mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:11:58 +00:00
ThearchyHelios
eddeded2a4
- 第一批（安全加固与生命周期止血）：隔离测试宿主机密钥与 .env 泄漏；修复 SQLite 取消查询挂死、:memory: 连接池报错与 EngineCache 世代竞态；补齐运行与评测删除时的 Checkpointer 线程、Artifact 磁盘文件与无主附件回收；加固三套沙箱路径越界、进程组清理与 Seatbelt 数据目录隔离，封堵 SSRF IPv4-mapped IPv6 元数据绕过与响应体 OOM；修复推理模型多轮 reasoning_content、结构化输出 Token 漏计与缓存折扣计费 - 第二批（数据库/集群/问数提速与首屏分包）：补齐核心表索引与 SQLite WAL 调优，列表接口按需 defer 大列，收敛启动扫描并打通跨进程 slot 槽位唤醒；优化跨源全局选表、哈希桶数字绑定、流式 CSV 导入及多方言分析算子；剥离首屏静态引入的 @xyflow/react 与全局样式（首屏主包降至 461.59 kB），补充流式 rAF 攒批、Store 精细订阅与表单无障碍绑定 - 第三批（架构演进与工程自动化闭环）：记录库超挂接阈值历史版本自动增量归档至单库多表（解除 SQLite 10 库 ATTACH 硬上限）；重构知识库为词面 BM25 与全集合归一化向量矩阵真·双路独立召回并支持游标分批重建；升级多方言感知 SQL 模板词法分析（E'...' / $$...$$ / q'[...]' / # / 嵌套注释）与 Decimal / bool / NaN 参数校验，消除 ledger 层对 api 路由层的反向依赖；拆分 3101 行 CaliberDrafts 页面为 7 个领域子模块，消除检查脚本本机路径硬编码并新增 GitHub Actions CI 门禁与三批回归测试套件
282 lines
10 KiB
Python
282 lines
10 KiB
Python
from __future__ import annotations
|
||
|
||
from datetime import datetime, timezone
|
||
from typing import Any
|
||
|
||
from sqlalchemy import delete, func, select
|
||
from sqlalchemy.ext.asyncio import AsyncSession
|
||
|
||
from app.db.models import MemoryItem
|
||
from app.memory.embeddings import (
|
||
cosine, embed_text, embed_texts, embedder_dim, embedder_id, from_blob, hybrid_rank,
|
||
to_blob, usable,
|
||
)
|
||
|
||
# 判重阈值：余弦相似度高于它才算同一件事。比排名分数严格得多，
|
||
# 因为这里比的是真实语义距离，不是"谁排第一"。
|
||
_DEDUPE_THRESHOLD = 0.92
|
||
|
||
|
||
async def remember(
|
||
session: AsyncSession,
|
||
*,
|
||
scope: str,
|
||
content: str,
|
||
kind: str = "fact",
|
||
meta: dict[str, Any] | None = None,
|
||
importance: float = 0.5,
|
||
dedupe: bool = True,
|
||
) -> MemoryItem:
|
||
"""写一条长期记忆。
|
||
|
||
dedupe 会先看同 scope 下有没有近乎重复的内容 —— agent 很容易把同一件事
|
||
反复记下来，不去重的话记忆库几轮就被同义句灌满了。
|
||
"""
|
||
content = content.strip()
|
||
vec = await embed_text(content)
|
||
|
||
if dedupe:
|
||
# 判重必须看**真实相似度**，不能用 recall 的分数。recall 走 hybrid_rank，
|
||
# 那里的分数是 min-max 排名归一化的结果：排第一的那条在两路信号上都会
|
||
# 被归成 1.0，与它和查询像不像毫无关系。于是 scope 里只要已有两条记忆，
|
||
# 新内容随便撞上一个字排到第一，就会被判成"重复"而丢弃，接口还返回 201
|
||
# 并把那条不相关的旧记忆当成"刚保存的"交回去——静默丢数据。
|
||
stmt = select(MemoryItem).where(MemoryItem.scope == scope)
|
||
if kind:
|
||
stmt = stmt.where(MemoryItem.kind == kind)
|
||
candidates = list((await session.execute(stmt)).scalars())
|
||
best, best_score = None, 0.0
|
||
for row in candidates:
|
||
# 对不上当前 embedder 的直接跳过：维度不同 cosine 会抛，维度碰巧
|
||
# 相同但模型不同则更糟——两个空间之间的相似度是个随机数，
|
||
# 拿它判重会把不相关的旧记忆认成"重复"然后丢掉新内容
|
||
if not usable(row.embed_model, row.embed_dim):
|
||
continue
|
||
other = from_blob(row.embedding, embedder_dim())
|
||
if other is None:
|
||
continue
|
||
score = cosine(vec, other)
|
||
if score > best_score:
|
||
best, best_score = row, score
|
||
if best is not None and best_score > _DEDUPE_THRESHOLD:
|
||
best.importance = max(best.importance, importance)
|
||
best.use_count += 1
|
||
await session.commit()
|
||
await session.refresh(best)
|
||
return best
|
||
|
||
item = MemoryItem(
|
||
scope=scope,
|
||
kind=kind,
|
||
content=content,
|
||
meta=meta or {},
|
||
importance=importance,
|
||
embedding=to_blob(vec),
|
||
embed_model=embedder_id(),
|
||
embed_dim=embedder_dim(),
|
||
)
|
||
session.add(item)
|
||
await session.commit()
|
||
await session.refresh(item)
|
||
return item
|
||
|
||
|
||
async def recall(
|
||
session: AsyncSession,
|
||
*,
|
||
scope: str,
|
||
query: str,
|
||
limit: int = 5,
|
||
kind: str | None = None,
|
||
min_score: float = 0.05,
|
||
on_degrade: Any = None,
|
||
track: bool = True,
|
||
) -> list[dict[str, Any]]:
|
||
"""按混合相关度取回记忆，并按重要性和使用次数做轻微加权。
|
||
|
||
存量向量和当前 embedder 对不上时整体退回纯关键词。以前这里退得悄无声息
|
||
——知识库那边至少会发一条 warn，记忆这边只是"最近想不起事"，而人根本
|
||
不会想到要去重建索引。
|
||
|
||
track=False 只看不记：调试台试一次回忆不该算"被召回过"，否则记忆卡上的
|
||
召回次数会被调试操作抬高，下次排序也跟着偏。
|
||
"""
|
||
stmt = select(MemoryItem).where(MemoryItem.scope == scope)
|
||
if kind:
|
||
stmt = stmt.where(MemoryItem.kind == kind)
|
||
rows = list((await session.execute(stmt)).scalars())
|
||
if not rows:
|
||
return []
|
||
|
||
query_vec = await embed_text(query)
|
||
# 和 kb.search 同一个道理：有对不上的就整体退回关键词，而不是让它们静默沉底
|
||
stale = [r for r in rows if not usable(r.embed_model, r.embed_dim)]
|
||
if stale:
|
||
if on_degrade:
|
||
names = sorted({r.embed_model or "未知模型" for r in stale})
|
||
on_degrade(
|
||
f"记忆域「{scope}」里有 {len(stale)}/{len(rows)} 条向量由"
|
||
f"「{'、'.join(names)}」建立，与当前的「{embedder_id()}」不一致，"
|
||
f"本次已退回纯关键词召回。重建索引后恢复。"
|
||
)
|
||
vectors: list[Any] = [None] * len(rows)
|
||
alpha = 0.0
|
||
else:
|
||
vectors = [from_blob(r.embedding, embedder_dim()) for r in rows]
|
||
alpha = 0.5
|
||
ranked = hybrid_rank(
|
||
query,
|
||
[r.content for r in rows],
|
||
vectors,
|
||
query_vec,
|
||
alpha=alpha,
|
||
)
|
||
|
||
results: list[dict[str, Any]] = []
|
||
picked_rows: list[MemoryItem] = []
|
||
now = datetime.now(timezone.utc)
|
||
for idx, score, parts in ranked:
|
||
row = rows[idx]
|
||
# 重要的、被用得多的记忆略微上浮
|
||
boosted = score * (0.8 + 0.4 * row.importance) * (1 + min(row.use_count, 10) * 0.01)
|
||
if boosted < min_score:
|
||
continue
|
||
picked_rows.append(row)
|
||
results.append(
|
||
{
|
||
"id": row.id,
|
||
"content": row.content,
|
||
"kind": row.kind,
|
||
"meta": row.meta,
|
||
"importance": row.importance,
|
||
"score": round(boosted, 4),
|
||
"signals": {k: round(v, 4) for k, v in parts.items()},
|
||
"created_at": row.created_at.isoformat() if row.created_at else None,
|
||
}
|
||
)
|
||
if len(results) >= limit:
|
||
break
|
||
|
||
if not track:
|
||
return results
|
||
# 记一下被召回过，作为下次排序的信号
|
||
for item in picked_rows:
|
||
item.use_count += 1
|
||
item.last_used_at = now
|
||
await session.commit()
|
||
return results
|
||
|
||
|
||
async def revise(
|
||
session: AsyncSession,
|
||
memory_id: str,
|
||
*,
|
||
content: str | None = None,
|
||
kind: str | None = None,
|
||
importance: float | None = None,
|
||
) -> MemoryItem | None:
|
||
"""原地改一条记忆。找不到返回 None。
|
||
|
||
内容变了向量必须跟着重算：不然召回按的还是旧话，改完的口径搜不出来，
|
||
搜得出来的却是已经被改掉的那句。
|
||
"""
|
||
item = await session.get(MemoryItem, memory_id)
|
||
if not item:
|
||
return None
|
||
if content is not None and content.strip() != item.content:
|
||
item.content = content.strip()
|
||
item.embedding = to_blob(await embed_text(item.content))
|
||
item.embed_model = embedder_id()
|
||
item.embed_dim = embedder_dim()
|
||
if kind is not None:
|
||
item.kind = kind
|
||
if importance is not None:
|
||
item.importance = importance
|
||
await session.commit()
|
||
await session.refresh(item)
|
||
return item
|
||
|
||
|
||
async def forget(session: AsyncSession, memory_id: str) -> bool:
|
||
item = await session.get(MemoryItem, memory_id)
|
||
if not item:
|
||
return False
|
||
await session.delete(item)
|
||
await session.commit()
|
||
return True
|
||
|
||
|
||
async def clear_scope(session: AsyncSession, scope: str) -> int:
|
||
result = await session.execute(delete(MemoryItem).where(MemoryItem.scope == scope))
|
||
await session.commit()
|
||
return result.rowcount or 0
|
||
|
||
|
||
async def list_memories(
|
||
session: AsyncSession, *, scope: str | None = None, limit: int = 200
|
||
) -> list[MemoryItem]:
|
||
stmt = select(MemoryItem).order_by(MemoryItem.created_at.desc(), MemoryItem.id.desc()).limit(limit)
|
||
if scope:
|
||
stmt = stmt.where(MemoryItem.scope == scope)
|
||
return list((await session.execute(stmt)).scalars())
|
||
|
||
|
||
async def stale_count(session: AsyncSession, scope: str | None = None) -> int:
|
||
"""有多少条记忆的向量和当前 embedder 对不上。"""
|
||
from sqlalchemy import or_
|
||
|
||
stmt = select(func.count(MemoryItem.id)).where(
|
||
or_(MemoryItem.embed_model != embedder_id(), MemoryItem.embed_dim != embedder_dim()))
|
||
if scope:
|
||
stmt = stmt.where(MemoryItem.scope == scope)
|
||
return int((await session.execute(stmt)).scalar_one() or 0)
|
||
|
||
|
||
async def reindex(
|
||
session: AsyncSession, scope: str | None = None, *, on_progress: Any = None,
|
||
) -> int:
|
||
"""用当前 embedder 重算记忆的向量，返回重算了多少条。
|
||
|
||
换 embedding 模型之后记忆和知识库一起失效，但在此之前只有知识库有重建
|
||
路径——记忆这边悄悄退回关键词，而且没有任何恢复手段。
|
||
on_progress(已完成, 总数) 每批提交之后回调一次。
|
||
"""
|
||
count_stmt = select(func.count(MemoryItem.id))
|
||
if scope:
|
||
count_stmt = count_stmt.where(MemoryItem.scope == scope)
|
||
total = int((await session.execute(count_stmt)).scalar_one() or 0)
|
||
if total == 0:
|
||
return 0
|
||
|
||
model_id, dim = embedder_id(), embedder_dim()
|
||
from app.memory import kb
|
||
|
||
batch = kb._EMBED_BATCH # 和 kb 那边统一：一次几千条全发给远端接口必被限流
|
||
done = 0
|
||
last_id = ""
|
||
while True:
|
||
stmt = select(MemoryItem).where(MemoryItem.id > last_id).order_by(MemoryItem.id.asc()).limit(batch)
|
||
if scope:
|
||
stmt = stmt.where(MemoryItem.scope == scope)
|
||
part = list((await session.execute(stmt)).scalars())
|
||
if not part:
|
||
break
|
||
last_id = part[-1].id
|
||
vectors = await embed_texts([r.content for r in part])
|
||
for row, vec in zip(part, vectors):
|
||
row.embedding = to_blob(vec)
|
||
row.embed_model = model_id
|
||
row.embed_dim = dim
|
||
await session.commit()
|
||
done += len(part)
|
||
if on_progress:
|
||
on_progress(min(done, total), total)
|
||
return done
|
||
|
||
|
||
async def list_scopes(session: AsyncSession) -> list[dict[str, Any]]:
|
||
rows = await session.execute(
|
||
select(MemoryItem.scope, func.count(MemoryItem.id)).group_by(MemoryItem.scope)
|
||
)
|
||
return [{"scope": scope, "count": count} for scope, count in rows]
|