159 lines
7.1 KiB
Python
159 lines
7.1 KiB
Python
"""批量补投「向量残留清理」:把**已过期但向量还在**的知识逐条送去清理(走管理接口)。
|
||||
|
|
|
|||
|
|
python tools/purge_expired_knowledge_vectors.py # 只读:列出要清理哪些
|
|||
|
|
python tools/purge_expired_knowledge_vectors.py --apply # 真投(每条一个 HTTP 调用)
|
|||
|
|
python tools/purge_expired_knowledge_vectors.py --apply --wait 30 # 投完等 Worker 清完再复核
|
|||
|
|
|
|||
|
|
## 它解决什么
|
|||
|
|
|
|||
|
|
`fin_knowledge_meta.status = 'expired'` 只说明"这行下架了",**不代表 Milvus 里的向量没了**。
|
|||
|
|
向量删除靠一条 `knowledge.vector_delete_requested` 事件,历史上有一批行是**别的途径**变成
|
|||
|
|
expired 的(早期脚本直接改库、导入侧幂等上线前的重复副本),它们**从来没投过**这条事件:
|
|||
|
|
|
|||
|
|
- 检索侧只过滤 `visibility`、**不看 `status`**(`app/service/knowledge_search_service.py`),
|
|||
|
|
所以这些死向量**继续参与排序**,和同题的活块抢答;
|
|||
|
|
- 而 `DELETE /api/v1/knowledge/{id}` 对已经 expired 的行**刻意返回 404**("不静默成功"),
|
|||
|
|
于是这些行在管理端口上**无路可走** —— 本脚本 + 新端点 `POST /{id}/vector-cleanups`
|
|||
|
|
就是补这个缺口。
|
|||
|
|
|
|||
|
|
## 与对账工具的关系
|
|||
|
|
|
|||
|
|
先用 `python tools/reconcile_knowledge_vectors.py` 看全局,再用本脚本动手。本脚本**只清理
|
|||
|
|
对账认定"确实还有向量"的那些 id**(`stale_vectors_ids`)—— 对一堆早就干净的 expired 行
|
|||
|
|
盲投事件,只会往 `domain_event_outbox` 里灌没有意义的删除事件,让"事件堆积"变成常态噪音。
|
|||
|
|
|
|||
|
|
## 为什么不直接连 Milvus 删
|
|||
|
|
|
|||
|
|
那样会**同时绕过**权限、审计与 Outbox 幂等语义(删除是不可逆动作,必须留痕、必须有唯一入口)。
|
|||
|
|
孤儿向量(MySQL 里连行都没有的那种)没有可挂事件的载体,本脚本**默认不处理**,
|
|||
|
|
只在对账报告里点出来 —— 真要删属于"直接操作向量库"的运维动作,须单独授权。
|
|||
|
|
"""
|
|||
|
|
|
|||
|
|
from __future__ import annotations
|
|||
|
|
|
|||
|
|
import argparse
|
|||
|
|
import importlib.util
|
|||
|
|
import sys
|
|||
|
|
import time
|
|||
|
|
import uuid
|
|||
|
|
from pathlib import Path
|
|||
|
|
from typing import Any
|
|||
|
|
|
|||
|
|
import httpx
|
|||
|
|
|
|||
|
|
if hasattr(sys.stdout, "reconfigure"):
|
|||
|
|
sys.stdout.reconfigure(errors="replace")
|
|||
|
|
|
|||
|
|
ROOT = Path(__file__).resolve().parents[1]
|
|||
|
|
sys.path.insert(0, str(ROOT))
|
|||
|
|
|
|||
|
|
BASE_DEFAULT = "http://127.0.0.1:8000"
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _load_reconcile_module() -> Any:
|
|||
|
|
"""按文件路径加载对账工具(`tools/` 不是包,没有 `__init__.py`,不能 `import tools.x`)。"""
|
|||
|
|
path = ROOT / "tools" / "reconcile_knowledge_vectors.py"
|
|||
|
|
spec = importlib.util.spec_from_file_location("reconcile_knowledge_vectors", path)
|
|||
|
|
if spec is None or spec.loader is None: # pragma: no cover - 只在文件缺失时发生
|
|||
|
|
raise SystemExit(f"找不到对账工具:{path}")
|
|||
|
|
module = importlib.util.module_from_spec(spec)
|
|||
|
|
spec.loader.exec_module(module)
|
|||
|
|
return module
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _headers(token: str) -> dict[str, str]:
|
|||
|
|
return {"Authorization": f"Bearer {token}", "Idempotency-Key": uuid.uuid4().hex}
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _login(client: httpx.Client, base: str, username: str, password: str) -> str:
|
|||
|
|
response = client.post(
|
|||
|
|
f"{base}/api/v1/auth/tokens", json={"username": username, "password": password}
|
|||
|
|
)
|
|||
|
|
if response.status_code != 200:
|
|||
|
|
raise SystemExit(f"登录失败:HTTP {response.status_code} {response.text[:160]}")
|
|||
|
|
return str(response.json()["data"]["access_token"])
|
|||
|
|
|
|||
|
|
|
|||
|
|
def stale_ids_by_collection() -> dict[str, list[str]]:
|
|||
|
|
"""本机对账一次,返回「还有向量的已过期行」:`{集合: [knowledge_id, ...]}`。"""
|
|||
|
|
reconcile = _load_reconcile_module()
|
|||
|
|
report = reconcile.reconcile(
|
|||
|
|
reconcile.mysql_rows(), reconcile.milvus_vectors(sorted(_collections()))
|
|||
|
|
)
|
|||
|
|
return {
|
|||
|
|
name: list(item["detail"]["stale_vectors_ids"])
|
|||
|
|
for name, item in report.get("collections", {}).items()
|
|||
|
|
if item["detail"]["stale_vectors_ids"]
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
|
|||
|
|
def _collections() -> list[str]:
|
|||
|
|
from app.core.knowledge_contracts import ALLOWED_COLLECTIONS
|
|||
|
|
|
|||
|
|
return sorted(ALLOWED_COLLECTIONS)
|
|||
|
|
|
|||
|
|
|
|||
|
|
def main(argv: list[str] | None = None) -> int:
|
|||
|
|
parser = argparse.ArgumentParser(description="补投已过期知识的向量清理(走管理接口)")
|
|||
|
|
parser.add_argument("--base", default=BASE_DEFAULT, help="平台地址")
|
|||
|
|
parser.add_argument("--username", default="admin_t")
|
|||
|
|
parser.add_argument("--password", default="88888888")
|
|||
|
|
parser.add_argument("--apply", action="store_true", help="真正投递(默认只列出要清理的)")
|
|||
|
|
parser.add_argument("--wait", type=int, default=0, metavar="秒",
|
|||
|
|
help="投完等 Worker 消费,再复核一次残留数")
|
|||
|
|
parser.add_argument("--limit", type=int, default=0, help="最多处理多少条(0=不限)")
|
|||
|
|
args = parser.parse_args(argv)
|
|||
|
|
|
|||
|
|
stale = stale_ids_by_collection()
|
|||
|
|
total = sum(len(ids) for ids in stale.values())
|
|||
|
|
if total == 0:
|
|||
|
|
print("对账结果:没有任何「已过期但向量还在」的知识行,无需清理。")
|
|||
|
|
print("(要看全局请跑 python tools/reconcile_knowledge_vectors.py)")
|
|||
|
|
return 0
|
|||
|
|
|
|||
|
|
print(f"对账发现 {total} 条已过期知识仍有向量(检索侧不看 status,它们仍在参与排序):")
|
|||
|
|
targets: list[str] = []
|
|||
|
|
for name, ids in stale.items():
|
|||
|
|
print(f" · {name}: {len(ids)} 条 → {ids[:12]}{' …' if len(ids) > 12 else ''}")
|
|||
|
|
targets.extend(ids)
|
|||
|
|
if args.limit > 0:
|
|||
|
|
targets = targets[: args.limit]
|
|||
|
|
print(f"(--limit {args.limit}:本次只处理前 {len(targets)} 条)")
|
|||
|
|
|
|||
|
|
if not args.apply:
|
|||
|
|
print("\n[dry-run] 未投递。加 --apply 真投(每条一次 POST /api/v1/knowledge/{id}/vector-cleanups)。")
|
|||
|
|
return 0
|
|||
|
|
|
|||
|
|
with httpx.Client(base_url=args.base, timeout=60) as client:
|
|||
|
|
token = _login(client, args.base, args.username, args.password)
|
|||
|
|
ok = 0
|
|||
|
|
failed: list[tuple[str, str]] = []
|
|||
|
|
for knowledge_id in targets:
|
|||
|
|
response = client.post(
|
|||
|
|
f"{args.base}/api/v1/knowledge/{knowledge_id}/vector-cleanups",
|
|||
|
|
headers=_headers(token),
|
|||
|
|
)
|
|||
|
|
if response.status_code == 200:
|
|||
|
|
ok += 1
|
|||
|
|
else:
|
|||
|
|
failed.append((knowledge_id, f"HTTP {response.status_code} {response.text[:120]}"))
|
|||
|
|
print(f"\n投递完成:成功 {ok} / 失败 {len(failed)}")
|
|||
|
|
for knowledge_id, reason in failed[:10]:
|
|||
|
|
print(f" ✗ {knowledge_id}: {reason}")
|
|||
|
|
|
|||
|
|
if args.wait > 0:
|
|||
|
|
print(f"等 {args.wait} 秒让 Worker 消费删除事件 …")
|
|||
|
|
time.sleep(args.wait)
|
|||
|
|
remaining = stale_ids_by_collection()
|
|||
|
|
left = sum(len(ids) for ids in remaining.values())
|
|||
|
|
print(f"复核:仍有向量的已过期行 {left} 条"
|
|||
|
|
+ ("" if left == 0 else f" → { {k: v[:8] for k, v in remaining.items()} }"))
|
|||
|
|
if left > 0:
|
|||
|
|
print(" 提示:确认 Worker 在跑(python -m app.worker),并看 domain_event_outbox 里"
|
|||
|
|
" knowledge.vector_delete_requested 是否卡在 failed/dead。")
|
|||
|
|
return 0
|
|||
|
|
|
|||
|
|
|
|||
|
|
if __name__ == "__main__":
|
|||
|
|
raise SystemExit(main())
|