Losita/mai-bot
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: 修复人物画像混入聊天摘要与机器人输出事实的问题

Browse Source
This commit is contained in:
DawnARC
2026-05-07 14:41:57 +08:00
committed by SengokuCola
parent b6808d4b73
commit adda11738e
7 changed files with 369 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
src/A_memorix/core/utils/person_profile_service.py
View File

@@ -340,11 +340,51 @@ class PersonProfileService:
"type": "paragraph",
"score": 1.1,
"content": content[:220],
"metadata": {},
"source": str(row.get("source", "") or source),
"metadata": dict(row.get("metadata", {}) or {}),
}
)
return self._filter_stale_paragraph_evidence(evidence)
@staticmethod
def _source_type_from_source(source: str) -> str:
token = str(source or "").strip()
if token.startswith("chat_summary:"):
return "chat_summary"
if token.startswith("person_fact:"):
return "person_fact"
return ""
def _enrich_paragraph_evidence_metadata(
self,
paragraph_hash: str,
metadata: Dict[str, Any],
) -> Tuple[Dict[str, Any], str]:
merged = dict(metadata or {})
source = str(merged.get("source", "") or "").strip()
try:
paragraph = self.metadata_store.get_paragraph(paragraph_hash)
except Exception:
paragraph = None
if isinstance(paragraph, dict):
paragraph_metadata = paragraph.get("metadata", {}) or {}
if isinstance(paragraph_metadata, dict):
merged = {**paragraph_metadata, **merged}
source = source or str(paragraph.get("source", "") or "").strip()
source_type = str(merged.get("source_type", "") or "").strip() or self._source_type_from_source(source)
if source_type:
merged["source_type"] = source_type
if source:
merged["source"] = source
return merged, source
@staticmethod
def _is_chat_summary_evidence(item: Dict[str, Any]) -> bool:
metadata = item.get("metadata", {}) if isinstance(item.get("metadata"), dict) else {}
source_type = str(metadata.get("source_type", "") or "").strip()
source = str(item.get("source", "") or metadata.get("source", "") or "").strip()
return source_type == "chat_summary" or source.startswith("chat_summary:")
def _filter_stale_paragraph_evidence(
self,
evidence: List[Dict[str, Any]],
@@ -417,7 +457,8 @@ class PersonProfileService:
"type": "paragraph",
"score": 0.0,
"content": str(para.get("content", ""))[:180],
"metadata": {},
"source": str(para.get("source", "") or ""),
"metadata": dict(para.get("metadata", {}) or {}),
}
)
return self._filter_stale_paragraph_evidence(fallback[:top_k])
@@ -443,13 +484,18 @@ class PersonProfileService:
if not h or h in seen_hash:
continue
seen_hash.add(h)
metadata, source = self._enrich_paragraph_evidence_metadata(
h,
dict(getattr(item, "metadata", {}) or {}),
)
evidence.append(
{
"hash": h,
"type": str(getattr(item, "result_type", "")),
"score": float(getattr(item, "score", 0.0) or 0.0),
"content": str(getattr(item, "content", "") or "")[:220],
"metadata": dict(getattr(item, "metadata", {}) or {}),
"source": source,
"metadata": metadata,
}
)
evidence.sort(key=lambda x: x.get("score", 0.0), reverse=True)
@@ -475,7 +521,7 @@ class PersonProfileService:
lines.append(f"记忆特征: {'; '.join(memory_traits[:6])}")
if relation_edges:
lines.append("关系证据:")
lines.append("稳定关系证据:")
for rel in relation_edges[:6]:
s = rel.get("subject", "")
p = rel.get("predicate", "")
@@ -483,9 +529,19 @@ class PersonProfileService:
conf = float(rel.get("confidence", 0.0))
lines.append(f"- {s} {p} {o} (conf={conf:.2f})")
if vector_evidence:
lines.append("向量证据摘要:")
for item in vector_evidence[:4]:
stable_evidence = [item for item in vector_evidence if not self._is_chat_summary_evidence(item)]
recent_interactions = [item for item in vector_evidence if self._is_chat_summary_evidence(item)]
if stable_evidence:
lines.append("稳定人物事实:")
for item in stable_evidence[:4]:
content = str(item.get("content", "")).strip()
if content:
lines.append(f"- {content}")
if recent_interactions:
lines.append("近期相关互动:")
for item in recent_interactions[:2]:
content = str(item.get("content", "")).strip()
if content:
lines.append(f"- {content}")

2
src/A_memorix/core/utils/summary_importer.py
View File

@@ -43,6 +43,7 @@ SUMMARY_PROMPT_TEMPLATE = """
请完成以下任务：
1. **生成总结**：以第三人称或机器人的视角，简洁明了地总结这段对话的主要内容、发生的事件或讨论的主题。
2. **提取实体与关系**：识别并提取对话中提到的重要实体以及它们之间的关系。
3. **区分事实来源**：用户自己明确表达的稳定人物事实可以记录；机器人发言只能作为上下文，不能单独作为用户画像事实来源。
请严格以 JSON 格式输出，格式如下：
{{
@@ -54,6 +55,7 @@ SUMMARY_PROMPT_TEMPLATE = """
}}
注意：总结应具有叙事性，能够作为长程记忆的一部分。直接使用实体的实际名称，不要使用 e1/e2 等代号。
不要把机器人提出的建议、猜测、玩笑、承诺或复述，写成用户的稳定偏好、身份或长期事实。
"""

13
src/person_info/person_info.py
View File

@@ -1,5 +1,5 @@
from datetime import datetime
from typing import Optional, Union
from typing import List, Optional, Union
import hashlib
import json
@@ -506,7 +506,14 @@ class Person:
logger.error(f"同步用户 {self.person_id} 信息到数据库时出错: {e}")
async def store_person_memory_from_answer(person_name: str, memory_content: str, chat_id: str) -> None:
async def store_person_memory_from_answer(
person_name: str,
memory_content: str,
chat_id: str,
*,
evidence_source: str = "user_supported",
evidence_message_ids: Optional[List[str]] = None,
) -> None:
"""将人物事实写入长期记忆系统。
Args:
@@ -569,6 +576,8 @@ async def store_person_memory_from_answer(person_name: str, memory_content: str,
"person_id": person_id,
"person_name": participant_name,
"writeback_source": "memory_flow_service",
"evidence_source": str(evidence_source or "user_supported"),
"evidence_message_ids": evidence_message_ids or [],
},
respect_filter=True,
user_id=session_user_id,

120
src/services/memory_flow_service.py
View File

@@ -84,7 +84,12 @@ class PersonFactWritebackService:
if target_person is None or not target_person.is_known:
return
facts = await self._extract_facts(target_person, reply_text)
user_evidence_messages = self._collect_user_evidence_messages(message, target_person)
if not user_evidence_messages:
return
user_evidence_text = self._format_user_evidence(user_evidence_messages)
facts = await self._extract_facts(target_person, reply_text, user_evidence_text)
if not facts:
return
@@ -104,8 +109,19 @@ class PersonFactWritebackService:
if not person_name:
return
evidence_message_ids = [
str(getattr(item, "message_id", "") or "").strip()
for item in user_evidence_messages
if str(getattr(item, "message_id", "") or "").strip()
]
for fact in facts:
await store_person_memory_from_answer(person_name, fact, session_id)
await store_person_memory_from_answer(
person_name,
fact,
session_id,
evidence_source="user_supported",
evidence_message_ids=evidence_message_ids,
)
def _resolve_target_person(self, message: Any) -> Optional[Person]:
session = getattr(message, "session", None)
@@ -140,22 +156,110 @@ class PersonFactWritebackService:
person = Person(person_id=person_id)
return person if person.is_known else None
async def _extract_facts(self, person: Person, reply_text: str) -> List[str]:
def _collect_user_evidence_messages(self, message: Any, person: Person) -> List[Any]:
session = getattr(message, "session", None)
session_id = str(
getattr(message, "session_id", "")
or getattr(session, "session_id", "")
or ""
).strip()
if not session_id:
return []
evidence: List[Any] = []
seen_ids = set()
reply_to = str(getattr(message, "reply_to", "") or "").strip()
if reply_to:
try:
replies = find_messages(message_id=reply_to, limit=1)
except Exception as exc:
logger.debug("查询人物事实 reply_to 证据失败: %s", exc)
replies = []
evidence.extend(self._filter_target_user_messages(replies, person, seen_ids))
if evidence:
return evidence[:3]
timestamp = self._extract_message_timestamp(message)
try:
candidates = find_messages(
session_id=session_id,
before_time=timestamp,
limit=6,
limit_mode="latest",
filter_bot=True,
)
except Exception as exc:
logger.debug("查询人物事实近期用户证据失败: %s", exc)
return []
return self._filter_target_user_messages(candidates, person, seen_ids)[:3]
@staticmethod
def _extract_message_timestamp(message: Any) -> float | None:
raw_timestamp = getattr(message, "timestamp", None)
if hasattr(raw_timestamp, "timestamp") and callable(raw_timestamp.timestamp):
try:
return float(raw_timestamp.timestamp())
except Exception:
return None
if isinstance(raw_timestamp, (int, float)):
return float(raw_timestamp)
return None
@staticmethod
def _filter_target_user_messages(messages: List[Any], person: Person, seen_ids: set) -> List[Any]:
filtered: List[Any] = []
target_person_id = str(getattr(person, "person_id", "") or "").strip()
for item in messages:
platform = str(getattr(item, "platform", "") or "").strip()
user_info = getattr(getattr(item, "message_info", None), "user_info", None)
user_id = str(getattr(user_info, "user_id", "") or getattr(item, "user_id", "") or "").strip()
if not platform or not user_id or is_bot_self(platform, user_id):
continue
if target_person_id and get_person_id(platform, user_id) != target_person_id:
continue
text = str(getattr(item, "processed_plain_text", "") or "").strip()
if not text:
continue
message_id = str(getattr(item, "message_id", "") or "").strip()
dedup_key = message_id or f"{platform}:{user_id}:{text}"
if dedup_key in seen_ids:
continue
seen_ids.add(dedup_key)
filtered.append(item)
return filtered
@staticmethod
def _format_user_evidence(messages: List[Any]) -> str:
lines: List[str] = []
for item in messages[:3]:
text = str(getattr(item, "processed_plain_text", "") or "").strip()
if text:
lines.append(f"- {text}")
return "\n".join(lines)
async def _extract_facts(self, person: Person, reply_text: str, user_evidence_text: str) -> List[str]:
person_name = str(getattr(person, "person_name", "") or getattr(person, "nickname", "") or person.person_id)
prompt = f"""你要从一条机器人刚刚发送的回复中，提取“关于{person_name}的稳定事实”。
prompt = f"""你要从用户原始发言中提取“关于{person_name}的稳定事实”。
目标人物：{person_name}
用户原始发言证据：
{user_evidence_text}
机器人回复：
{reply_text}
请只提取满足以下条件的事实：
1. 明确是关于目标人物本人的信息。
2. 具有相对稳定性，可以作为长期记忆保存。
3. 用简洁中文陈述句表达。
4. 如果回复是在直接对目标人物说话，出现“你/你的/你自己”时，默认都指目标人物，请先改写成关于目标人物的第三人称事实再输出。
1. 必须能被“用户原始发言证据”直接支持，不能只来自机器人回复。
2. 明确是关于目标人物本人的信息。
3. 具有相对稳定性，可以作为长期记忆保存。
4. 用简洁中文陈述句表达。
5. 如果用户原始发言中出现“我/我的/自己”，默认指目标人物，请先改写成关于目标人物的第三人称事实再输出。
不要提取：
- 机器人的情绪、计划、临时动作、客套话
- 仅由机器人提出的建议、猜测、玩笑、回忆或承诺
- 只适用于当前时刻的短期安排
- 不确定、猜测、反问
- 与目标人物无关的信息