78ab3e65c2
- 将成员画像能力进一步收敛到插件内部，强化按群启用、后台异步刷新、后台查看的完整链路 - 新增群维度按日批量提取能力：以群为单位按天处理一次，统一提取当天活跃成员的日级画像摘要 - 日级画像输出扩展为更适合长期累计的结构化信号，补充身份线索、技能信号、家庭线索、阶段线索、价值偏好、群内角色、决策风格等字段 - 优化提示词设计，明确要求优先提取可复用、可累计、可验证的行为线索，减少一次性情绪和短期噪声对长期画像的干扰 - 打通日 -> 周 -> 月 -> 最终画像 的分层汇总链路，让后续月度画像直接消费日/周级结构化摘要，而不是重复回扫长窗口原始消息 - 新增/完善画像融合策略：identity_traits、skill_profile、family_profile、life_stage_profile、value_profile 也纳入长期分数累计，不再仅依赖最近一次结果覆盖旧结果 - 将活跃群、活跃成员、辅助消息样本等口径统一调整为只处理已结束日期，避免当天未完结数据进入画像计算 - 调整日级批处理逻辑，默认只处理昨天及更早日期，确保不会处理当天消息 - 修复重复执行时仍然先调用 AI 再跳过的问题，改为先检查当天候选成员是否已完成生成，全部已存在时直接跳过，减少无效 AI 请求和耗时 - 增加群日批处理、周摘要、月摘要、群刷新进度等日志，方便后台定位当前刷新到哪些群、哪些成员、进度如何 - 丰富后台画像展示字段，支持查看更完整的长期画像维度与摘要统计 - 更新插件配置默认值，收敛为近 60 天启动窗口、每日滚动处理与群级日摘要模式 - 补充 message_storage 读取能力，支持按群按日提取消息，为群日批量画像与后续周期汇总提供底层数据支撑
249 lines
16 KiB
Python
249 lines
16 KiB
Python
# -*- coding: utf-8 -*-
|
||
import json
|
||
from typing import Dict, List
|
||
|
||
|
||
class MemberContextPromptBuilder:
|
||
"""成员分层画像提示词构建器"""
|
||
|
||
@staticmethod
|
||
def build_group_daily_digest_prompt(chatroom_id: str, digest_date: str,
|
||
member_labels: List[str], compressed_chat: str) -> str:
|
||
return (
|
||
"你是微信群后台的成员日观察批量提取器。\n"
|
||
"请基于给定的一天群聊记录，识别当天有参与发言的成员，并分别提取中性、克制的成员日观察摘要。\n"
|
||
"不要做心理诊断、隐私猜测、负面评价，不要输出未在候选名单中的成员。\n"
|
||
"你的输出将被后续系统按周、按月持续累积，因此优先提取可复用、可累计、可验证的行为信号，而不是一次性的情绪和玩笑。\n"
|
||
"输出严格 JSON，不要 markdown。\n"
|
||
"{"
|
||
"\"members\":["
|
||
"{"
|
||
"\"wxid\":\"成员wxid\","
|
||
"\"display_name\":\"成员显示名\","
|
||
"\"topics\":[\"主题1\"],"
|
||
"\"identity_clues\":[\"身份线索1\"],"
|
||
"\"skill_signals\":[\"技能信号1\"],"
|
||
"\"family_signals\":[\"家庭线索1\"],"
|
||
"\"life_stage_signals\":[\"阶段线索1\"],"
|
||
"\"value_preferences\":[\"价值偏好1\"],"
|
||
"\"interaction_style\":\"一句中文\","
|
||
"\"message_pattern\":\"一句中文\","
|
||
"\"response_style_hint\":\"一句中文\","
|
||
"\"habit_signals\":[\"信号1\"],"
|
||
"\"engagement_traits\":[\"特征1\"],"
|
||
"\"decision_style\":\"一句中文\","
|
||
"\"social_role\":\"一句中文，描述当天在群中的角色表现\","
|
||
"\"reply_taboos\":[\"避坑1\"],"
|
||
"\"temperament_signal\":\"一句中文，描述当天沟通倾向，必须克制\","
|
||
"\"summary_text\":\"一段不超过100字的日摘要\","
|
||
"\"representative_messages\":[\"原话1\",\"原话2\"],"
|
||
"\"confidence\":0.0"
|
||
"}"
|
||
"]"
|
||
"}\n"
|
||
"要求：\n"
|
||
"1. 只输出当天真正参与发言且能看出明确行为信号的成员；发言极少的人可以不输出。\n"
|
||
"2. 每个成员的 topics、identity_clues、skill_signals、family_signals、life_stage_signals、value_preferences、habit_signals、engagement_traits 最多4个，reply_taboos 最多3个。\n"
|
||
"3. representative_messages 只保留最能代表当天表达方式的短句，最多3条。\n"
|
||
"4. 必须严格使用候选成员列表中的 wxid 和显示名。\n"
|
||
"5. identity_clues、family_signals、life_stage_signals 只能写公开聊天中出现的线索，不可把弱线索写成确定事实。\n"
|
||
"6. skill_signals 重点提炼成员解决问题、提供信息、组织表达、专业能力等信号。\n"
|
||
"7. social_role 只描述当天在群里的角色表现，例如：问题提出者、信息补充者、气氛调节者、组织推进者。\n"
|
||
"8. topics 更偏向持续关注的话题方向；habit_signals 更偏向重复表达或互动习惯；engagement_traits 更偏向参与方式。\n"
|
||
"9. value_preferences 只记录公开表达出的偏好，如效率优先、成本敏感、谨慎验证、乐于助人，不要写抽象大词。\n"
|
||
"10. summary_text 应是后台观察摘要，不要写成对用户说的话。\n"
|
||
f"群ID: {chatroom_id}\n"
|
||
f"日期: {digest_date}\n"
|
||
"候选成员:\n" + "\n".join(member_labels[:80]) + "\n"
|
||
"压缩后的群聊记录:\n" + compressed_chat
|
||
)
|
||
|
||
@staticmethod
|
||
def build_daily_digest_prompt(chatroom_id: str, wxid: str, display_name: str,
|
||
digest_date: str, messages: List[Dict]) -> str:
|
||
lines = []
|
||
for msg in messages[-80:]:
|
||
ts = str(msg.get("timestamp", ""))[11:16]
|
||
content = (msg.get("content") or "").replace("\n", " ").strip()[:180]
|
||
if content:
|
||
lines.append(f"[{ts}] {content}")
|
||
|
||
return (
|
||
"你是微信群后台的成员日观察摘要生成器。\n"
|
||
"请仅基于给定的当日公开聊天记录，提取对后续互动有帮助的中性行为观察。\n"
|
||
"不要做人格诊断、隐私猜测、负面评价，不要脑补群外信息。\n"
|
||
"这些日观察会被后续系统按周、按月持续累积，所以应优先输出长期可验证的行为线索，而不是一次性情绪。\n"
|
||
"输出严格 JSON，不要 markdown。\n"
|
||
"{"
|
||
"\"topics\":[\"主题1\"],"
|
||
"\"identity_clues\":[\"身份线索1\"],"
|
||
"\"skill_signals\":[\"技能信号1\"],"
|
||
"\"family_signals\":[\"家庭线索1\"],"
|
||
"\"life_stage_signals\":[\"阶段线索1\"],"
|
||
"\"value_preferences\":[\"价值偏好1\"],"
|
||
"\"interaction_style\":\"一句中文\","
|
||
"\"message_pattern\":\"一句中文\","
|
||
"\"response_style_hint\":\"一句中文\","
|
||
"\"habit_signals\":[\"信号1\"],"
|
||
"\"engagement_traits\":[\"特征1\"],"
|
||
"\"decision_style\":\"一句中文\","
|
||
"\"social_role\":\"一句中文\","
|
||
"\"reply_taboos\":[\"避坑1\"],"
|
||
"\"temperament_signal\":\"一句中文，描述当天显露的沟通倾向，必须克制\","
|
||
"\"summary_text\":\"一段不超过100字的日摘要\","
|
||
"\"representative_messages\":[\"原话1\",\"原话2\"],"
|
||
"\"confidence\":0.0"
|
||
"}\n"
|
||
"要求：\n"
|
||
"1. topics、identity_clues、skill_signals、family_signals、life_stage_signals、value_preferences、habit_signals、engagement_traits 最多4个，reply_taboos 最多3个。\n"
|
||
"2. temperament_signal 只能写当日可观察到的沟通倾向，不可上升为长期性格判断。\n"
|
||
"3. representative_messages 保留最能代表当天风格的短句，最多3条。\n"
|
||
"4. identity_clues、family_signals、life_stage_signals 只能写线索，不可写成确定事实。\n"
|
||
"5. skill_signals 重点描述专业能力、工具熟练度、信息组织能力、问题解决能力等当天显露出的信号。\n"
|
||
"6. topics 尽量写持续关注方向，避免写一次性插话；habit_signals 只写当天已明显出现的表达或互动习惯。\n"
|
||
"7. value_preferences 只保留公开表达出的判断偏好，如效率优先、先验证再决策、重成本、重稳定。\n"
|
||
"8. summary_text 要像后台备注，不要像对话回复。\n"
|
||
f"成员: {display_name} ({wxid})\n"
|
||
f"群ID: {chatroom_id}\n"
|
||
f"日期: {digest_date}\n"
|
||
"当日消息:\n" + ("\n".join(lines) or "暂无")
|
||
)
|
||
|
||
@staticmethod
|
||
def build_period_digest_prompt(digest_type: str, chatroom_id: str, wxid: str,
|
||
display_name: str, period_key: str, items: List[Dict]) -> str:
|
||
structured_lines = []
|
||
for item in items:
|
||
structured = item.get("structured", {}) or {}
|
||
payload = {
|
||
"period_key": item.get("period_key"),
|
||
"summary_text": item.get("summary_text", ""),
|
||
"topics": structured.get("topics") or structured.get("stable_topics") or structured.get("long_term_topics") or [],
|
||
"identity_clues": structured.get("identity_clues") or structured.get("identity_traits") or [],
|
||
"skill_signals": structured.get("skill_signals") or structured.get("skill_profile") or [],
|
||
"family_signals": structured.get("family_signals") or structured.get("family_profile") or [],
|
||
"life_stage_signals": structured.get("life_stage_signals") or structured.get("life_stage_profile") or [],
|
||
"value_preferences": structured.get("value_preferences") or structured.get("value_profile") or [],
|
||
"habit_signals": structured.get("habit_signals") or structured.get("habit_patterns") or [],
|
||
"engagement_traits": structured.get("engagement_traits") or structured.get("stable_traits") or [],
|
||
"reply_preferences": structured.get("reply_preferences") or structured.get("long_term_reply_preferences") or [],
|
||
"social_role": structured.get("social_role") or structured.get("group_role") or "",
|
||
"decision_style": structured.get("decision_style") or structured.get("decision_profile") or "",
|
||
"temperament_signal": structured.get("temperament_signal") or structured.get("temperament_tendency") or "",
|
||
"recent_state": structured.get("recent_state") or [],
|
||
}
|
||
structured_lines.append(json.dumps(payload, ensure_ascii=False))
|
||
|
||
if digest_type == "weekly":
|
||
schema = (
|
||
"{"
|
||
"\"stable_topics\":[\"主题1\"],"
|
||
"\"identity_traits\":[\"身份特征1\"],"
|
||
"\"skill_profile\":[\"技能画像1\"],"
|
||
"\"family_profile\":[\"家庭线索1\"],"
|
||
"\"life_stage_profile\":[\"阶段线索1\"],"
|
||
"\"value_profile\":[\"价值偏好1\"],"
|
||
"\"stable_traits\":[\"特征1\"],"
|
||
"\"habit_patterns\":[\"习惯1\"],"
|
||
"\"reply_preferences\":[\"偏好1\"],"
|
||
"\"group_role\":\"一句中文\","
|
||
"\"decision_profile\":\"一句中文\","
|
||
"\"recent_state\":[\"状态1\"],"
|
||
"\"temperament_tendency\":\"一句中文\","
|
||
"\"summary_text\":\"一段不超过120字的周摘要\","
|
||
"\"confidence\":0.0"
|
||
"}"
|
||
)
|
||
extra = "请从多个日摘要中提炼本周重复出现的模式，过滤单日噪音。"
|
||
else:
|
||
schema = (
|
||
"{"
|
||
"\"long_term_topics\":[\"主题1\"],"
|
||
"\"identity_traits\":[\"身份特征1\"],"
|
||
"\"skill_profile\":[\"技能画像1\"],"
|
||
"\"family_profile\":[\"家庭线索1\"],"
|
||
"\"life_stage_profile\":[\"阶段线索1\"],"
|
||
"\"value_profile\":[\"价值偏好1\"],"
|
||
"\"stable_traits\":[\"特征1\"],"
|
||
"\"habit_patterns\":[\"习惯1\"],"
|
||
"\"long_term_reply_preferences\":[\"偏好1\"],"
|
||
"\"group_role\":\"一句中文\","
|
||
"\"decision_profile\":\"一句中文\","
|
||
"\"phase_state\":[\"状态1\"],"
|
||
"\"temperament_tendency\":\"一句中文\","
|
||
"\"summary_text\":\"一段不超过140字的月摘要\","
|
||
"\"confidence\":0.0"
|
||
"}"
|
||
)
|
||
extra = "请从多个周摘要中提炼阶段性稳定特征，只有反复出现的模式才能进入长期层。"
|
||
|
||
return (
|
||
f"你是微信群后台的成员{digest_type}摘要生成器。\n"
|
||
f"{extra}\n"
|
||
"不可做心理诊断、负面评价、隐私猜测。输出严格 JSON，不要 markdown。\n"
|
||
f"{schema}\n"
|
||
"要求：\n"
|
||
"1. 所有列表字段最多5项，必须中性克制。\n"
|
||
"2. 只有多个下级摘要反复出现的特征，才允许写进 stable_traits / habit_patterns / long_term_reply_preferences。\n"
|
||
"3. recent_state / phase_state 只描述当前阶段状态，不要冒充长期人格。\n"
|
||
"4. identity_traits、family_profile、life_stage_profile 只能保留反复出现的公开线索，不可编造事实。\n"
|
||
"5. skill_profile 要优先提炼稳定出现的能力、专业方向、擅长处理的问题类型。\n"
|
||
"6. group_role 描述其在群中的长期角色位置，decision_profile 描述其决策与判断风格。\n"
|
||
"7. value_profile 需要优先保留真正反复出现的判断偏好，如效率优先、成本敏感、风险谨慎、愿意分享。\n"
|
||
f"成员: {display_name} ({wxid})\n"
|
||
f"群ID: {chatroom_id}\n"
|
||
f"周期: {period_key}\n"
|
||
"下级摘要:\n" + ("\n".join(structured_lines) or "暂无")
|
||
)
|
||
|
||
@staticmethod
|
||
def build_final_context_prompt(chatroom_id: str, wxid: str, display_name: str,
|
||
monthly_digests: List[Dict], weekly_digests: List[Dict],
|
||
daily_digests: List[Dict]) -> str:
|
||
monthly_lines = [json.dumps(item.get("structured", {}), ensure_ascii=False) for item in monthly_digests[:6]]
|
||
weekly_lines = [json.dumps(item.get("structured", {}), ensure_ascii=False) for item in weekly_digests[:4]]
|
||
daily_lines = [json.dumps(item.get("structured", {}), ensure_ascii=False) for item in daily_digests[:6]]
|
||
|
||
return (
|
||
"你是微信群后台的最终成员交互画像整理器。\n"
|
||
"请结合月级、周级、日级摘要，输出一个既有长期层又有近期层的后台交互画像。\n"
|
||
"不要做敏感推断、心理诊断、隐私猜测。输出严格 JSON，不要 markdown。\n"
|
||
"{"
|
||
"\"activity_level\":\"高活跃|中活跃|低活跃|观察中\","
|
||
"\"message_pattern\":\"一句中文\","
|
||
"\"interaction_style\":\"一句中文\","
|
||
"\"response_style_hint\":\"一句中文\","
|
||
"\"topics_of_interest\":[\"主题1\"],"
|
||
"\"recent_focus\":[\"近期主题1\"],"
|
||
"\"identity_traits\":[\"身份线索1\"],"
|
||
"\"skill_profile\":[\"技能画像1\"],"
|
||
"\"family_profile\":[\"家庭线索1\"],"
|
||
"\"life_stage_profile\":[\"阶段线索1\"],"
|
||
"\"value_profile\":[\"价值偏好1\"],"
|
||
"\"stable_traits\":[\"长期特征1\"],"
|
||
"\"habit_patterns\":[\"习惯1\"],"
|
||
"\"long_term_reply_preferences\":[\"偏好1\"],"
|
||
"\"group_role\":\"一句中文\","
|
||
"\"decision_profile\":\"一句中文\","
|
||
"\"recent_state\":[\"近期状态1\"],"
|
||
"\"temperament_tendency\":\"一句中文\","
|
||
"\"summary_text\":\"一段不超过150字的后台摘要\","
|
||
"\"confidence\":0.0,"
|
||
"\"engagement_traits\":[\"特征1\"],"
|
||
"\"reply_taboos\":[\"避坑1\"]"
|
||
"}\n"
|
||
"要求：\n"
|
||
"1. stable_traits、habit_patterns、long_term_reply_preferences 只从月级和多次重复证据中提取。\n"
|
||
"2. recent_focus、recent_state 更依赖最近周级和日级。\n"
|
||
"3. summary_text 要像后台备注，不要明显暴露在给用户做画像。\n"
|
||
"4. identity_traits、family_profile、life_stage_profile 必须写成公开线索或长期观察，不得伪造事实。\n"
|
||
"5. skill_profile 要尽量覆盖专业能力、问题解决能力、表达组织能力、资源协调能力等维度。\n"
|
||
"6. group_role 要描述其在群中的角色定位，decision_profile 要描述其决策/判断方式。\n"
|
||
"7. 如果月级与周级证据不足，宁可少写，也不要把短期状态写成长期人格。\n"
|
||
f"成员: {display_name} ({wxid})\n"
|
||
f"群ID: {chatroom_id}\n"
|
||
"月级摘要:\n" + ("\n".join(monthly_lines) or "暂无")
|
||
+ "\n周级摘要:\n" + ("\n".join(weekly_lines) or "暂无")
|
||
+ "\n日级摘要:\n" + ("\n".join(daily_lines) or "暂无")
|
||
)
|