选这些 + 排序依据"}}
/no_think"""
EmitStepFn = Callable[[dict], None]
def _summarize_skill_content(content: str, limit: int = 500) -> str:
"""从 SKILL.md 提取摘要:跳过 frontmatter,取正文开头 limit 字。"""
if not content:
return ""
body = content
m = re.match(r"^---\s*\n(.*?\n)---\s*\n(.*)$", content, flags=re.DOTALL)
if m:
body = m.group(2)
body = re.sub(r"^#\s+.*?\n", "", body.lstrip(), count=1)
body = body.strip()
if len(body) > limit:
return body[:limit].rstrip() + "…"
return body
class SkillSearchProvider(BaseSearchProvider):
"""从启用的 Skills 中筛选并返回作为素材。
构造参数:
llm_select: 异步 LLM 调用 (system, user, fallback) -> dict。
emit_step: 可选;推送时间线步骤事件(list_skills / select_skills / invoke_skill)。
max_select: LLM 单次最多挑几个技能。
user_query: 用户在素材确认卡输入框写的自然语言(可能包含技能名 / 来源 / 补充意图)。
"""
def __init__(
self,
llm_select: Callable,
*,
emit_step: Optional[EmitStepFn] = None,
max_select: int = 5,
user_query: str = "",
) -> None:
self._llm_select = llm_select
self._emit_step = emit_step or (lambda _step: None)
self._max_select = max_select
self._user_query = (user_query or "").strip()
self._errors: list[str] = []
# 单次 researcher 调用内复用结果,避免每个 keyword 都问一次 LLM
self._selected_cache: list[dict] | None = None
self._enabled_skills: list | None = None
# 跨 keyword 调用时同一个 skill 只暴露一次
self._yielded_skills: set[str] = set()
@property
def errors(self) -> list[str]:
return list(self._errors)
def _load_enabled_skills(self) -> list:
if self._enabled_skills is not None:
return self._enabled_skills
try:
from deerflow.skills import get_or_new_skill_storage
storage = get_or_new_skill_storage()
skills = storage.load_skills(enabled_only=True)
except Exception as e:
logger.exception("ai_writing.skill_search: 加载启用技能失败")
self._errors.append(f"skill list error: {e}")
skills = []
self._enabled_skills = skills
self._emit_step({
"type": "list_skills",
"message": f"发现 {len(skills)} 个可用技能",
"skill_names": [s.name for s in skills],
})
return skills
async def _select_skills(self, query: str, user_intent: str) -> list[dict]:
"""让 LLM 从启用的技能中按相关度排序选出若干个。同一次检索复用结果。"""
if self._selected_cache is not None:
return self._selected_cache
skills = self._load_enabled_skills()
if not skills:
self._selected_cache = []
return []
skill_list_text = "\n".join(
f"- {s.name}: {s.description or '(无描述)'}" for s in skills
)
try:
result = await self._llm_select(
_SELECT_SKILLS_SYSTEM,
_SELECT_SKILLS_USER.format(
user_intent=user_intent or query,
user_query=self._user_query or "(无)",
keywords=query,
skill_count=len(skills),
skill_list=skill_list_text,
max_select=self._max_select,
),
fallback={"selected_skills": [s.name for s in skills[: self._max_select]]},
)
except Exception as e:
logger.exception("ai_writing.skill_search: LLM 选择技能失败")
self._errors.append(f"skill select error: {e}")
result = {"selected_skills": [s.name for s in skills[: self._max_select]]}
wanted_names = result.get("selected_skills") or []
if not isinstance(wanted_names, list):
wanted_names = []
by_name = {s.name: s for s in skills}
chosen: list[dict] = []
for name in wanted_names:
if not isinstance(name, str):
continue
sk = by_name.get(name.strip())
if sk:
chosen.append({"skill": sk, "reason": result.get("reason", "")})
if len(chosen) >= self._max_select:
break
self._emit_step({
"type": "select_skills",
"message": f"按相关度排序,候选 {len(chosen)} 个技能(将按顺序调用)",
"skill_names": [c["skill"].name for c in chosen],
"reason": result.get("reason", ""),
})
self._selected_cache = chosen
return chosen
@staticmethod
def _read_skill_content(skill) -> str:
try:
return skill.skill_file.read_text(encoding="utf-8")
except Exception as e:
logger.warning("ai_writing.skill_search: 读 SKILL.md 失败 %s: %s", skill.name, e)
return skill.description or ""
def _to_search_result(self, skill, reason: str) -> SearchResult:
raw = self._read_skill_content(skill)
summary = _summarize_skill_content(raw, limit=500)
category_value = skill.category.value if hasattr(skill.category, "value") else str(skill.category)
return SearchResult(
title=skill.name,
content=summary or (skill.description or skill.name),
source=f"技能库 / {category_value}",
published_at="",
url="",
relevance_score=1.0,
source_type="skill",
recUuid="",
extra_metadata={
"skill_name": skill.name,
"skill_description": skill.description or "",
"skill_path": skill.skill_path,
"container_path": skill.get_container_file_path(),
"category": category_value,
"select_reason": reason,
# 完整 SKILL.md 内容,供 writer_draft 节点按需注入到 prompt
"skill_full_content": raw,
},
)
async def search(
self,
query: str,
max_results: int = 5,
date_from: str = "",
date_to: str = "",
user_intent: str = "",
) -> List[SearchResult]:
"""按相关度顺序产出候选 skills 对应的 SearchResult。
说明:每个 skill 直接产出一条 SearchResult。researcher 节点上游会按
skill_name 去重,所以同一个 SkillSearchProvider 多次调用本方法时不会
产出重复 skill 素材(_yielded_skills 兜底)。
"""
chosen = await self._select_skills(query, user_intent=user_intent)
if not chosen:
return []
results: List[SearchResult] = []
for entry in chosen:
sk = entry["skill"]
if sk.name in self._yielded_skills:
continue
self._emit_step({
"type": "invoke_skill",
"message": f"调用技能:{sk.name}",
"skill_name": sk.name,
"matched_count": len(self._yielded_skills) + 1,
})
results.append(self._to_search_result(sk, entry.get("reason", "")))
self._yielded_skills.add(sk.name)
if len(results) >= max_results:
break
return results
async def health_check(self) -> bool:
try:
return bool(self._load_enabled_skills())
except Exception:
return False