✨ feat: WP-04 记忆与知识——候选版本闭环、检索管线与大周天(ADR-011/018/023)

- 迁移 9:memory_candidate(来源/敏感度/作用域/状态检查约束)、memory(resource_version 作 ETag 源、有效期、当前版本指针)、memory_version((memory_id, version) 唯一、取代链)、big_cycle(cycle_ 聚合根、phase_memory、产物版本引用)
- 候选闭环:手动提交;白名单自动批准仅限 auto+fact+低敏感(撤销窗 72h 记录于审计);敏感/高影响默认人工审核;批准同事务写 Memory+Version+Audit+Outbox(契约 4.4)
- 纠正:If-Match 校验 resource_version(缺失/过期 412),创建新版本不原地覆盖(ADR-011)
- 删除:立即阻断召回 + 清除全部版本原文(墓碑不存被删原文,契约 4.4)
- 检索管线:数据库层域/角色预过滤 → 状态/有效期过滤 → 文本匹配 → 响应级召回证据(语义检索待 embedding 端点)
- 大周天最小实现:手动触发 → 后台四阶段(回顾/对照/提议/提交)→ 模型真实评估产出 pending persona_candidate(候选先于写入);无端点 fail-closed
- 测试 60 项全绿(+6):审批流、白名单三例、ETag 正/缺失/过期、检索隔离与召回证据、删除传播与墓碑原文清除、大周天端到端
This commit is contained in:
chenxin committed 2026-09-11 10:54:20 +08:00
1 parent b4242f3844
commit e333dc5ea5
8 files changed
+1069

No files matched your search

+267
View File
@@ -0,0 +1,267 @@
"""记忆域:候选、审批、不可变版本、纠正、删除与检索(WP-04 / ADR-011)。"""
import hashlib
import json
from datetime import UTC, datetime
from typing import Any
from fastapi import HTTPException, status
from sqlalchemy import select
from sqlalchemy.ext.asyncio import AsyncSession
from app.models import (
AuditEvent,
Memory,
MemoryCandidate,
MemoryVersion,
OutboxEvent,
)
# 白名单自动批准:仅自动提取、fact、低敏感(契约 4.2:撤销窗随批准记录)
AUTO_APPROVE_KINDS = {"fact"}
AUTO_APPROVE_SENSITIVITY = {"none", "low"}
def _content_hash(payload: dict[str, Any]) -> str:
canonical = json.dumps(payload, sort_keys=True, ensure_ascii=False)
return hashlib.sha256(canonical.encode("utf-8")).hexdigest()
async def require_memory(db: AsyncSession, domain_id: str, memory_id: str) -> Memory:
memory = await db.get(Memory, memory_id)
if memory is None or memory.owner_personal_domain_id != domain_id:
raise HTTPException(status.HTTP_404_NOT_FOUND, "记忆不存在")
return memory
async def submit_candidate(
db: AsyncSession,
domain_id: str,
body: dict[str, Any],
trace_id: str | None,
) -> tuple[MemoryCandidate, Memory | None]:
"""提交候选;白名单(auto+fact+低敏感)自动批准并直接创建 Memory(撤销窗记录于候选)。"""
kind = body.get("kind", "")
sensitivity = body.get("sensitivity", "none")
payload = body.get("payload", {})
source = body.get("source", "manual")
if not isinstance(payload, dict) or not payload:
raise HTTPException(status.HTTP_422_UNPROCESSABLE_ENTITY, "payload 不能为空")
from datetime import timedelta
window_until = None
auto_approved = (
source == "auto" and kind in AUTO_APPROVE_KINDS and sensitivity in AUTO_APPROVE_SENSITIVITY
)
if auto_approved:
# 撤销窗:白名单自动写入后 72 小时内可一键撤销(产品策略可调)
window_until = datetime.now(UTC) + timedelta(hours=72)
candidate = MemoryCandidate(
owner_personal_domain_id=domain_id,
character_id=body.get("character_id"),
kind=kind,
sensitivity=sensitivity,
realm=body.get("realm", "reality"),
payload=payload,
source=source,
source_ref=body.get("source_ref", {}),
content_hash=_content_hash(payload),
status="pending",
)
db.add(candidate)
await db.flush()
memory = None
if auto_approved:
memory = await _approve_into_memory(db, domain_id, candidate, trace_id, window_until)
else:
db.add(
AuditEvent(
actor_type="user", actor_id=domain_id, action="memory.candidate_submitted",
aggregate_type="memory_candidate", aggregate_id=candidate.id,
payload={"kind": kind, "sensitivity": sensitivity}, trace_id=trace_id,
)
)
await db.flush()
return candidate, memory
async def _approve_into_memory(
db: AsyncSession,
domain_id: str,
candidate: MemoryCandidate,
trace_id: str | None,
window_until: datetime | None = None,
) -> Memory:
"""批准候选:同一事务写 Memory + MemoryVersion(v1) + Audit + Outbox(契约 4.4)。"""
candidate.status = "approved"
candidate.decided_at = datetime.now(UTC)
memory = Memory(
owner_personal_domain_id=candidate.owner_personal_domain_id,
character_id=candidate.character_id,
kind=candidate.kind,
sensitivity=candidate.sensitivity,
realm=candidate.realm,
status="active",
)
db.add(memory)
await db.flush()
version = MemoryVersion(memory_id=memory.id, version=1, payload=candidate.payload)
db.add(version)
await db.flush()
memory.current_version_id = version.id
db.add(
AuditEvent(
actor_type="user", actor_id=domain_id, action="memory.created",
aggregate_type="memory", aggregate_id=memory.id,
payload={
"candidate_id": candidate.id, "auto": window_until is not None,
"window_until": window_until.isoformat() if window_until else None,
},
trace_id=trace_id,
)
)
db.add(
OutboxEvent(
aggregate_type="memory", aggregate_id=memory.id,
event_type="memory.created", payload={"memory_id": memory.id},
)
)
await db.flush()
return memory
async def decide_candidate(
db: AsyncSession,
domain_id: str,
candidate_id: str,
approve: bool,
trace_id: str | None,
) -> tuple[MemoryCandidate, Memory | None]:
"""人工审批:敏感与高影响默认走此路径(契约:候选先于长期写入)。"""
candidate = await db.get(MemoryCandidate, candidate_id)
if candidate is None or candidate.owner_personal_domain_id != domain_id:
raise HTTPException(status.HTTP_404_NOT_FOUND, "候选不存在")
if candidate.status != "pending":
raise HTTPException(status.HTTP_409_CONFLICT, "候选不在待审状态")
if not approve:
candidate.status = "rejected"
candidate.decided_at = datetime.now(UTC)
db.add(
AuditEvent(
actor_type="user", actor_id=domain_id, action="memory.candidate_rejected",
aggregate_type="memory_candidate", aggregate_id=candidate.id, trace_id=trace_id,
)
)
await db.flush()
return candidate, None
memory = await _approve_into_memory(db, domain_id, candidate, trace_id)
return candidate, memory
async def correct_memory(
db: AsyncSession,
domain_id: str,
memory_id: str,
if_match: str | None,
payload: dict[str, Any],
trace_id: str | None,
) -> Memory:
"""纠正:If-Match 校验 resource_version,创建新版本(不原地覆盖)。"""
memory = await require_memory(db, domain_id, memory_id)
expected = str(memory.resource_version)
if if_match is None or if_match.strip('"') != expected:
raise HTTPException(
status.HTTP_412_PRECONDITION_FAILED,
"版本不匹配:请以最新 resource_version 重试",
)
last = (
await db.execute(
select(MemoryVersion.version)
.where(MemoryVersion.memory_id == memory.id)
.order_by(MemoryVersion.version.desc())
.limit(1)
)
).scalar_one()
version = MemoryVersion(
memory_id=memory.id,
version=last + 1,
payload=payload,
supersedes_version_id=memory.current_version_id,
)
db.add(version)
await db.flush()
memory.current_version_id = version.id
memory.resource_version += 1
db.add(
AuditEvent(
actor_type="user", actor_id=domain_id, action="memory.corrected",
aggregate_type="memory", aggregate_id=memory.id,
payload={"new_version": version.version}, trace_id=trace_id,
)
)
await db.flush()
return memory
async def delete_memory(
db: AsyncSession, domain_id: str, memory_id: str, trace_id: str | None
) -> Memory:
"""删除:立即阻止召回;清除全部版本原文(墓碑不存被删原文,契约 4.4)。"""
memory = await require_memory(db, domain_id, memory_id)
memory.status = "deleted"
versions = (
await db.execute(select(MemoryVersion).where(MemoryVersion.memory_id == memory.id))
).scalars().all()
for v in versions:
v.payload = {} # 原文清除:墓碑只留元数据
db.add(
AuditEvent(
actor_type="user", actor_id=domain_id, action="memory.deleted",
aggregate_type="memory", aggregate_id=memory.id, trace_id=trace_id,
)
)
db.add(
OutboxEvent(
aggregate_type="memory", aggregate_id=memory.id,
event_type="memory.deleted", payload={"memory_id": memory.id},
)
)
await db.flush()
return memory
async def search_memories(
db: AsyncSession,
domain_id: str,
character_id: str | None,
query: str | None,
limit: int,
cursor: str | None,
) -> tuple[list[tuple[Memory, MemoryVersion | None]], str | None, bool]:
"""检索管线(最小):数据库层授权预过滤 → 状态/有效期过滤 → 文本匹配 → 召回证据。
语义检索(pgvector)随 embedding 端点接入;当前文本匹配以个人域规模为前提
(域过滤前置 + ix_memory_domain_status_kind 索引)。
"""
stmt = select(Memory, MemoryVersion).outerjoin(
MemoryVersion, Memory.current_version_id == MemoryVersion.id
).where(
Memory.owner_personal_domain_id == domain_id,
Memory.status == "active",
)
if character_id is not None:
stmt = stmt.where(Memory.character_id == character_id)
now = datetime.now(UTC)
stmt = stmt.where(
(Memory.expires_at.is_(None)) | (Memory.expires_at > now),
)
if query:
stmt = stmt.where(MemoryVersion.payload["text"].astext.ilike(f"%{query}%"))
if cursor:
stmt = stmt.where(Memory.id > cursor)
stmt = stmt.order_by(Memory.id.asc()).limit(limit + 1)
rows = list((await db.execute(stmt)).all())
has_more = len(rows) > limit
page: list[tuple[Memory, MemoryVersion | None]] = [(m, v) for m, v in rows[:limit]]
next_cursor = page[-1][0].id if has_more and page else None
return page, next_cursor, has_more