liuwei_x_x/abot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

增强成员锐评的历史画像分析

Browse Source 
1. 为锐评插件新增近60天成员历史画像窗口，汇总周/月/少量日摘要形成长期人设。

2. 新增近60天群历史总结背景汇总，让锐评同时参考群长期语境与近期发言。

3. 调整提示词，明确要求模型结合当前画像与历史画像的反差进行分析。
This commit is contained in:
liuwei
2026-04-27 14:39:00 +08:00
parent f783e8ba0c
commit 290e44986e
2 changed files with 190 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
plugins/member_roast/config.toml
View File

@@ -27,6 +27,13 @@ message_limit = 50
min_message_count = 8
context_stale_hours = 24
name_match_min_chars = 2
# 历史画像窗口：
# 1. 用户要求明确使用“历史两个月画像 + 当前本人画像”；
# 2. 这里统一按 60 天窗口汇总成员历史摘要与群历史总结；
# 3. 只抽取结构化关键信息，不会把全部原文暴力塞给模型。
history_profile_days = 60
history_member_digest_limit = 16
history_group_summary_limit = 10
[style]
# 文案风格参数：

183
plugins/member_roast/main.py
View File

@@ -11,8 +11,10 @@ from base.plugin_common.message_plugin_interface import MessagePluginInterface
from base.plugin_common.plugin_interface import PluginStatus
from db.connection import DBConnectionManager
from db.contacts_db import ContactsDBOperator
from db.member_digest_db import MemberDigestDBOperator
from db.member_context_db import MemberContextDBOperator
from db.message_storage import MessageStorageDB
from db.message_summary_db import MessageSummaryDBOperator
from plugins.ai_auto_response.memory.group_memory_profile import GroupMemoryService
from plugins.member_context.service import MemberContextService
from utils.ai.unified_llm import UnifiedLLMClient
@@ -39,8 +41,10 @@ class MemberRoastService:
self.db_manager = db_manager
self.plugin_config = plugin_config or {}
self.contacts_db = ContactsDBOperator(db_manager)
self.member_digest_db = MemberDigestDBOperator(db_manager)
self.member_context_db = MemberContextDBOperator(db_manager)
self.message_db = MessageStorageDB(db_manager)
self.message_summary_db = MessageSummaryDBOperator(db_manager)
self.member_context_service = MemberContextService(db_manager, plugin_config or {})
# 群画像服务本身已经带“按天/按总结刷新”的快照逻辑，直接复用即可。
self.group_memory_service = GroupMemoryService(db_manager, {})
@@ -51,6 +55,9 @@ class MemberRoastService:
self.message_limit = max(int(profile_cfg.get("message_limit", 50) or 50), 1)
self.min_message_count = max(int(profile_cfg.get("min_message_count", 8) or 8), 1)
self.context_stale_hours = max(int(profile_cfg.get("context_stale_hours", 24) or 24), 1)
self.history_profile_days = max(int(profile_cfg.get("history_profile_days", 60) or 60), 1)
self.history_member_digest_limit = max(int(profile_cfg.get("history_member_digest_limit", 16) or 16), 1)
self.history_group_summary_limit = max(int(profile_cfg.get("history_group_summary_limit", 10) or 10), 1)
def build_roast_payload(self, group_id: str, target_wxid: str) -> Tuple[bool, Dict[str, Any]]:
"""构建锐评所需的完整画像载荷。"""
@@ -81,7 +88,9 @@ class MemberRoastService:
}
member_context = self._load_or_refresh_member_context(group_id, target_wxid)
historical_member_profile = self._build_historical_member_profile(group_id, target_wxid)
group_memory_profile = self.group_memory_service.build_group_memory_profile(group_id, group_name=group_name)
historical_group_profile = self._build_historical_group_profile(group_id)
active_dates = self.message_db.get_member_active_dates(group_id, target_wxid, days=min(self.sample_days, 180)) or []
payload = {
@@ -90,7 +99,9 @@ class MemberRoastService:
"target_wxid": target_wxid,
"display_name": display_name,
"member_context": member_context or {},
"historical_member_profile": historical_member_profile or {},
"group_memory_profile": group_memory_profile or {},
"historical_group_profile": historical_group_profile or {},
"recent_messages": recent_messages,
"message_count": len(recent_messages),
"active_days": len(active_dates),
@@ -159,6 +170,151 @@ class MemberRoastService:
continue
return None
def _build_historical_member_profile(self, group_id: str, target_wxid: str) -> Dict[str, Any]:
"""构建近两个月成员历史画像摘要。
数据来源优先级：
1. 成员周/月摘要，代表更稳定的阶段画像；
2. 少量近期日摘要，用来补充“最近几次明显状态变化”；
3. 最终输出压缩成结构化摘要，避免把几十条 digest 原文直接塞给模型。
"""
since_time = datetime.now() - timedelta(days=self.history_profile_days)
weekly_digests = self.member_digest_db.list_digests(group_id, target_wxid, "weekly", limit=self.history_member_digest_limit) or []
monthly_digests = self.member_digest_db.list_digests(group_id, target_wxid, "monthly", limit=max(self.history_member_digest_limit, 8)) or []
daily_digests = self.member_digest_db.list_digests(group_id, target_wxid, "daily", limit=min(self.history_member_digest_limit, 8)) or []
def in_range(item: Dict[str, Any]) -> bool:
end_time = self._safe_parse_datetime(str(item.get("period_end", "") or ""))
if not end_time:
return False
return end_time >= since_time
weekly_digests = [item for item in weekly_digests if in_range(item)]
monthly_digests = [item for item in monthly_digests if in_range(item)]
daily_digests = [item for item in daily_digests if in_range(item)]
focus_keys = [
"long_term_topics", "stable_topics", "topics", "identity_traits", "skill_profile",
"problem_solving_profile", "habit_patterns", "expression_profile", "recent_state",
"engagement_traits", "reply_preferences", "long_term_reply_preferences",
]
phrase_keys = [
"group_role", "temperament_tendency", "decision_profile", "message_pattern",
"interaction_style", "phase_state",
]
list_counter: Dict[str, Dict[str, int]] = {key: {} for key in focus_keys}
phrase_counter: Dict[str, Dict[str, int]] = {key: {} for key in phrase_keys}
timeline: List[str] = []
def push_count(bucket: Dict[str, Dict[str, int]], key: str, value: str, weight: int = 1) -> None:
normalized = str(value or "").strip()
if not normalized:
return
bucket[key][normalized] = bucket[key].get(normalized, 0) + weight
# 月摘要权重大于周摘要，周摘要又高于日摘要。
weighted_sources = [
(monthly_digests, 3),
(weekly_digests, 2),
(daily_digests, 1),
]
for digest_list, weight in weighted_sources:
for item in digest_list:
structured = item.get("structured", {}) or {}
summary_text = str(item.get("summary_text", "") or "").strip()
period_key = str(item.get("period_key", "") or "")
if summary_text:
timeline.append(f"{period_key}: {summary_text[:80]}")
for key in focus_keys:
values = structured.get(key, []) or []
if not isinstance(values, list):
values = [values]
for value in values:
push_count(list_counter, key, str(value or ""), weight=weight)
for key in phrase_keys:
push_count(phrase_counter, key, str(structured.get(key, "") or ""), weight=weight)
def top_items(counter_map: Dict[str, int], limit: int = 4) -> List[str]:
ordered = sorted(counter_map.items(), key=lambda item: (-item[1], len(item[0])))
return [name for name, _ in ordered[:limit]]
history_summary = {
"history_days": self.history_profile_days,
"source_counts": {
"monthly": len(monthly_digests),
"weekly": len(weekly_digests),
"daily": len(daily_digests),
},
"stable_topics": top_items(list_counter["long_term_topics"] or list_counter["stable_topics"] or list_counter["topics"]),
"identity_traits": top_items(list_counter["identity_traits"]),
"skill_profile": top_items(list_counter["skill_profile"]),
"problem_solving_profile": top_items(list_counter["problem_solving_profile"]),
"habit_patterns": top_items(list_counter["habit_patterns"]),
"expression_profile": top_items(list_counter["expression_profile"]),
"recent_state": top_items(list_counter["recent_state"]),
"engagement_traits": top_items(list_counter["engagement_traits"]),
"reply_preferences": top_items(list_counter["reply_preferences"] or list_counter["long_term_reply_preferences"]),
"group_role": top_items(phrase_counter["group_role"], limit=2),
"temperament_tendency": top_items(phrase_counter["temperament_tendency"], limit=2),
"decision_profile": top_items(phrase_counter["decision_profile"], limit=2),
"message_pattern": top_items(phrase_counter["message_pattern"], limit=2),
"interaction_style": top_items(phrase_counter["interaction_style"], limit=2),
"phase_state": top_items(phrase_counter["phase_state"], limit=3),
"timeline": timeline[:8],
}
return history_summary
def _build_historical_group_profile(self, group_id: str) -> Dict[str, Any]:
"""构建近两个月群历史总结摘要。"""
rows = self._load_recent_group_summaries(group_id)
topic_counter: Dict[str, int] = {}
timeline: List[str] = []
for row in rows:
period_key = str(row.get("period_key", "") or "")
summary_text = str(row.get("summary_text", "") or "").strip()
if summary_text:
timeline.append(f"{period_key}: {summary_text[:90]}")
meta = row.get("meta", {}) or {}
for key in ("topics", "focus_topics", "top_topics", "keywords"):
values = meta.get(key, []) or []
if not isinstance(values, list):
values = [values]
for value in values:
normalized = str(value or "").strip()
if not normalized:
continue
topic_counter[normalized] = topic_counter.get(normalized, 0) + 1
ordered_topics = sorted(topic_counter.items(), key=lambda item: (-item[1], len(item[0])))
return {
"history_days": self.history_profile_days,
"summary_count": len(rows),
"focus_topics": [name for name, _ in ordered_topics[:8]],
"timeline": timeline[:8],
}
def _load_recent_group_summaries(self, group_id: str) -> List[Dict[str, Any]]:
"""读取近两个月群总结记录。"""
since_time = datetime.now() - timedelta(days=self.history_profile_days)
sql = """
SELECT *
FROM t_message_summary
WHERE chatroom_id = %s
AND summary_type IN ('daily', 'manual')
AND period_end >= %s
ORDER BY period_end DESC, update_time DESC
LIMIT %s
"""
rows = self.message_summary_db.execute_query(
sql,
(group_id, since_time.strftime("%Y-%m-%d %H:%M:%S"), self.history_group_summary_limit),
) or []
return [self.message_summary_db._deserialize_row(dict(row)) or {} for row in rows]
class MemberRoastPlugin(MessagePluginInterface):
"""成员锐评插件。
@@ -496,6 +652,8 @@ class MemberRoastPlugin(MessagePluginInterface):
"""拼装用户提示词。"""
member_context = payload.get("member_context", {}) or {}
group_memory_profile = payload.get("group_memory_profile", {}) or {}
historical_member_profile = payload.get("historical_member_profile", {}) or {}
historical_group_profile = payload.get("historical_group_profile", {}) or {}
recent_messages = payload.get("recent_messages", []) or []
meta = member_context.get("meta", {}) or {}
group_style = group_memory_profile.get("style_profile", {}) or {}
@@ -540,17 +698,42 @@ class MemberRoastPlugin(MessagePluginInterface):
"气质倾向": meta.get("temperament_tendency", ""),
"近期状态": meta.get("recent_state", []),
},
"成员近两个月历史画像": {
"窗口天数": historical_member_profile.get("history_days", self.history_profile_days),
"摘要来源数": historical_member_profile.get("source_counts", {}),
"稳定主题": historical_member_profile.get("stable_topics", []),
"身份特征": historical_member_profile.get("identity_traits", []),
"技能画像": historical_member_profile.get("skill_profile", []),
"处理问题方式": historical_member_profile.get("problem_solving_profile", []),
"习惯模式": historical_member_profile.get("habit_patterns", []),
"表达标签": historical_member_profile.get("expression_profile", []),
"互动特征": historical_member_profile.get("engagement_traits", []),
"长期回复偏好": historical_member_profile.get("reply_preferences", []),
"长期群内角色": historical_member_profile.get("group_role", []),
"长期气质倾向": historical_member_profile.get("temperament_tendency", []),
"长期发言模式": historical_member_profile.get("message_pattern", []),
"长期互动风格": historical_member_profile.get("interaction_style", []),
"阶段变化轨迹": historical_member_profile.get("phase_state", []),
"历史时间线": historical_member_profile.get("timeline", []),
},
"群聊背景": {
"群名": payload.get("group_name", ""),
"长期主题": group_memory_profile.get("focus_topics", []),
"群风格": group_style,
"群摘要": group_memory_profile.get("summary_text", ""),
},
"群近两个月历史背景": {
"窗口天数": historical_group_profile.get("history_days", self.history_profile_days),
"总结条数": historical_group_profile.get("summary_count", 0),
"历史关注主题": historical_group_profile.get("focus_topics", []),
"历史时间线": historical_group_profile.get("timeline", []),
},
"最近发言样本": recent_lines,
"额外要求": [
"要像熟人看破不说破，不要像机器写分析。",
"可以先抬后杀，也可以连续抓几个典型症状。",
"如果这人明显偏技术、答疑、摸鱼、抽象、嘴硬、复读机、群气氛组，请点出来。",
"必须同时参考“当前成员画像”和“近两个月历史画像”，如果两者有反差，要把这种反差写出来。",
f"发起请求的人是：{requester_name}" if requester_name else "",
],
}