cfdaily
292 lines
9.4 KiB
Python
292 lines
9.4 KiB
Python
"""Experience Distillation — 经验蒸馏
|
||
|
||
从已完成的任务产出中提取经验：
|
||
1. 模式识别（成功/失败 pattern）
|
||
2. 经验分类（pitfall / best_practice / environment）
|
||
3. 向量索引 + Markdown 持久化
|
||
4. 自动推荐给后续相似任务
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import logging
|
||
import re
|
||
from datetime import datetime
|
||
from pathlib import Path
|
||
from typing import Any, Dict, List, Optional
|
||
|
||
logger = logging.getLogger("moziplus-v2.experience")
|
||
|
||
|
||
class ExperienceCategory(str):
|
||
PITFALL = "pitfall"
|
||
BEST_PRACTICE = "best_practice"
|
||
ENVIRONMENT = "environment"
|
||
PATTERN = "pattern"
|
||
DECISION = "decision"
|
||
|
||
|
||
class Experience:
|
||
"""一条经验"""
|
||
|
||
def __init__(
|
||
self,
|
||
category: str,
|
||
summary: str,
|
||
source_task_id: Optional[str] = None,
|
||
agent_id: Optional[str] = None,
|
||
evidence: Optional[str] = None,
|
||
tags: Optional[List[str]] = None,
|
||
confidence: float = 0.8,
|
||
experience_id: Optional[str] = None,
|
||
created_at: Optional[str] = None,
|
||
):
|
||
self.id = experience_id or f"exp-{datetime.utcnow().strftime('%Y%m%d%H%M%S')}-{hash(summary) % 10000:04d}"
|
||
self.category = category
|
||
self.summary = summary
|
||
self.source_task_id = source_task_id
|
||
self.agent_id = agent_id
|
||
self.evidence = evidence
|
||
self.tags = tags or []
|
||
self.confidence = confidence
|
||
self.created_at = created_at or datetime.utcnow().isoformat()
|
||
|
||
def to_dict(self) -> Dict[str, Any]:
|
||
return {
|
||
"id": self.id,
|
||
"category": self.category,
|
||
"summary": self.summary,
|
||
"source_task_id": self.source_task_id,
|
||
"agent_id": self.agent_id,
|
||
"evidence": self.evidence,
|
||
"tags": self.tags,
|
||
"confidence": self.confidence,
|
||
"created_at": self.created_at,
|
||
}
|
||
|
||
@classmethod
|
||
def from_dict(cls, data: Dict[str, Any]) -> Experience:
|
||
return cls(**{k: v for k, v in data.items() if k != "id"},
|
||
experience_id=data.get("id"))
|
||
|
||
|
||
class ExperienceStore:
|
||
"""经验存储（JSONL 持久化）"""
|
||
|
||
def __init__(self, store_path: Optional[Path] = None):
|
||
self.store_path = store_path
|
||
self._experiences: Dict[str, Experience] = {}
|
||
|
||
if store_path and store_path.exists():
|
||
self._load()
|
||
|
||
def _load(self):
|
||
if not self.store_path:
|
||
return
|
||
for line in self.store_path.read_text().strip().split("\n"):
|
||
line = line.strip()
|
||
if not line:
|
||
continue
|
||
try:
|
||
exp = Experience.from_dict(json.loads(line))
|
||
self._experiences[exp.id] = exp
|
||
except Exception:
|
||
logger.warning("Failed to parse experience: %s", line[:80])
|
||
|
||
def _save(self):
|
||
if not self.store_path:
|
||
return
|
||
lines = [json.dumps(exp.to_dict(), ensure_ascii=False)
|
||
for exp in self._experiences.values()]
|
||
self.store_path.write_text("\n".join(lines) + "\n")
|
||
|
||
def add(self, experience: Experience) -> str:
|
||
self._experiences[experience.id] = experience
|
||
self._save()
|
||
return experience.id
|
||
|
||
def get(self, experience_id: str) -> Optional[Experience]:
|
||
return self._experiences.get(experience_id)
|
||
|
||
def list_all(self) -> List[Experience]:
|
||
return list(self._experiences.values())
|
||
|
||
def search(
|
||
self,
|
||
category: Optional[str] = None,
|
||
tags: Optional[List[str]] = None,
|
||
query: Optional[str] = None,
|
||
limit: int = 10,
|
||
) -> List[Experience]:
|
||
results = list(self._experiences.values())
|
||
|
||
if category:
|
||
results = [e for e in results if e.category == category]
|
||
|
||
if tags:
|
||
results = [e for e in results
|
||
if any(t in e.tags for t in tags)]
|
||
|
||
if query:
|
||
q_lower = query.lower()
|
||
results = [e for e in results
|
||
if q_lower in e.summary.lower()
|
||
or q_lower in (e.evidence or "").lower()]
|
||
|
||
return results[:limit]
|
||
|
||
def delete(self, experience_id: str) -> bool:
|
||
if experience_id in self._experiences:
|
||
del self._experiences[experience_id]
|
||
self._save()
|
||
return True
|
||
return False
|
||
|
||
def count(self) -> int:
|
||
return len(self._experiences)
|
||
|
||
|
||
class ExperienceDistiller:
|
||
"""经验蒸馏器"""
|
||
|
||
# 模式关键词映射
|
||
PATTERNS = {
|
||
"pitfall": [
|
||
r"bug|error|fail|broken|crash|wrong|incorrect",
|
||
r"doesn'?t work|not working|issue",
|
||
r"forgot to|missed|overlooked",
|
||
r"陷阱|踩坑|错误|失败",
|
||
],
|
||
"best_practice": [
|
||
r"should|recommend|best practice|always|never",
|
||
r"tip|trick|pro tip|remember",
|
||
r"最佳实践|建议|推荐|必须",
|
||
],
|
||
"environment": [
|
||
r"install|configure|setup|deploy|version|compatibility",
|
||
r"环境|配置|安装|部署|版本",
|
||
],
|
||
}
|
||
|
||
def __init__(self, store: Optional[ExperienceStore] = None):
|
||
self.store = store or ExperienceStore()
|
||
|
||
def distill_from_task(
|
||
self,
|
||
task_id: str,
|
||
task_title: str,
|
||
task_type: Optional[str] = None,
|
||
outputs: Optional[List[Dict[str, Any]]] = None,
|
||
review_result: Optional[Dict[str, Any]] = None,
|
||
agent_id: Optional[str] = None,
|
||
) -> List[Experience]:
|
||
"""从任务产出中蒸馏经验"""
|
||
experiences = []
|
||
|
||
# 从 review 结果提取
|
||
if review_result:
|
||
if review_result.get("verdict") == "fail":
|
||
for step_result in review_result.get("results", []):
|
||
if step_result.get("verdict") == "fail":
|
||
exp = Experience(
|
||
category="pitfall",
|
||
summary=f"[{task_title}] {step_result.get('step', '?')}: {step_result.get('details', '')}",
|
||
source_task_id=task_id,
|
||
agent_id=agent_id,
|
||
tags=[task_type or "unknown", "review-failure"],
|
||
)
|
||
self.store.add(exp)
|
||
experiences.append(exp)
|
||
|
||
# 从 suggestions 提取
|
||
for step_result in review_result.get("results", []):
|
||
for suggestion in step_result.get("suggestions", []):
|
||
category = self._classify_text(suggestion)
|
||
exp = Experience(
|
||
category=category,
|
||
summary=suggestion,
|
||
source_task_id=task_id,
|
||
agent_id=agent_id,
|
||
tags=[task_type or "unknown"],
|
||
)
|
||
self.store.add(exp)
|
||
experiences.append(exp)
|
||
|
||
# 从产出文本提取
|
||
if outputs:
|
||
for out in outputs:
|
||
content = out.get("content", "")
|
||
if not content and out.get("path"):
|
||
try:
|
||
content = Path(out["path"]).read_text()
|
||
except Exception:
|
||
pass
|
||
|
||
if content:
|
||
extracted = self._extract_from_text(
|
||
content, task_id, task_title, task_type, agent_id
|
||
)
|
||
for exp in extracted:
|
||
self.store.add(exp)
|
||
experiences.extend(extracted)
|
||
|
||
return experiences
|
||
|
||
def _extract_from_text(
|
||
self,
|
||
text: str,
|
||
task_id: str,
|
||
task_title: str,
|
||
task_type: Optional[str],
|
||
agent_id: Optional[str],
|
||
) -> List[Experience]:
|
||
"""从文本中提取经验（基于段落模式匹配）"""
|
||
experiences = []
|
||
# 简单段落分割
|
||
paragraphs = [p.strip() for p in text.split("\n\n") if p.strip()]
|
||
|
||
for para in paragraphs:
|
||
category = self._classify_text(para)
|
||
if category:
|
||
# 只提取有模式匹配的段落
|
||
exp = Experience(
|
||
category=category,
|
||
summary=para[:200],
|
||
source_task_id=task_id,
|
||
agent_id=agent_id,
|
||
tags=[task_type or "unknown", category],
|
||
evidence=para[:500],
|
||
)
|
||
experiences.append(exp)
|
||
|
||
return experiences
|
||
|
||
def _classify_text(self, text: str) -> Optional[str]:
|
||
"""文本模式分类"""
|
||
text_lower = text.lower()
|
||
for category, patterns in self.PATTERNS.items():
|
||
for pattern in patterns:
|
||
if re.search(pattern, text_lower):
|
||
return category
|
||
return None
|
||
|
||
def recommend(
|
||
self,
|
||
task_type: Optional[str] = None,
|
||
tags: Optional[List[str]] = None,
|
||
query: Optional[str] = None,
|
||
limit: int = 5,
|
||
) -> List[Experience]:
|
||
"""推荐相关经验"""
|
||
all_tags = list(tags or [])
|
||
if task_type:
|
||
all_tags.append(task_type)
|
||
|
||
results = self.store.search(tags=all_tags if all_tags else None,
|
||
query=query, limit=limit)
|
||
|
||
# 按置信度排序
|
||
results.sort(key=lambda e: e.confidence, reverse=True)
|
||
return results[:limit]
|