feat: 装配投影删除客户端(记忆失效/销户清理图与向量),并修复画像字段残留

一、装配 projection_cleaner
新增 app/service/projection_cleanup_service.py,并在**组装层**(app/worker/__main__.py)
注入。此前该客户端一直未装配,清理链路只记录降级(skipped_no_client),实际后果是
**销户后图里仍留着偏好关系**——投顾仍能通过关系网络看到这个人。

清理策略是"以画像为准"而不是按标识直接删边:先删掉该记忆对应的长期事实,再重建画像,
最后用对账修复让图跟着画像收敛。这样即使一条记忆影响多条派生边也能删干净,不依赖
"记得它当初投影成了什么"。Milvus 侧按 memory_uuid 删除;集合不存在(语义召回未启用)
时视为无需清理,客户端不可用则如实报告未清理,绝不伪造成功。

放在组装层而不是 runtime 内部兜底:组件内部给默认实现会把"尚未装配"这一事实悄悄盖住,
而"未注入即显式降级并留痕"是 runtime 刻意保留的语义。最初的改法写成内部兜底,被 5 个
既有单测拦下——那些测试是对的,因此改为在入口处注入。

二、顺带修复:事实消失后画像字段残留
实测:让 preference:horizon 记忆失效并清理后,user_facts 与图边都正确清除,但
fin_customer_profile.investment_horizon 仍是"长期(5年以上)"。根因是 rebuild_profile
只写"本轮有新值"的字段;事实被删后该字段没有新值,旧值就留在画像里——于是记忆已经作废,
投顾还能看到一个客户从未授权继续生效的投资期限。

改为:本轮没有对应事实时**清空**该字段;investor_type 例外——问卷是唯一权威,本轮没有
问卷记录时保持原值(该列 NOT NULL,且清空会在重测空档抹掉开户时的等级)。

三、实测
· 让 preference:horizon 失效后清理:cleaned=True,
  detail=profile_rebuilt; graph_cleaned; vector_collection_absent;
· user_facts 2→1;图边由 ['HAS_GOAL','PREFERS'] 变为 ['PREFERS'];
· 画像 investment_horizon 由"长期(5年以上)"变为 None;investor_type 保持 C2、
  risk_tags 不受影响(与 horizon 无关);
· 恢复记忆状态后重建,画像与图边均正确复原;
· ruff 通过、mypy 113 文件无错、unit+contract 447 passed、integration 29 passed。
This commit is contained in:
2026-09-10 21:58:42 +08:00
parent 89f889b350
commit d4836ed4df
4 changed files with 166 additions and 6 deletions
+10 -2
View File
@@ -190,8 +190,16 @@ class ProfileAssemblyService:
"promoted": len(facts),
}
for field in PROFILE_OWNED_FIELDS:
if field in values:
setattr(profile, field, values[field])
if field == "investor_type":
# 问卷是唯一权威:本轮没有问卷记录时**保持原值**(既不写入也不清空)。
# 否则重测前的空档会把开户时的等级抹掉,而该列是 NOT NULL。
if field in values:
setattr(profile, field, values[field])
continue
# 其余字段由本服务独占:本轮没有对应事实即清空。
# 这不只是洁癖——记忆失效后若不清理,画像会留着一个已经作废的投资期限,
# 投顾据此给建议,而客户从未授权这条信息继续生效(实测踩到过)。
setattr(profile, field, values.get(field))
profile.updated_at = now
await self.session.flush()
+136
View File
@@ -0,0 +1,136 @@
"""投影清理:记忆失效或删除时,清掉它在图库与向量库里的派生数据。
**为什么以画像为准,而不是按标识直接删边**:一条记忆可能对应多条派生边(同一 key 的
证据累积会让 tag 取值变化),直接按 tag_key 删容易删不干净或误删。这里走
「删掉该记忆对应的长期事实 → 重建画像 → 对账修复」,让投影跟着权威源收敛——
最终图里剩下的内容必然与画像一致,而不依赖"我记得这条记忆当初投影成了什么"。
**Milvus 向量**按 `memory_uuid` 删除;集合不存在或客户端不可用时**如实报告未清理**,
绝不伪造成功。这与 `_cleanup_projection` 的审计语义一致:审计只记录适配器返回的真实结论。
"""
import logging
from dataclasses import dataclass
from sqlalchemy import delete, select
from app.core.config import get_settings
from app.infrastructure.db import SessionFactory
from app.model.memory import MemoryUnit
from app.model.profile import UserFact
from app.service.profile_assembly_service import ProfileAssemblyService
from app.service.profile_graph_projection_service import ProfileGraphProjectionService
from app.service.relationship_service import RelationshipService
logger = logging.getLogger(__name__)
@dataclass(frozen=True)
class ProjectionCleanupOutcome:
"""与 `app/worker/runtime.py` 的同名结构保持字段一致(该处只按属性取用)。
`cleaned=True` 表示派生数据已确实清理;为假时 `detail` 必须说明真实缺口,
审计会原样记录它。
"""
cleaned: bool
detail: str = ""
class ProjectionCleanupService:
def __init__(self, relationships: RelationshipService | None = None) -> None:
self.relationships = relationships
async def cleanup(self, *, memory_uuid: str, operation: str) -> ProjectionCleanupOutcome:
del operation # 失效与删除对投影的处理一致:都以权威源为准重建
async with SessionFactory() as session:
memory = await session.scalar(
select(MemoryUnit).where(MemoryUnit.memory_uuid == memory_uuid)
)
if memory is None:
# 记忆已经不在库里:投影清理的目标已消失,视为完成(幂等)
return ProjectionCleanupOutcome(True, "memory_not_found")
customer_id = int(memory.customer_id)
fact_key = str(memory.memory_key)
# 该记忆已失效/删除,它对应的事实不应再参与画像
async with SessionFactory() as session, session.begin():
await session.execute(delete(UserFact).where(
UserFact.customer_id == customer_id, UserFact.fact_key == fact_key
))
# 以剩余证据重建画像,再让图跟着画像收敛
details: list[str] = []
try:
async with SessionFactory() as session, session.begin():
await ProfileAssemblyService(session).rebuild(customer_id)
details.append("profile_rebuilt")
except Exception:
logger.warning("profile rebuild failed during cleanup customer_id=%s",
customer_id, exc_info=True)
details.append("profile_rebuild_failed")
graph_ok = await self._cleanup_graph(customer_id, details)
vector_ok = await self._cleanup_vector(memory_uuid, details)
return ProjectionCleanupOutcome(
cleaned=graph_ok and vector_ok and "profile_rebuild_failed" not in details,
detail="; ".join(details),
)
async def _cleanup_graph(self, customer_id: int, details: list[str]) -> bool:
if self.relationships is None:
details.append("graph_client_unavailable")
return False
try:
async with SessionFactory() as session:
outcome = await ProfileGraphProjectionService(
session, self.relationships
).reconcile_customer(customer_id, repair=True)
except Exception:
logger.warning("graph cleanup failed customer_id=%s", customer_id, exc_info=True)
details.append("graph_cleanup_failed")
return False
if outcome.degraded:
details.append(f"graph_degraded:{outcome.reason}")
return False
if not outcome.consistent:
# repair 之后仍不一致:如实报告,不写成清理成功
details.append(
f"graph_still_inconsistent:missing={len(outcome.missing)},"
f"orphaned={len(outcome.orphaned)}"
)
return False
details.append("graph_cleaned")
return True
async def _cleanup_vector(self, memory_uuid: str, details: list[str]) -> bool:
"""删除该记忆的向量。集合不存在(未启用语义召回)时视为无需清理。"""
settings = get_settings()
if not settings.milvus_uri:
details.append("vector_store_not_configured")
return True
try:
from pymilvus import MilvusClient # type: ignore[import-untyped]
client = MilvusClient(
uri=settings.milvus_uri, token=settings.milvus_token or None
)
except Exception:
logger.warning("milvus client unavailable during cleanup", exc_info=True)
details.append("vector_client_unavailable")
return False
try:
if settings.milvus_collection not in set(client.list_collections()):
# 记忆向量集合尚未启用:没有需要清理的派生数据
details.append("vector_collection_absent")
return True
client.delete(
collection_name=settings.milvus_collection,
filter=f'memory_uuid == "{memory_uuid}"',
)
details.append("vector_cleaned")
return True
except Exception:
logger.warning("milvus delete failed memory_uuid=%s", memory_uuid, exc_info=True)
details.append("vector_delete_failed")
return False
+16 -1
View File
@@ -4,6 +4,8 @@ import logging
from app.core.config import get_settings
from app.infrastructure.db import engine
from app.service.agent.bootstrap import get_relationship_service
from app.service.projection_cleanup_service import ProjectionCleanupService
from app.worker.runtime import WorkerRuntime
logger = logging.getLogger(__name__)
@@ -11,7 +13,20 @@ logger = logging.getLogger(__name__)
async def serve(*, once: bool = False) -> None:
settings = get_settings()
runtime = WorkerRuntime(settings=settings)
relationships = get_relationship_service()
# 在**组装层**注入投影删除客户端:记忆失效/销户时清理图库与向量库里的派生数据。
# 放在这里而不是 runtime 内部兜底,是为了保留"未注入即显式降级并留痕"的语义
# (有单测守着这一点),也让"生产装配了什么"在入口处一眼可见。
# 该服务返回自己模块里的 ProjectionCleanupOutcome(字段与 runtime 的同名结构一致),
# 结构化契约成立但名义类型不同,故显式忽略:为此把结构体抽到共享模块会造成
# service 与 worker 两个层次互相导入,不值得。
runtime = WorkerRuntime(
settings=settings,
relationships=relationships,
projection_cleaner=ProjectionCleanupService( # type: ignore[arg-type]
relationships=relationships
),
)
try:
while True:
try:
+4 -3
View File
@@ -79,7 +79,7 @@ class WorkerRuntime:
endpoint_resolver: ExtractionEndpointResolver | None = None,
memory_cache: CacheDeleteAdapter | None = None,
projection_cleaner: ProjectionCleaner | None = None,
relationships: Any | None = None,
relationships: Any = None,
) -> None:
self.factory = factory if factory is not None else get_agent_factory()
self.settings = settings or get_settings()
@@ -97,8 +97,9 @@ class WorkerRuntime:
self.memory_cache: CacheDeleteAdapter | None = (
memory_cache if memory_cache is not None else get_memory_cache_adapter()
)
# Milvus/Neo4j 删除客户端:当前组装层没有提供(bootstrap 只装配召回用的读适配器),
# 因此默认 None = 投影清理显式降级并留痕,绝不写成"删除成功"。
# Milvus/Neo4j 删除客户端:由**组装层**(app/worker/__main__.py)注入生产实现,
# 这里不兜底。组件内部给默认实现会把"尚未装配"这一事实悄悄盖住——而"未注入即显式
# 降级并留痕"是本模块刻意保留的语义(有单测守着),因此默认值保持 None。
self.projection_cleaner = projection_cleaner
# 图关系服务:画像投影用它写入节点与关系(投顾的多跳推荐、风控的关系网络都读它)。
# 默认取生产装配;图库不可用时该值为 None,投影如实降级而不是失败。