Merge branch 'dhk-1' of http://192.168.0.200/knowledge-base-bj/deerflow-server into dhk-1
This commit is contained in:
commit
016f921cf0
34 files changed
+1921
-54
No files matched your search
@@ -209,6 +209,8 @@ _CONTEXT_CONFIGURABLE_KEYS: frozenset[str] = frozenset(
|
||||
"max_concurrent_subagents",
|
||||
"agent_name",
|
||||
"is_bootstrap",
|
||||
"is_writing_setup",
|
||||
"valid_article_types",
|
||||
"excluded_tools",
|
||||
"skill_stop_names",
|
||||
"skill_stop_message",
|
||||
|
||||
+120
-3
@@ -7,11 +7,19 @@ import logging
|
||||
from langchain_core.messages import HumanMessage, SystemMessage
|
||||
from langchain_core.runnables import RunnableConfig
|
||||
|
||||
from ..state import AIWritingState, KeyPoint, Outline, OutlineSection
|
||||
from ..outline_parser import parse_user_outline
|
||||
from ..prompts.writer_prompts import (
|
||||
OUTLINE_SYSTEM, OUTLINE_USER, OUTLINE_REVISE_SYSTEM, OUTLINE_REVISE_USER,
|
||||
OUTLINE_MATERIAL_ASSIGN_SYSTEM,
|
||||
OUTLINE_MATERIAL_ASSIGN_USER,
|
||||
OUTLINE_NORMALIZE_SYSTEM,
|
||||
OUTLINE_NORMALIZE_USER,
|
||||
OUTLINE_REVISE_SYSTEM,
|
||||
OUTLINE_REVISE_USER,
|
||||
OUTLINE_SYSTEM,
|
||||
OUTLINE_USER,
|
||||
)
|
||||
from ._utils import get_writing_model, parse_json, _content_to_text, reasoning_delta, ThinkTagParser, safe_get_stream_writer
|
||||
from ..state import AIWritingState, KeyPoint, Outline, OutlineSection
|
||||
from ._utils import ThinkTagParser, _content_to_text, get_writing_model, llm_json, parse_json, reasoning_delta, safe_get_stream_writer
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
@@ -80,6 +88,108 @@ def _parse_outline(raw: dict) -> Outline:
|
||||
return Outline(title=raw.get("title", "未命名文章"), sections=sections)
|
||||
|
||||
|
||||
def _derive_title(state: AIWritingState) -> str:
|
||||
"""用户大纲没给总标题时,用写作意图兜底一个标题。"""
|
||||
intent = (state.get("user_intent") or "").strip()
|
||||
if not intent:
|
||||
return "未命名文章"
|
||||
return intent[:30].rstrip() + ("…" if len(intent) > 30 else "")
|
||||
|
||||
|
||||
async def _normalize_user_outline_via_llm(state: AIWritingState, config: RunnableConfig, raw: str) -> Outline:
|
||||
"""确定性解析失败时,调 LLM 把用户大纲原文归一成三级结构(严格保真、不改写)。
|
||||
|
||||
再失败则把每个非空行当作一个章节,保证永远产出一份可用大纲、不丢用户内容。
|
||||
"""
|
||||
model_name = (state.get("model_name") or "").strip() or None
|
||||
line_fallback = {
|
||||
"title": "",
|
||||
"sections": [
|
||||
{"section_title": ln.strip(), "key_points": [], "material_ids": []}
|
||||
for ln in raw.splitlines() if ln.strip()
|
||||
] or [{"section_title": "正文", "key_points": [], "material_ids": []}],
|
||||
}
|
||||
result = await llm_json(
|
||||
OUTLINE_NORMALIZE_SYSTEM,
|
||||
OUTLINE_NORMALIZE_USER.format(raw_outline=raw),
|
||||
line_fallback,
|
||||
config,
|
||||
model_name=model_name,
|
||||
)
|
||||
return _parse_outline(result)
|
||||
|
||||
|
||||
async def _assign_materials_to_outline(
|
||||
state: AIWritingState, config: RunnableConfig, outline: Outline, pkg: dict
|
||||
) -> Outline:
|
||||
"""为用户大纲的各章节回填 material_ids(结构锁定:只取编号、按章节序号嫁接)。
|
||||
|
||||
LLM 只负责「哪个章节配哪些素材」,标题与层级一律以解析结果为准,杜绝改写。失败则
|
||||
保持 material_ids 为空(writer_draft 自带「取前 2 条素材」兜底)。
|
||||
"""
|
||||
sections = outline.get("sections", [])
|
||||
materials = pkg.get("materials") or []
|
||||
if not sections or not materials:
|
||||
return outline
|
||||
|
||||
valid_ids = {m["id"] for m in materials}
|
||||
sections_list = "\n".join(
|
||||
f"{i + 1}. {s.get('section_title', '')}" for i, s in enumerate(sections)
|
||||
)
|
||||
model_name = (state.get("model_name") or "").strip() or None
|
||||
result = await llm_json(
|
||||
OUTLINE_MATERIAL_ASSIGN_SYSTEM,
|
||||
OUTLINE_MATERIAL_ASSIGN_USER.format(
|
||||
user_intent=state.get("user_intent", ""),
|
||||
sections_list=sections_list,
|
||||
materials_summary=_materials_summary(pkg),
|
||||
),
|
||||
{"assignments": []},
|
||||
config,
|
||||
model_name=model_name,
|
||||
)
|
||||
|
||||
assignments = result.get("assignments")
|
||||
if not isinstance(assignments, list):
|
||||
return outline
|
||||
for i, section in enumerate(sections):
|
||||
if i >= len(assignments):
|
||||
break
|
||||
ids = assignments[i]
|
||||
if isinstance(ids, list):
|
||||
section["material_ids"] = [mid for mid in ids if mid in valid_ids]
|
||||
return outline
|
||||
|
||||
|
||||
async def _build_outline_from_user(
|
||||
state: AIWritingState, config: RunnableConfig, pkg: dict, raw: str
|
||||
) -> dict:
|
||||
"""用户自带大纲分支:解析用户大纲 → 关联素材 → 直接进确认,不调 LLM 重新规划。"""
|
||||
writer = safe_get_stream_writer()
|
||||
writer({"type": "outline_chunk", "text": "正在整理你提供的大纲…\n", "is_thinking": True})
|
||||
|
||||
outline = parse_user_outline(raw)
|
||||
if outline is None:
|
||||
# 确定性解析拿不准(散文化/格式混乱)→ LLM 兜底归一
|
||||
outline = await _normalize_user_outline_via_llm(state, config, raw)
|
||||
|
||||
if not outline.get("title"):
|
||||
outline = {**outline, "title": _derive_title(state)}
|
||||
|
||||
# 仍走了检索 → 把已检索素材关联到用户章节(结构锁定)
|
||||
outline = await _assign_materials_to_outline(state, config, outline, pkg)
|
||||
|
||||
section_count = len(outline.get("sections", []))
|
||||
return {
|
||||
"current_outline": outline,
|
||||
"status": "awaiting_outline_confirm",
|
||||
"progress_events": [
|
||||
{"type": "step_start", "agent_name": "作家", "message": "采用你提供的大纲"},
|
||||
{"type": "outline_ready", "agent_name": "作家", "message": f"已整理你的大纲,共 {section_count} 个章节"},
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
async def writer_outline_node(state: AIWritingState, config: RunnableConfig) -> dict:
|
||||
pkg = state.get("material_package") or {"keywords": [], "materials": [], "summary": ""}
|
||||
intervention = state.get("last_user_intervention") or {}
|
||||
@@ -95,6 +205,13 @@ async def writer_outline_node(state: AIWritingState, config: RunnableConfig) ->
|
||||
current_outline = state.get("current_outline")
|
||||
is_revise = bool(feedback) and bool(current_outline)
|
||||
|
||||
# 用户自带大纲:首轮(尚无大纲、非重规划)且表单粘了大纲原文 → 走「采用用户大纲」分支,
|
||||
# 跳过 LLM 重新规划。后续若用户在确认卡里点「重新规划」,current_outline 已存在 → 落到
|
||||
# 下方 is_revise 编辑模式,按反馈调整用户大纲(符合预期)。
|
||||
user_outline_raw = (state.get("user_outline_raw") or "").strip()
|
||||
if user_outline_raw and not current_outline and not is_revise:
|
||||
return await _build_outline_from_user(state, config, pkg, user_outline_raw)
|
||||
|
||||
if is_revise:
|
||||
system_prompt = OUTLINE_REVISE_SYSTEM
|
||||
user_prompt = OUTLINE_REVISE_USER.format(
|
||||
|
||||
+220
@@ -0,0 +1,220 @@
|
||||
"""把用户粘贴的大纲文本解析成三级 ``Outline`` 结构(确定性解析,不依赖 LLM)。
|
||||
|
||||
用户在写作表单里可能直接粘贴一份已有大纲,要求「完全按我的大纲生成」。本模块负责
|
||||
把各种常见的大纲文本格式严格、保真地转成后端三级结构:
|
||||
|
||||
Outline.sections[] = Level 1 章节(section_title)
|
||||
OutlineSection.key_points[] = Level 2 小节(point_title)
|
||||
KeyPoint.sub_points[] = Level 3 子要点
|
||||
|
||||
兼容的输入格式:
|
||||
- Markdown 标题:``#`` / ``##`` / ``###`` / ``####``(首个孤立 ``#`` 视为文章标题)
|
||||
- 中文编号:``一、`` / ``(一)`` / ``第一章`` / ``第一节``
|
||||
- 阿拉伯编号:``1.`` / ``1、`` / ``1.1`` / ``1.1.1``
|
||||
- 缩进 bullet:``-`` / ``*`` / ``•`` / ``·``(用缩进深度表示层级)
|
||||
- 纯文本平铺:每行一个一级标题(短标题、无句末标点)
|
||||
|
||||
设计原则:
|
||||
- **保真不改写**:只识别层级、原样保留标题文本,绝不增删或改写用户内容。
|
||||
- **多级混排自适应**:把文本里实际出现的各种「层级标记」收集起来排序,按相对外层
|
||||
顺序映射到 1/2/3 级——所以不依赖固定的标记搭配,``##``+``-`` 或 ``一、``+缩进
|
||||
都能正确分层。
|
||||
- **拿不准就交给上层**:识别不出可靠层级(如整段散文)时返回 ``None``,由调用方走
|
||||
LLM 兜底归一,避免硬把散文拆成错误结构。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
|
||||
from .state import KeyPoint, Outline, OutlineSection
|
||||
|
||||
# ── 行首标记识别 ──────────────────────────────────────────────────────────────
|
||||
|
||||
_CN_NUM = "一二三四五六七八九十百千零〇两"
|
||||
|
||||
# Markdown 标题:## 标题
|
||||
_MD_HEADING_RE = re.compile(r"^(#{1,6})\s+(.*\S)\s*$")
|
||||
# 第一章 / 第1节 / 第一部分
|
||||
_CN_CHAPTER_RE = re.compile(rf"^第\s*[{_CN_NUM}\d]+\s*[章节部分篇讲回卷](?:[\s::、..))]\s*|\s+)?(.*)$")
|
||||
# 一、 / 十二、 / 一. / 一.
|
||||
_CN_DUN_RE = re.compile(rf"^[{_CN_NUM}]+\s*[、..]\s*(.*)$")
|
||||
# (一) / (一) / (1) / (1)
|
||||
_PAREN_RE = re.compile(rf"^[((]\s*[{_CN_NUM}\d]+\s*[))]\s*(.*)$")
|
||||
# 1.1 / 1.1.1(带至少一个点的多级阿拉伯编号)
|
||||
_ARABIC_DOTTED_RE = re.compile(r"^(\d+(?:\.\d+)+)\s*[、..))]?\s*(.*)$")
|
||||
# 1. / 1、 / 1) / 1.(单级阿拉伯编号)
|
||||
_ARABIC_SINGLE_RE = re.compile(r"^\d+\s*[、..))]\s*(.*)$")
|
||||
# 带圈数字 ①②③ / 罗马 ⅰ
|
||||
_CIRCLED_RE = re.compile(r"^[①②③④⑤⑥⑦⑧⑨⑩⑪⑫⑬⑭⑮]\s*(.*)$")
|
||||
# bullet:- / * / • / ·(要求标记后有空白,避免误吞「*强调*」之类)
|
||||
_BULLET_RE = re.compile(r"^[-*•·]\s+(.*\S)\s*$")
|
||||
|
||||
# 标记分组的「外层度」基准值(越小越外层)。group: 0=标题 1=编号 2=缩进。
|
||||
# 同组内 order 决定相对层级;不同组之间靠 group 排序(标题 < 编号 < 缩进)。
|
||||
# 编号顺序遵循中文公文惯例:第一章 > 一、 > (一) > 1. > 1.1 > ①。
|
||||
_NUM_ORDER = {
|
||||
"cn_chapter": 0.0, # 第一章
|
||||
"cn_dun": 1.0, # 一、
|
||||
"paren": 2.0, # (一)
|
||||
"arabic_single": 3.0, # 1.
|
||||
"circled": 8.0, # ①
|
||||
}
|
||||
|
||||
|
||||
def _indent_width(line: str) -> int:
|
||||
"""计算行首缩进宽度(tab 记 4 列)。"""
|
||||
width = 0
|
||||
for ch in line:
|
||||
if ch == " ":
|
||||
width += 1
|
||||
elif ch == "\t":
|
||||
width += 4
|
||||
else:
|
||||
break
|
||||
return width
|
||||
|
||||
|
||||
def _classify(stripped: str, indent: int) -> tuple[tuple[int, float], str] | None:
|
||||
"""识别一行的层级标记,返回 ``((group, order), 纯文本)``;无法识别返回 None。
|
||||
|
||||
``(group, order)`` 是该行的「层级键」:把全文出现过的不同键排序后,按位置映射到
|
||||
1/2/3 级。这样不必预设标记搭配,相对顺序对了层级就对。
|
||||
"""
|
||||
if not stripped:
|
||||
return None
|
||||
|
||||
# 1) Markdown 标题(group 0,order = 井号数)
|
||||
m = _MD_HEADING_RE.match(stripped)
|
||||
if m:
|
||||
depth = len(m.group(1))
|
||||
text = m.group(2).strip()
|
||||
return (0, float(depth)), text
|
||||
|
||||
# 2) 各类编号(group 1)
|
||||
m = _CN_CHAPTER_RE.match(stripped)
|
||||
if m and stripped.startswith("第"):
|
||||
return (1, _NUM_ORDER["cn_chapter"]), (m.group(1) or "").strip()
|
||||
m = _CN_DUN_RE.match(stripped)
|
||||
if m:
|
||||
return (1, _NUM_ORDER["cn_dun"]), m.group(1).strip()
|
||||
m = _ARABIC_DOTTED_RE.match(stripped)
|
||||
if m:
|
||||
depth = m.group(1).count(".") + 1 # 1.1 → 2, 1.1.1 → 3
|
||||
# 多级阿拉伯:order 随层级递增,落在单级(3.0)与圈号(8.0)之间
|
||||
return (1, 3.0 + depth), m.group(2).strip()
|
||||
m = _PAREN_RE.match(stripped)
|
||||
if m:
|
||||
return (1, _NUM_ORDER["paren"]), m.group(1).strip()
|
||||
m = _ARABIC_SINGLE_RE.match(stripped)
|
||||
if m:
|
||||
return (1, _NUM_ORDER["arabic_single"]), m.group(1).strip()
|
||||
m = _CIRCLED_RE.match(stripped)
|
||||
if m:
|
||||
return (1, _NUM_ORDER["circled"]), m.group(1).strip()
|
||||
|
||||
# 3) bullet:用缩进深度区分层级(group 2)
|
||||
m = _BULLET_RE.match(stripped)
|
||||
if m:
|
||||
return (2, float(indent)), m.group(1).strip()
|
||||
|
||||
# 4) 纯文本:用缩进深度区分层级(group 3,最内层基准)
|
||||
return (3, float(indent)), stripped
|
||||
|
||||
|
||||
def _looks_like_prose(text: str) -> bool:
|
||||
"""判断一行是否更像散文句子而非大纲标题(长、或带句末标点)。"""
|
||||
if len(text) > 50:
|
||||
return True
|
||||
return text.rstrip()[-1:] in "。!?!?;;"
|
||||
|
||||
|
||||
def parse_user_outline(raw: str) -> Outline | None:
|
||||
"""把用户粘贴的大纲文本解析成三级 ``Outline``;无法可靠解析时返回 ``None``。
|
||||
|
||||
返回 ``None`` 的情形:空白输入、只有一行且无任何标记、或整体更像散文段落——这些交给
|
||||
上层用 LLM 兜底归一,避免误拆。
|
||||
"""
|
||||
if not raw or not raw.strip():
|
||||
return None
|
||||
|
||||
rows: list[tuple[tuple[int, float], str]] = []
|
||||
for line in raw.splitlines():
|
||||
if not line.strip():
|
||||
continue
|
||||
indent = _indent_width(line)
|
||||
classified = _classify(line.strip(), indent)
|
||||
if classified is None:
|
||||
continue
|
||||
key, text = classified
|
||||
if text:
|
||||
rows.append((key, text))
|
||||
|
||||
if not rows:
|
||||
return None
|
||||
|
||||
marked = any(key[0] != 3 for key, _ in rows) # 是否出现过标题/编号/bullet 标记
|
||||
|
||||
# 纯文本且无任何标记:可能是「平铺标题列表」也可能是散文。
|
||||
# 单行、或任一行像句子 → 判为散文,返回 None 交给 LLM。
|
||||
if not marked:
|
||||
if len(rows) < 2:
|
||||
return None
|
||||
if any(_looks_like_prose(text) for _, text in rows):
|
||||
return None
|
||||
|
||||
# 文章标题:仅在「Markdown 首行是孤立的最浅井号、且存在更深井号」时认定。
|
||||
title = ""
|
||||
body = rows
|
||||
heading_depths = [key[1] for key, _ in rows if key[0] == 0]
|
||||
if heading_depths:
|
||||
min_depth = min(heading_depths)
|
||||
first_key = rows[0][0]
|
||||
is_first_top_heading = first_key == (0, min_depth)
|
||||
only_one_at_top = sum(1 for d in heading_depths if d == min_depth) == 1
|
||||
has_deeper = any(d > min_depth for d in heading_depths)
|
||||
if is_first_top_heading and only_one_at_top and has_deeper:
|
||||
title = rows[0][1]
|
||||
body = rows[1:]
|
||||
|
||||
if not body:
|
||||
return None
|
||||
|
||||
# 把 body 里出现的不同「层级键」排序,按位置映射到 1/2/3 级(超过 3 级压到 3)。
|
||||
distinct_keys = sorted({key for key, _ in body})
|
||||
level_of: dict[tuple[int, float], int] = {
|
||||
key: min(idx + 1, 3) for idx, key in enumerate(distinct_keys)
|
||||
}
|
||||
|
||||
sections: list[OutlineSection] = []
|
||||
cur_section: OutlineSection | None = None
|
||||
cur_point: KeyPoint | None = None
|
||||
|
||||
for key, text in body:
|
||||
level = level_of[key]
|
||||
if level <= 1:
|
||||
cur_section = OutlineSection(section_title=text, key_points=[], material_ids=[])
|
||||
sections.append(cur_section)
|
||||
cur_point = None
|
||||
elif level == 2:
|
||||
if cur_section is None: # 孤立 L2:提升为章节,避免空标题
|
||||
cur_section = OutlineSection(section_title=text, key_points=[], material_ids=[])
|
||||
sections.append(cur_section)
|
||||
cur_point = None
|
||||
continue
|
||||
cur_point = KeyPoint(point_title=text, sub_points=[])
|
||||
cur_section["key_points"].append(cur_point)
|
||||
else: # level == 3
|
||||
if cur_section is None: # 孤立 L3:提升为章节
|
||||
cur_section = OutlineSection(section_title=text, key_points=[], material_ids=[])
|
||||
sections.append(cur_section)
|
||||
cur_point = None
|
||||
elif cur_point is None: # 章节下尚无小节:提升为小节
|
||||
cur_point = KeyPoint(point_title=text, sub_points=[])
|
||||
cur_section["key_points"].append(cur_point)
|
||||
else:
|
||||
cur_point["sub_points"].append(text)
|
||||
|
||||
if not sections:
|
||||
return None
|
||||
|
||||
return Outline(title=title, sections=sections)
|
||||
+54
@@ -123,6 +123,60 @@ OUTLINE_REVISE_USER = """写作意图:{user_intent}
|
||||
|
||||
请在「当前大纲」基础上落实上述「用户修改要求」,未提到的章节保持原样,输出调整后的完整大纲 JSON。"""
|
||||
|
||||
# ── 用户自带大纲:LLM 兜底归一(确定性解析失败时才用)──────────────────────────
|
||||
|
||||
OUTLINE_NORMALIZE_SYSTEM = """你是一名大纲格式整理助手。
|
||||
用户粘贴了一份已有的文章大纲(格式可能不规范),你的任务是把它原样转成结构化 JSON。
|
||||
|
||||
【最高原则:保真,绝不改写】
|
||||
- 严格保留用户写的每一条标题文字,逐字照搬,不要润色、改写、翻译、增加或删除任何条目。
|
||||
- 你只判断层级(哪条是章节、哪条是小节、哪条是子要点),不创作内容。
|
||||
- 层级判断依据用户原文里的编号(一、/(一)/1./1.1)、Markdown 井号数、或缩进深度。
|
||||
|
||||
【三级结构】
|
||||
- Level 1 章节 → section_title
|
||||
- Level 2 小节 → key_points[].point_title
|
||||
- Level 3 子要点 → key_points[].sub_points[](字符串数组)
|
||||
- 用户大纲只有一级时,每条都是章节,key_points 留空 []。
|
||||
- 用户大纲只有两级时,sub_points 留空 []。
|
||||
|
||||
【输出要求】
|
||||
- 只输出一个 JSON 对象,不要解释文字,不要 markdown 代码块,必须用英文双引号。
|
||||
- material_ids 一律输出 [](此步不关联素材)。
|
||||
- 结构:{"title": "...", "sections": [{"section_title": "...", "key_points": [{"point_title": "...", "sub_points": ["..."]}], "material_ids": []}]}
|
||||
- 用户大纲若包含一个总标题,放进 title;没有就留空字符串。"""
|
||||
|
||||
OUTLINE_NORMALIZE_USER = """用户粘贴的大纲原文:
|
||||
{raw_outline}
|
||||
|
||||
请逐条保留原文,仅判断层级,输出结构化大纲 JSON。"""
|
||||
|
||||
# ── 用户自带大纲:素材关联(结构锁定,只挑 material_ids)──────────────────────────
|
||||
|
||||
OUTLINE_MATERIAL_ASSIGN_SYSTEM = """你是一名资料编辑。
|
||||
给定一份【已确定的文章大纲】(章节顺序固定)和一批检索到的素材,你的任务是为每个章节
|
||||
挑选最相关的素材编号。
|
||||
|
||||
【规则】
|
||||
- 不要修改大纲的任何标题或结构,你只输出每个章节对应的素材编号列表。
|
||||
- 严格按大纲给出的章节顺序输出,一个章节对应一个编号数组;该章节没有相关素材就给 []。
|
||||
- material_id 必须来自给定素材清单里出现过的编号,不要编造。
|
||||
- 一条素材可以分配给多个章节,也可以一个都不分配。
|
||||
|
||||
【输出要求】
|
||||
- 只输出一个 JSON 对象,不要解释文字,不要 markdown 代码块,必须用英文双引号。
|
||||
- 结构:{"assignments": [["m001","m002"], [], ["m003"]]},数组长度等于章节数、顺序一致。"""
|
||||
|
||||
OUTLINE_MATERIAL_ASSIGN_USER = """写作意图:{user_intent}
|
||||
|
||||
文章大纲(章节按顺序编号):
|
||||
{sections_list}
|
||||
|
||||
可用素材:
|
||||
{materials_summary}
|
||||
|
||||
请按章节顺序,为每个章节挑选相关的 material_id,输出 assignments 数组。"""
|
||||
|
||||
DRAFT_SECTION_SYSTEM = """你是一名专业内容创作者(作家)。
|
||||
你的目标:根据大纲和素材,写出一段结构清晰、高质量的文章内容。
|
||||
你的工作背景:你有十年内容创作经验,文字流畅,善于用事实说话。
|
||||
|
||||
@@ -125,6 +125,9 @@ class BlockedSection(TypedDict):
|
||||
class AIWritingState(TypedDict):
|
||||
# 写作配置
|
||||
user_intent: str
|
||||
# 用户在写作表单粘贴的「已有大纲」原文(可选)。非空时 writer_outline 跳过 LLM 规划,
|
||||
# 直接解析这份大纲并按用户结构生成,做到「完全按我的大纲写」。留空走正常自动规划。
|
||||
user_outline_raw: str
|
||||
article_type: str
|
||||
target_audience: str
|
||||
word_count_target: int
|
||||
|
||||
+24
-3
@@ -11,6 +11,7 @@ from deerflow.agents.middlewares.memory_middleware import MemoryMiddleware
|
||||
from deerflow.agents.middlewares.prompt_prefix_middleware import PromptPrefixMiddleware
|
||||
from deerflow.agents.middlewares.rag_citation_middleware import RagCitationMiddleware
|
||||
from deerflow.agents.middlewares.setup_agent_approval_middleware import SetupAgentApprovalMiddleware
|
||||
from deerflow.agents.middlewares.setup_writing_approval_middleware import SetupWritingApprovalMiddleware
|
||||
from deerflow.agents.middlewares.subagent_limit_middleware import SubagentLimitMiddleware
|
||||
from deerflow.agents.middlewares.summarization_middleware import BeforeSummarizationHook, DeerFlowSummarizationMiddleware
|
||||
from deerflow.agents.middlewares.title_middleware import TitleMiddleware
|
||||
@@ -272,6 +273,7 @@ def _build_middlewares(
|
||||
# Add TodoList middleware if plan mode is enabled
|
||||
is_plan_mode = cfg.get("is_plan_mode", False)
|
||||
is_bootstrap = cfg.get("is_bootstrap", False)
|
||||
is_writing_setup = cfg.get("is_writing_setup", False)
|
||||
todo_list_middleware = _create_todo_list_middleware(is_plan_mode)
|
||||
if todo_list_middleware is not None:
|
||||
middlewares.append(todo_list_middleware)
|
||||
@@ -285,7 +287,7 @@ def _build_middlewares(
|
||||
|
||||
# Add MemoryMiddleware (after TitleMiddleware) — skip in bootstrap mode (agent creation flow)
|
||||
memory_injection_enabled = cfg.get("memory_injection_enabled", True)
|
||||
if not is_bootstrap:
|
||||
if not is_bootstrap and not is_writing_setup:
|
||||
middlewares.append(MemoryMiddleware(agent_name=agent_name, memory_config=resolved_app_config.memory, injection_enabled=memory_injection_enabled))
|
||||
|
||||
# Add ViewImageMiddleware only if the current model supports vision.
|
||||
@@ -314,6 +316,7 @@ def _build_middlewares(
|
||||
middlewares.extend(custom_middlewares)
|
||||
|
||||
middlewares.append(SetupAgentApprovalMiddleware())
|
||||
middlewares.append(SetupWritingApprovalMiddleware())
|
||||
|
||||
# SkillStopMiddleware — intercepts skill invocations and halts execution
|
||||
# when a configured skill is about to run. Request-level options
|
||||
@@ -361,7 +364,7 @@ def make_lead_agent(config: RunnableConfig):
|
||||
def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
|
||||
# Lazy import to avoid circular dependency
|
||||
from deerflow.tools import get_available_tools
|
||||
from deerflow.tools.builtins import setup_agent
|
||||
from deerflow.tools.builtins import setup_agent, setup_writing
|
||||
|
||||
cfg = _get_runtime_config(config)
|
||||
resolved_app_config = app_config
|
||||
@@ -373,6 +376,10 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
|
||||
subagent_enabled = cfg.get("subagent_enabled", False)
|
||||
max_concurrent_subagents = cfg.get("max_concurrent_subagents", 3)
|
||||
is_bootstrap = cfg.get("is_bootstrap", False)
|
||||
# 写作配置 Q&A:与 is_bootstrap 同构的另一种「最小化 bootstrap」——挂 setup_writing 工具、
|
||||
# 走写作配置专用 prompt,只负责帮用户把写作表单聊清并提交,不真正开始写作。
|
||||
is_writing_setup = cfg.get("is_writing_setup", False)
|
||||
writing_setup_article_types = cfg.get("valid_article_types") if is_writing_setup else None
|
||||
memory_injection_enabled = bool(cfg.get("memory_injection_enabled", True))
|
||||
rag_mode_enabled = bool(cfg.get("rag_mode_enabled", False))
|
||||
writing_mode = bool(cfg.get("writing_mode", False))
|
||||
@@ -383,7 +390,7 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
|
||||
requested_excluded_tools = cfg.get("excluded_tools") or None
|
||||
|
||||
try:
|
||||
agent_config = load_agent_config(agent_id) if not is_bootstrap else None
|
||||
agent_config = load_agent_config(agent_id) if not is_bootstrap and not is_writing_setup else None
|
||||
except FileNotFoundError:
|
||||
logger.warning("Agent directory not found for '%s'; running with default config", agent_id)
|
||||
agent_config = None
|
||||
@@ -443,6 +450,20 @@ def _make_lead_agent(config: RunnableConfig, *, app_config: AppConfig):
|
||||
}
|
||||
)
|
||||
|
||||
if is_writing_setup:
|
||||
# 写作配置 Q&A:最小化 agent —— **只挂 setup_writing 一个工具** + 一份职责单一的独立提示。
|
||||
# 关键:绝不给全套工具、绝不复用 lead 超级体提示,否则模型会把「新能源汽车市场分析」之类
|
||||
# 长得像任务的输入直接做掉,而不是转成写作配置提交。
|
||||
from deerflow.agents.lead_agent.prompt import build_writing_setup_prompt
|
||||
|
||||
return create_agent(
|
||||
model=create_chat_model(name=model_name, thinking_enabled=thinking_enabled, app_config=resolved_app_config),
|
||||
tools=[setup_writing],
|
||||
middleware=_build_middlewares(config, model_name=model_name, app_config=resolved_app_config),
|
||||
system_prompt=build_writing_setup_prompt(writing_setup_article_types),
|
||||
state_schema=ThreadState,
|
||||
)
|
||||
|
||||
if is_bootstrap:
|
||||
# Special bootstrap agent with minimal prompt for initial custom agent creation flow
|
||||
return create_agent(
|
||||
|
||||
+58
-18
@@ -703,14 +703,14 @@ def _get_cached_skills_prompt_section(
|
||||
|
||||
display_pattern = (
|
||||
"""
|
||||
**Skill Result Display Pattern:**
|
||||
- If RAG/reference display mode is enabled, do not number each skill result independently. Treat all retrieval results from the same user turn as one combined `referenceBatch` and cite with `[1]`, `[2]`, etc. from the system-provided reference prompt.
|
||||
- In RAG/reference display mode, do not output `<skill-display .../>`; the frontend renders cards, tables, lists, and citation details in the right reference panel.
|
||||
- If RAG/reference display mode is not enabled and a skill has `<display mode="citation">`, use the skill's returned result order as reference numbers. Cite facts from the first result as `[1]`, the second as `[2]`, and so on. Use only numbers that correspond to returned results.
|
||||
- If a skill has `<display mode="list">`, summarize the returned items as a concise list and keep item order stable.
|
||||
- If a skill has `<display mode="table">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="table" />` where the frontend should render the real table. Do not hand-write the table rows yourself.
|
||||
- If a skill has `<display mode="card">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="card" />` where the frontend should render the cards.
|
||||
- If a skill has `<display mode="custom">`, follow the custom display instruction in the `<display>` tag when presenting the skill result.
|
||||
**Skill Result Display Pattern:**
|
||||
- If RAG/reference display mode is enabled, do not number each skill result independently. Treat all retrieval results from the same user turn as one combined `referenceBatch` and cite with `[1]`, `[2]`, etc. from the system-provided reference prompt.
|
||||
- In RAG/reference display mode, do not output `<skill-display .../>`; the frontend renders cards, tables, lists, and citation details in the right reference panel.
|
||||
- If RAG/reference display mode is not enabled and a skill has `<display mode="citation">`, use the skill's returned result order as reference numbers. Cite facts from the first result as `[1]`, the second as `[2]`, and so on. Use only numbers that correspond to returned results.
|
||||
- If a skill has `<display mode="list">`, summarize the returned items as a concise list and keep item order stable.
|
||||
- If a skill has `<display mode="table">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="table" />` where the frontend should render the real table. Do not hand-write the table rows yourself.
|
||||
- If a skill has `<display mode="card">`, summarize the result and output `<skill-display tool="SKILL_OR_TOOL_NAME" mode="card" />` where the frontend should render the cards.
|
||||
- If a skill has `<display mode="custom">`, follow the custom display instruction in the `<display>` tag when presenting the skill result.
|
||||
"""
|
||||
if display_signature
|
||||
else ""
|
||||
@@ -904,6 +904,46 @@ def _build_bootstrap_section() -> str:
|
||||
"""
|
||||
|
||||
|
||||
def build_writing_setup_prompt(article_types: list[dict] | None = None) -> str:
|
||||
"""写作配置 Q&A 的**独立**系统提示(不复用 lead-agent 超级体提示)。
|
||||
|
||||
关键:写作配置助手绝不能像 lead agent 那样「自己把任务做掉」——用户输入常常长得像一个
|
||||
可执行任务(如「新能源汽车市场分析」),若复用 lead 提示 + 全套工具,模型会直接去研究/写
|
||||
分析。这里用一份职责单一、口吻强硬的独立提示,配合「只挂 setup_writing 一个工具」,把模型
|
||||
死死框在「聊清需求 → 调 setup_writing」上。
|
||||
"""
|
||||
if article_types:
|
||||
type_lines = "\n".join(
|
||||
f" - `{t.get('key', '')}`:{t.get('label', '')}" for t in article_types if t.get("key")
|
||||
)
|
||||
type_block = f"【可选文章类型】article_type 只能填下列 key 之一:\n{type_lines}"
|
||||
else:
|
||||
type_block = "【文章类型】若不确定 article_type,可留空,交给用户在表单里选。"
|
||||
|
||||
return f"""你是「AI 写作」功能的**配置助手**。你的唯一职责:通过简短对话把用户的写作需求聊清楚,然后调用 `setup_writing` 工具把写作配置提交给用户确认。提交后由用户在表单里应用、再点「开始写作」。
|
||||
|
||||
【铁律——最高优先级】
|
||||
- 你**绝对不撰写**文章正文、分析内容、大纲内容或任何成稿文字。即使用户的输入看起来就是一个写作任务(例如「新能源汽车市场分析」「帮我写一篇……」),你也**只把它当作"要配置的写作目标"**,绝不替用户动笔、绝不直接完成这篇写作。
|
||||
- 你能做的动作只有两个:① 用自然语言向用户提问或确认;② 调用 `setup_writing` 工具。**除此之外你没有任何其它工具**,也不要假装调用别的工具、不要做检索、不要做研究。
|
||||
- 不要在回复里输出文章段落、要点清单或大纲(用户主动粘贴的大纲除外——原样放进 `outline` 参数)。
|
||||
|
||||
【工作流程】
|
||||
1. 读用户输入。如果「写什么」已经明确,关键信息(文章类型、字数、读者)即使缺失也可以用合理默认补全——**信息基本够了就直接进入第 3 步,不要为了凑齐而反复追问**。
|
||||
2. 只有当「写什么」本身都不清楚时,才用一句话问 1-2 个最关键的问题(最多一轮)。
|
||||
3. 调用 `setup_writing` 提交配置:
|
||||
- `user_intent`:精炼成一句话的写作目标(必填)。
|
||||
- `article_type`:见下方可选类型;不确定就留空。
|
||||
- `word_count_target`:正整数;用户没明说就给合理默认(如 800)。
|
||||
- `target_audience`:如「普通读者」「行业专家」;没提就填「普通读者」。
|
||||
- `writing_mode`:用户强调「严格依据素材/不要编造」→ `strict`,否则 `loose`。
|
||||
- `material_source`:用户提到用「技能/技能检索」→ `skill`,否则 `knowledge_base`。
|
||||
- `outline`:用户提供的大纲原文(保留其层级与标题),否则留空。
|
||||
4. 调用 `setup_writing` 前,用**一句话**说明你将提交的配置要点;调用之后不要再输出任何文章内容。
|
||||
|
||||
{type_block}
|
||||
"""
|
||||
|
||||
|
||||
def apply_prompt_template(
|
||||
subagent_enabled: bool = False,
|
||||
max_concurrent_subagents: int = 3,
|
||||
@@ -995,16 +1035,16 @@ Writing mode is enabled for this run.
|
||||
network_policy_section = _build_network_policy_section()
|
||||
rag_mode_section = (
|
||||
"""
|
||||
<rag_citation_mode>
|
||||
RAG/reference display mode is enabled.
|
||||
When a search, knowledge, skill, or MCP tool returns a result list:
|
||||
- If multiple search terms are needed, prefer calling the retrieval tools for all known terms in the same step, then answer after the tool results are returned.
|
||||
- Treat every retrieval result from the same user turn as one combined `referenceBatch`; do not restart numbering for each search term, tool call, skill, or display adapter.
|
||||
- Use only the system-provided merged reference numbers in the final answer, formatted as `[1]`, `[2]`, `[1][3][5]`.
|
||||
- Do not output `<skill-display .../>` in RAG/reference mode. The frontend renders configured cards, tables, lists, and citation details in the right reference panel.
|
||||
- If no display adapter exists, the frontend defaults the right reference panel to numbered citations.
|
||||
</rag_citation_mode>
|
||||
"""
|
||||
<rag_citation_mode>
|
||||
RAG/reference display mode is enabled.
|
||||
When a search, knowledge, skill, or MCP tool returns a result list:
|
||||
- If multiple search terms are needed, prefer calling the retrieval tools for all known terms in the same step, then answer after the tool results are returned.
|
||||
- Treat every retrieval result from the same user turn as one combined `referenceBatch`; do not restart numbering for each search term, tool call, skill, or display adapter.
|
||||
- Use only the system-provided merged reference numbers in the final answer, formatted as `[1]`, `[2]`, `[1][3][5]`.
|
||||
- Do not output `<skill-display .../>` in RAG/reference mode. The frontend renders configured cards, tables, lists, and citation details in the right reference panel.
|
||||
- If no display adapter exists, the frontend defaults the right reference panel to numbered citations.
|
||||
</rag_citation_mode>
|
||||
"""
|
||||
if rag_mode_enabled
|
||||
else ""
|
||||
)
|
||||
|
||||
+129
@@ -0,0 +1,129 @@
|
||||
"""Approval middleware for the AI-writing config setup tool.
|
||||
|
||||
与 ``SetupAgentApprovalMiddleware`` 同构:模型在「写作配置 Q&A」里调用 ``setup_writing``
|
||||
时,于 ``after_model`` 拦截,附上一条结构化审批 ToolMessage(建议的写作配置打包在
|
||||
``additional_kwargs["setup_writing_approval"]``)并 ``jump_to="end"`` 立即结束本轮运行——
|
||||
前端据此渲染审批卡,用户点「应用到表单」后把这些字段回填到左侧写作表单。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
from hashlib import sha256
|
||||
from typing import Any, override
|
||||
|
||||
from langchain.agents import AgentState
|
||||
from langchain.agents.middleware import AgentMiddleware, hook_config
|
||||
from langchain_core.messages import AIMessage, ToolMessage
|
||||
from langgraph.runtime import Runtime
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class SetupWritingApprovalMiddlewareState(AgentState):
|
||||
"""Compatible with the DeerFlow thread state schema."""
|
||||
|
||||
pass
|
||||
|
||||
|
||||
class SetupWritingApprovalMiddleware(AgentMiddleware[SetupWritingApprovalMiddlewareState]):
|
||||
"""Stop before ``setup_writing`` and ask the user to apply the proposed config.
|
||||
|
||||
See ``SetupAgentApprovalMiddleware`` for the rationale behind intercepting in
|
||||
``after_model`` + ``jump_to="end"`` (robust against parallel tool calls).
|
||||
"""
|
||||
|
||||
state_schema = SetupWritingApprovalMiddlewareState
|
||||
|
||||
def _is_approved(self, runtime: Runtime) -> bool:
|
||||
context = getattr(runtime, "context", None) or {}
|
||||
if not isinstance(context, dict):
|
||||
return False
|
||||
return context.get("setup_writing_approved") is True
|
||||
|
||||
def _stable_message_id(self, tool_call_id: str, args: dict[str, Any]) -> str:
|
||||
if tool_call_id:
|
||||
return f"setup-writing-approval:{tool_call_id}"
|
||||
digest = sha256(repr(sorted(args.items())).encode("utf-8")).hexdigest()[:16]
|
||||
return f"setup-writing-approval:{digest}"
|
||||
|
||||
def _coerce_word_count(self, raw: Any) -> int | None:
|
||||
if isinstance(raw, bool): # bool 是 int 子类,先排除
|
||||
return None
|
||||
if isinstance(raw, int):
|
||||
return raw if raw > 0 else None
|
||||
if isinstance(raw, float):
|
||||
return int(raw) if raw > 0 else None
|
||||
if isinstance(raw, str):
|
||||
import re
|
||||
|
||||
m = re.search(r"\d[\d,]*", raw)
|
||||
if m:
|
||||
try:
|
||||
value = int(m.group().replace(",", ""))
|
||||
return value if value > 0 else None
|
||||
except ValueError:
|
||||
return None
|
||||
return None
|
||||
|
||||
def _build_payload(self, args: dict[str, Any]) -> dict[str, Any]:
|
||||
return {
|
||||
"user_intent": str(args.get("user_intent") or "").strip(),
|
||||
"article_type": str(args.get("article_type") or "").strip(),
|
||||
"word_count_target": self._coerce_word_count(args.get("word_count_target")),
|
||||
"target_audience": str(args.get("target_audience") or "").strip(),
|
||||
"writing_mode": str(args.get("writing_mode") or "").strip(),
|
||||
"material_source": str(args.get("material_source") or "").strip(),
|
||||
"outline": str(args.get("outline") or "").strip(),
|
||||
}
|
||||
|
||||
def _intercept(self, last: AIMessage) -> dict[str, Any] | None:
|
||||
"""Return a state update if setup_writing must be gated; None to passthrough."""
|
||||
tool_calls = list(last.tool_calls or [])
|
||||
setup_call = next((tc for tc in tool_calls if tc.get("name") == "setup_writing"), None)
|
||||
if setup_call is None:
|
||||
return None
|
||||
|
||||
args = setup_call.get("args") or {}
|
||||
tool_call_id = setup_call.get("id") or ""
|
||||
payload = self._build_payload(args)
|
||||
logger.info("Setup writing approval requested before applying writing config")
|
||||
|
||||
approval_msg = ToolMessage(
|
||||
id=self._stable_message_id(tool_call_id, args),
|
||||
content="请确认是否把这份写作配置应用到表单。",
|
||||
tool_call_id=tool_call_id,
|
||||
name="setup_writing",
|
||||
additional_kwargs={"setup_writing_approval": payload},
|
||||
)
|
||||
|
||||
return {
|
||||
"messages": [approval_msg],
|
||||
"jump_to": "end",
|
||||
}
|
||||
|
||||
@hook_config(can_jump_to=["end"])
|
||||
@override
|
||||
def after_model(
|
||||
self,
|
||||
state: SetupWritingApprovalMiddlewareState,
|
||||
runtime: Runtime,
|
||||
) -> dict[str, Any] | None:
|
||||
if self._is_approved(runtime):
|
||||
return None
|
||||
messages = state.get("messages") or []
|
||||
if not messages:
|
||||
return None
|
||||
last = messages[-1]
|
||||
if not isinstance(last, AIMessage):
|
||||
return None
|
||||
return self._intercept(last)
|
||||
|
||||
@hook_config(can_jump_to=["end"])
|
||||
@override
|
||||
async def aafter_model(
|
||||
self,
|
||||
state: SetupWritingApprovalMiddlewareState,
|
||||
runtime: Runtime,
|
||||
) -> dict[str, Any] | None:
|
||||
return self.after_model(state, runtime)
|
||||
@@ -4,11 +4,13 @@ from .memory_tool import memory_tool
|
||||
from .present_file_tool import present_file_tool
|
||||
from .scheduled_task_tool import scheduled_task_tool
|
||||
from .setup_agent_tool import setup_agent
|
||||
from .setup_writing_tool import setup_writing
|
||||
from .task_tool import task_tool
|
||||
from .view_image_tool import view_image_tool
|
||||
|
||||
__all__ = [
|
||||
"setup_agent",
|
||||
"setup_writing",
|
||||
"present_file_tool",
|
||||
"ask_clarification_tool",
|
||||
"agent_orchestration_tool",
|
||||
|
||||
+53
@@ -0,0 +1,53 @@
|
||||
"""写作配置助手用的提交工具。
|
||||
|
||||
与 ``setup_agent`` 同构:模型在「写作配置 Q&A」里与用户聊清需求后,调用本工具把建议的
|
||||
写作配置「提交」出来。实际它总是被 ``SetupWritingApprovalMiddleware`` 在 ``after_model``
|
||||
拦截(不真正执行)——拦截后把建议字段打包进审批 ToolMessage,前端据此回填左侧写作表单。
|
||||
|
||||
因此本工具的函数体基本是兜底:万一未被拦截(如审批被预先放行),也只回一句确认、不落任何盘。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import logging
|
||||
|
||||
from langchain_core.messages import ToolMessage
|
||||
from langchain_core.tools import tool
|
||||
from langgraph.prebuilt import ToolRuntime
|
||||
from langgraph.types import Command
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
@tool
|
||||
def setup_writing(
|
||||
user_intent: str,
|
||||
runtime: ToolRuntime,
|
||||
article_type: str | None = None,
|
||||
word_count_target: int | None = None,
|
||||
target_audience: str | None = None,
|
||||
writing_mode: str | None = None,
|
||||
material_source: str | None = None,
|
||||
outline: str | None = None,
|
||||
) -> Command:
|
||||
"""把与用户聊清的写作配置提交给前端表单(不直接开始写作)。
|
||||
|
||||
Args:
|
||||
user_intent: 用户的写作意图(精炼成一句话的写作目标)。
|
||||
article_type: 文章类型的 key(须是系统给定的合法类型之一;不确定可留空)。
|
||||
word_count_target: 目标字数(正整数)。
|
||||
target_audience: 目标读者(如「行业专家」「普通读者」)。
|
||||
writing_mode: 写作模式,'strict'(严格依据素材)或 'loose'(可结合通用知识)。
|
||||
material_source: 素材来源,'knowledge_base'(知识库)或 'skill'(技能检索)。
|
||||
outline: 若用户已提供大纲,原样填入其大纲文本(保留其层级与标题),否则留空。
|
||||
"""
|
||||
logger.info("[setup_writing] proposed writing config submitted (intercepted by approval middleware)")
|
||||
return Command(
|
||||
update={
|
||||
"messages": [
|
||||
ToolMessage(
|
||||
content="写作配置已生成,请在右侧确认后应用到表单。",
|
||||
tool_call_id=runtime.tool_call_id,
|
||||
)
|
||||
]
|
||||
}
|
||||
)
|
||||
Reference in new issue
Block a user