阶段三(治理):
- sensitive.py:写入前脱敏 token/密钥/手机号/身份证/邮箱(sensitive_redaction_enabled)
- 版本快照:更新前写 knowledge_note_versions;GET /notes/{id}/versions
- dedup.py:标题/摘要 Jaccard + 内容哈希查重;GET /duplicates、POST /merge(来源并入主条目、重复项归档)
- auto_ingest.py:独立后台队列/线程沉淀,不阻塞问答;中间件 after_agent 在 auto_ingest_enabled 时入队;价值判断(最小字符)+ skip_existing 去重;自动沉淀落 draft 待审
阶段四(图谱):
- LLM 抽取 entities/relations;service 存 knowledge_entities/knowledge_relations 并连边
- exporter.build_graph 支持实体节点 + note→entity/entity→entity 边;graph.html/json 增强
- GET /api/knowledge/graph 返回产品图谱数据
前端:知识库页「查重合并」对话框 + 相关 api/hooks
配置:sensitive_redaction_enabled / dedup_similarity_threshold / entity_extraction_enabled
测试:tests/test_knowledge.py 20 passed
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
215 lines
8.4 KiB
Python
215 lines
8.4 KiB
Python
"""LLM-based knowledge extraction (wiki-capture style).
|
||
|
||
Turns a conversation into a *declarative* knowledge note instead of dumping the
|
||
raw chat. Follows the obsidian-wiki ``wiki-capture`` skill: classify the note,
|
||
keep only high-value findings/decisions, drop greetings / intermediate
|
||
reasoning / verbatim transcripts, mark inferred content, and emit structured
|
||
JSON that we render into the Context / Finding / Reasoning / Implications /
|
||
Related body.
|
||
|
||
Falls back to the rule-based :func:`deerflow.knowledge.extractor.extract_thread`
|
||
when the LLM is unavailable or returns something unusable, so capture never
|
||
hard-fails.
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import logging
|
||
import re
|
||
from typing import Any
|
||
|
||
from deerflow.knowledge import extractor
|
||
from deerflow.knowledge.schemas import NoteDraft, SourceDraft
|
||
from deerflow.knowledge.skill_templates import WIKI_CAPTURE_GUIDANCE
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
_VALID_CATEGORIES = {"synthesis", "concepts", "references", "decision", "journal"}
|
||
|
||
_SYSTEM_PROMPT = """\
|
||
你是知识库整理专家,遵循 obsidian-wiki 的 wiki-capture 规范,把一次对话提炼成\
|
||
**陈述式知识(declarative knowledge)**,而不是聊天流水账。
|
||
|
||
要求:
|
||
- 只保留高价值内容:明确结论、方案、决策、关键事实、被引用的来源。
|
||
- 丢弃:寒暄、客套、中间推理过程、重复内容、无意义的原文堆砌。
|
||
- 不要写“用户问了什么、AI 回答了什么”,要写“结论是什么、为什么成立”。
|
||
- 推断内容用 ^[inferred] 标记,不确定/冲突用 ^[ambiguous] 标记。
|
||
- 分类 category 取值之一:synthesis(分析/方案/结论)、concepts(概念/框架)、\
|
||
references(外部来源汇总)、decision(架构/设计决策)、journal(会话纪要)。
|
||
- 用中文输出。
|
||
|
||
只输出一个 JSON 对象,不要任何额外文字或 Markdown 代码块,结构如下:
|
||
{
|
||
"should_ingest": true,
|
||
"reason": "为什么值得/不值得沉淀",
|
||
"title": "简洁的知识标题(不超过 40 字)",
|
||
"summary": "1-2 句话概括这页知识",
|
||
"category": "synthesis",
|
||
"context": "问题背景或触发场景,1-3 句",
|
||
"findings": ["结论/要点(陈述式,可含 ^[inferred])", "..."],
|
||
"reasoning": "为什么成立、取舍、不确定性",
|
||
"implications": ["后续建议/风险/下一步", "..."],
|
||
"tags": ["标签1", "标签2"],
|
||
"entities": [{"name": "实体名", "type": "technology|concept|person|org|product|other"}],
|
||
"relations": [{"from": "实体A", "to": "实体B", "type": "uses|relates_to|part_of|depends_on"}],
|
||
"confidence": 0.0
|
||
}"""
|
||
|
||
|
||
def _build_transcript(turns: list[dict[str, str]], sources: list[SourceDraft], *, max_chars: int) -> str:
|
||
parts: list[str] = []
|
||
for i, t in enumerate(turns, 1):
|
||
if t.get("user"):
|
||
parts.append(f"[用户 {i}] {t['user']}")
|
||
if t.get("ai"):
|
||
parts.append(f"[助手 {i}] {t['ai']}")
|
||
if sources:
|
||
parts.append("\n[来源]")
|
||
for s in sources[:20]:
|
||
label = s.title or s.url or s.tool_name or "来源"
|
||
parts.append(f"- {label}" + (f" ({s.url})" if s.url else ""))
|
||
text = "\n\n".join(parts)
|
||
if len(text) > max_chars:
|
||
text = text[:max_chars] + "\n…(已截断)"
|
||
return text
|
||
|
||
|
||
def _parse_json(raw: str) -> dict[str, Any] | None:
|
||
if not raw:
|
||
return None
|
||
text = raw.strip()
|
||
# Strip ```json fences if present.
|
||
fence = re.search(r"```(?:json)?\s*(.*?)```", text, re.DOTALL)
|
||
if fence:
|
||
text = fence.group(1).strip()
|
||
# Otherwise grab the outermost { ... }.
|
||
if not text.startswith("{"):
|
||
start = text.find("{")
|
||
end = text.rfind("}")
|
||
if start >= 0 and end > start:
|
||
text = text[start : end + 1]
|
||
try:
|
||
data = json.loads(text)
|
||
return data if isinstance(data, dict) else None
|
||
except json.JSONDecodeError:
|
||
logger.warning("Knowledge LLM returned non-JSON output; falling back to rule-based")
|
||
return None
|
||
|
||
|
||
def _render_body(data: dict[str, Any], title: str, related: list[str], sources: list[SourceDraft]) -> str:
|
||
def _bullets(items: Any) -> str:
|
||
if isinstance(items, str):
|
||
items = [items]
|
||
if not isinstance(items, list):
|
||
return ""
|
||
return "\n".join(f"- {str(x).strip()}" for x in items if str(x).strip())
|
||
|
||
context = str(data.get("context") or "").strip() or "- 来自一次对话沉淀。^[inferred]"
|
||
findings = _bullets(data.get("findings")) or "_暂无可提炼的结论。_"
|
||
reasoning = str(data.get("reasoning") or "").strip() or "(未提供推理说明)"
|
||
implications = _bullets(data.get("implications")) or "- 如内容过期或冲突,请编辑或归档。"
|
||
|
||
related_block = "\n".join(f"- {link}" for link in related)
|
||
if sources:
|
||
related_block += "\n\n### Sources\n\n"
|
||
for s in sources[:20]:
|
||
label = s.title or s.url or "来源"
|
||
related_block += f"- {label}" + (f" — {s.url}" if s.url else "") + "\n"
|
||
|
||
return (
|
||
f"# {title}\n\n"
|
||
f"## Context\n\n{context}\n\n"
|
||
f"## Finding / Decision\n\n{findings}\n\n"
|
||
f"## Reasoning\n\n{reasoning}\n\n"
|
||
f"## Implications\n\n{implications}\n\n"
|
||
f"## Related\n\n{related_block}\n"
|
||
)
|
||
|
||
|
||
async def extract_thread_llm(
|
||
messages: list[dict[str, Any]],
|
||
*,
|
||
thread_id: str,
|
||
thread_title: str | None = None,
|
||
mode: str = "summary",
|
||
title: str | None = None,
|
||
tags: list[str] | None = None,
|
||
project_name: str | None = "deerflow",
|
||
sources: list[SourceDraft] | None = None,
|
||
max_messages: int = 80,
|
||
model_name: str | None = None,
|
||
max_input_chars: int = 12000,
|
||
) -> NoteDraft | None:
|
||
"""Distill a thread into a :class:`NoteDraft` via the LLM.
|
||
|
||
Returns ``None`` when there is nothing to sediment **or** the LLM path
|
||
fails — callers should fall back to the rule-based extractor.
|
||
"""
|
||
turns = extractor.split_turns(messages, max_messages=max_messages)
|
||
if not turns:
|
||
return None
|
||
sources = sources or []
|
||
|
||
try:
|
||
from langchain_core.messages import HumanMessage, SystemMessage
|
||
|
||
from deerflow.models import create_chat_model
|
||
|
||
model = create_chat_model(name=model_name or None, thinking_enabled=False)
|
||
transcript = _build_transcript(turns, sources, max_chars=max_input_chars)
|
||
prompt = [
|
||
SystemMessage(content=_SYSTEM_PROMPT + "\n\n" + WIKI_CAPTURE_GUIDANCE),
|
||
HumanMessage(content=f"以下是一次对话,请按规范提炼为知识 JSON:\n\n{transcript}"),
|
||
]
|
||
response = await model.ainvoke(prompt)
|
||
raw = response.content if isinstance(response.content, str) else str(response.content)
|
||
data = _parse_json(raw)
|
||
except Exception:
|
||
logger.exception("Knowledge LLM extraction failed for thread %s; using rule-based fallback", thread_id)
|
||
return None
|
||
|
||
if not data:
|
||
return None
|
||
|
||
category = str(data.get("category") or "synthesis").strip().lower()
|
||
if category not in _VALID_CATEGORIES:
|
||
category = "synthesis"
|
||
|
||
final_title = (title or "").strip() or str(data.get("title") or "").strip() or extractor._auto_title(thread_title, turns)
|
||
summary = str(data.get("summary") or "").strip()
|
||
related = extractor.related_links(project_name)
|
||
body = _render_body(data, final_title, related, sources)
|
||
|
||
base_tags = list(tags or [])
|
||
llm_tags = data.get("tags") if isinstance(data.get("tags"), list) else []
|
||
for tag in [*llm_tags, project_name or "deerflow", "knowledge-base"]:
|
||
tag = str(tag).strip()
|
||
if tag and tag not in base_tags:
|
||
base_tags.append(tag)
|
||
|
||
try:
|
||
confidence = float(data.get("confidence") or 0.7)
|
||
except (TypeError, ValueError):
|
||
confidence = 0.7
|
||
|
||
entities = data.get("entities") if isinstance(data.get("entities"), list) else []
|
||
relations = data.get("relations") if isinstance(data.get("relations"), list) else []
|
||
|
||
return NoteDraft(
|
||
title=final_title,
|
||
summary=summary or final_title,
|
||
content_md=body,
|
||
category=category,
|
||
source_type="thread",
|
||
source_id=thread_id,
|
||
tags=base_tags,
|
||
confidence=max(0.0, min(confidence, 1.0)),
|
||
related=related,
|
||
sources=sources,
|
||
entities=[e for e in entities if isinstance(e, dict) and e.get("name")],
|
||
relations=[r for r in relations if isinstance(r, dict) and r.get("from") and r.get("to")],
|
||
source_key=f"thread:{thread_id}",
|
||
)
|