Skip to content

fix: review learned context before persona writes #190

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
EterUltimate merged 1 commit into from
Jun 9, 2026
Merged

fix: review learned context before persona writes #190

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions services/analysis/ml_analyzer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -441,10 +441,12 @@ async def replay_memory(self, group_id: str, new_messages: List[Dict[str, Any]],
await self.temporary_persona_updater.apply_comprehensive_update_to_system_prompt(
group_id, insights_data
)
logger.info(f"成功将强化学习结果集成到system_prompt: {group_id}")
logger.info(
f"强化学习结果已提交 WebUI 人格审查: {group_id}"
)

except Exception as e:
logger.error(f"集成强化学习结果到system_prompt失败: {e}")
logger.error(f"提交强化学习结果到人格审查失败: {e}")


# 这里可以将 refined_data 传递给 PersonaUpdater 进行人格更新
Expand Down
6 changes: 4 additions & 2 deletions services/analysis/multidimensional_analyzer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -570,10 +570,12 @@ async def analyze_message_context(self, event: AstrMessageEvent, message_text: s
await self.temporary_persona_updater.apply_comprehensive_update_to_system_prompt(
group_id, update_data
)
logger.info(f"成功将多维度分析结果集成到system_prompt: {group_id}")
logger.info(
f"多维度分析结果已提交 WebUI 人格审查: {group_id}"
)

except Exception as e:
logger.error(f"集成分析结果到system_prompt失败: {e}")
logger.error(f"提交分析结果到人格审查失败: {e}")
# 这里不抛出异常，让分析结果正常返回

return analysis_result
Expand Down
59 changes: 4 additions & 55 deletions services/learning/realtime_processor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -257,9 +257,6 @@ async def _process_expression_style_learning(
group_id, limit=5
)
if learned_patterns:
await self._apply_style_to_prompt_temporarily(
group_id, learned_patterns
)
few_shots_content = (
await self._dialog_analyzer.generate_few_shots_dialog(
group_id, message_data_list
Expand All @@ -270,70 +267,22 @@ async def _process_expression_style_learning(
group_id, learned_patterns, few_shots_content
)
logger.info(
f"群组 {group_id} 表达风格学习结果已临时应用到prompt，"
"并已提交人格审查"
f"群组 {group_id} 表达风格学习结果已提交人格审查，"
"等待批准后再写入 begin_dialogs"
)
else:
logger.info(
f"群组 {group_id} 表达风格学习结果已临时应用到prompt"
f"群组 {group_id} 表达风格学习成功，"
"但未生成可审查的 few-shot 示例，跳过人格写入"
)
except Exception as e:
logger.error(f"处理表达风格学习结果失败: {e}")

self._learning_stats.style_updates += 1

if self.update_system_prompt_callback:
await self.update_system_prompt_callback(group_id)
logger.info(
f"群组 {group_id} 表达风格学习结果已应用到system_prompt"
)

except Exception as e:
logger.error(f"群组 {group_id} 表达风格学习处理失败: {e}")

# Temporary style application — injected as begin_dialogs example conversations

async def _apply_style_to_prompt_temporarily(
self, group_id: str, learned_patterns: List[Any]
) -> None:
"""Inject learned style patterns as begin_dialogs example conversations."""
try:
if not learned_patterns:
return

dialog_pairs: list = []
for pattern in learned_patterns[:5]:
situation = (
pattern.situation
if hasattr(pattern, "situation")
else pattern.get("situation", "")
)
expression = (
pattern.expression
if hasattr(pattern, "expression")
else pattern.get("expression", "")
)
if situation and expression:
dialog_pairs.append((situation, expression))

if not dialog_pairs:
return

success = await self._temporary_persona_updater.apply_style_as_begin_dialogs(
group_id, dialog_pairs
)

if success:
logger.info(
f"群组 {group_id} 风格示例已注入 begin_dialogs，"
f"包含 {len(dialog_pairs)} 组示例对话"
)
else:
logger.warning(f"群组 {group_id} begin_dialogs 风格注入失败")

except Exception as e:
logger.error(f"追加 few-shot 风格示例失败: {e}")

# Helpers

@staticmethod
Expand Down
224 changes: 187 additions & 37 deletions services/persona/temporary_persona_updater.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,13 +5,15 @@
import json
import time
import asyncio
import hashlib
from typing import Dict, List, Optional, Any, Tuple
from datetime import datetime, timedelta

from astrbot.api import logger
from astrbot.api.star import Context

from ...config import PluginConfig
from ...constants import UPDATE_TYPE_PROGRESSIVE_PERSONA_LEARNING

from ...core.interfaces import IPersonaUpdater, IPersonaBackupManager

Expand Down Expand Up @@ -1268,9 +1270,138 @@ async def _apply_context_awareness_update_to_system_prompt(self, group_id: str,
logger.error(f"直接更新上下文感知到system prompt失败: {e}")
return False

@staticmethod
def _format_review_value(value: Any) -> str:
"""Convert runtime context values to readable review text."""
if value is None:
return ""
if isinstance(value, str):
return value.strip()
if isinstance(value, (list, tuple, set)):
return "、".join(str(item).strip() for item in value if str(item).strip())
if isinstance(value, dict):
return json.dumps(value, ensure_ascii=False, sort_keys=True)
return str(value).strip()

@staticmethod
def _append_text_block(base_text: str, appended_text: str) -> str:
base = (base_text or '').strip()
appended = (appended_text or '').strip()
if base and appended:
return f"{base}\n\n{appended}"
return appended or base

def _build_comprehensive_review_content(self, update_data: dict) -> str:
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M')
sections: List[str] = []

def add_section(title: str, fields: List[Tuple[str, Any]], instruction: str) -> None:
lines = [f"【{title} - {timestamp}】"]
for label, value in fields:
formatted = self._format_review_value(value)
if formatted:
lines.append(f"• {label}: {formatted}")
if len(lines) == 1:
return
lines.append(instruction)
sections.append("\n".join(lines))

if 'mood' in update_data:
add_section(
'当前情绪状态',
[('情绪描述', update_data.get('mood'))],
'• 请根据此情绪调整回复的语气和风格，保持语言与心情对应',
)

relationship_info = update_data.get('social_relationship')
if isinstance(relationship_info, dict):
add_section(
'当前社交关系状态',
[
('用户关系', relationship_info.get('user_relationships')),
('群体氛围', relationship_info.get('group_atmosphere')),
('互动风格', relationship_info.get('interaction_style')),
],
'• 请根据当前社交关系状态调整回复方式和互动风格',
)

profile_info = update_data.get('user_profile')
if isinstance(profile_info, dict):
add_section(
'当前用户档案',
[
('用户偏好', profile_info.get('preferences')),
('兴趣爱好', profile_info.get('interests')),
('沟通风格', profile_info.get('communication_style')),
('性格特征', profile_info.get('personality_traits')),
],
'• 请根据用户档案信息调整回复内容和方式以更好地适应用户',
)

insights_info = update_data.get('learning_insights')
if isinstance(insights_info, dict):
add_section(
'当前学习洞察',
[
('交互模式', insights_info.get('interaction_patterns')),
('改进建议', insights_info.get('improvement_suggestions')),
('有效策略', insights_info.get('effective_strategies')),
('学习重点', insights_info.get('learning_focus')),
],
'• 请根据学习洞察调整回复策略和改进交互质量',
)

context_info = update_data.get('context_awareness')
if isinstance(context_info, dict):
add_section(
'当前上下文状态',
[
('当前话题', context_info.get('current_topic')),
('对话状态', context_info.get('conversation_state')),
('近期关注', context_info.get('recent_focus')),
('对话流向', context_info.get('dialogue_flow')),
],
'• 请根据当前上下文状态保持话题连贯性并提供相关回复',
)

return "\n\n".join(sections).strip()

@staticmethod
def _runtime_context_signature(update_data: dict) -> str:
normalized = json.dumps(
update_data or {},
ensure_ascii=False,
sort_keys=True,
default=str,
)
return hashlib.sha256(normalized.encode('utf-8')).hexdigest()

async def _has_pending_runtime_context_review(
self, group_id: str, signature: str, proposed_content: str
) -> bool:
if not self.db_manager or not hasattr(self.db_manager, 'get_pending_persona_learning_reviews'):
return False
try:
pending_reviews = await self.db_manager.get_pending_persona_learning_reviews()
for review in pending_reviews or []:
if review.get('group_id') != group_id:
continue
metadata = review.get('metadata') or {}
if metadata.get('runtime_context_review'):
return True
if review.get('proposed_content') == proposed_content:
return True
except Exception as exc:
logger.debug(f"检查运行时上下文审查重复项失败: {exc}")
return False

async def apply_comprehensive_update_to_system_prompt(self, group_id: str, update_data: dict) -> bool:
"""
综合应用多种类型的增量更新到 system prompt
将多种运行时上下文更新提交到 WebUI 人格审查队列。

这里沿用 MaiBot 的 learn-before-upsert 思路：自动学习阶段只生成
待审查候选，不直接修改 AstrBot 人格。用户在 WebUI 批准后才会
通过 PersonaReviewService 追加到 system_prompt。

Args:
group_id: 群组ID
Expand All @@ -1280,44 +1411,63 @@ async def apply_comprehensive_update_to_system_prompt(self, group_id: str, updat
bool: 是否应用成功
"""
try:
success_count = 0
total_updates = 0

# 应用情绪更新
if 'mood' in update_data:
total_updates += 1
if await self._apply_mood_update_to_system_prompt(group_id, update_data['mood']):
success_count += 1

# 应用社交关系更新
if 'social_relationship' in update_data:
total_updates += 1
if await self._apply_social_relationship_update_to_system_prompt(group_id, update_data['social_relationship']):
success_count += 1

# 应用用户档案更新
if 'user_profile' in update_data:
total_updates += 1
if await self._apply_user_profile_update_to_system_prompt(group_id, update_data['user_profile']):
success_count += 1

# 应用学习洞察更新
if 'learning_insights' in update_data:
total_updates += 1
if await self._apply_learning_insights_update_to_system_prompt(group_id, update_data['learning_insights']):
success_count += 1

# 应用上下文感知更新
if 'context_awareness' in update_data:
total_updates += 1
if await self._apply_context_awareness_update_to_system_prompt(group_id, update_data['context_awareness']):
success_count += 1

logger.info(f"综合更新完成: {success_count}/{total_updates} 项更新成功应用")
return success_count > 0
proposed_content = self._build_comprehensive_review_content(update_data or {})
if not proposed_content:
logger.debug(f"群组 {group_id} 运行时上下文为空，跳过人格审查创建")
return False

signature = self._runtime_context_signature(update_data or {})
if await self._has_pending_runtime_context_review(
group_id, signature, proposed_content
):
logger.info(f"群组 {group_id} 已存在运行时上下文待审查记录，跳过重复创建")
return True

original_prompt = ''
try:
persona = await self._get_framework_persona(group_id)
if persona:
original_prompt = persona.get('prompt', '') or persona.get('system_prompt', '')
except Exception as exc:
logger.debug(f"获取当前人格用于审查快照失败: {exc}")

new_content = self._append_text_block(original_prompt, proposed_content)
metadata = {
'runtime_context_review': True,
'runtime_context_signature': signature,
'update_categories': list((update_data or {}).keys()),
'raw_update_data': update_data or {},
'application_mode': 'append_system_prompt',
'review_flow': 'maibot_web_review_before_apply',
}

if not self.db_manager or not hasattr(self.db_manager, 'add_persona_learning_review'):
logger.warning("数据库管理器不可用，无法创建运行时上下文人格审查记录")
return False

review_id = await self.db_manager.add_persona_learning_review(
group_id=group_id,
proposed_content=proposed_content,
learning_source=UPDATE_TYPE_PROGRESSIVE_PERSONA_LEARNING,
confidence_score=0.7,
raw_analysis="运行时上下文候选已提交 WebUI 审查，批准后才会写入人格",
metadata=metadata,
original_content=original_prompt,
new_content=new_content,
)

if review_id:
logger.info(
f"群组 {group_id} 运行时上下文已提交人格审查 "
f"(ID: {review_id})，等待批准后再写入 system_prompt"
)
return True

logger.warning(f"群组 {group_id} 运行时上下文审查记录创建失败")
return False

except Exception as e:
logger.error(f"综合应用增量更新失败: {e}")
logger.error(f"提交运行时上下文人格审查失败: {e}")
return False

async def _create_mood_backup_persona(self, group_id: str, mood_type: str) -> bool:
Expand Down
Loading
Loading