mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-17 09:32:00 +00:00
a027261701
以前只有用户明说「记住」才会写:17 轮有图的对话里只有 4 轮带记忆节点,全在
那个用户主动谈记忆的会话里。于是他说过一次「我叫张三」,换个话题再问就
谁也不知道了。
现在 answer 路径多一步判断:这句话里有没有值得长期记住的东西,有就在主链路
前面加一个 memory(write) 节点。边界比指令本身要紧,所以写死了四类不该记的:
这一轮的问题本身、查出来的数(会过期且能重查)、一次性指令、以及**用户没说过
而你推测出来的**。另外两条是从召回的实际情况倒推的——content 要脱离上下文也
成立(召回时没有这轮对话,「他说他叫张三」读起来不知道「他」是谁),以及
拿不准就不记:记错了污染以后每一次对话,漏记一条用户再说一遍就是了。
**动手前先修了一个前提**:写记忆在界面上是隐形的。memory 节点发的是 info 级
日志,而解码层明确丢弃所有 info(decode.ts),也就是说自动记忆 = 系统背着
用户往一个跨会话、长期有效的地方存关于他的信息,而他看不见。
新增 memory.end 事件,解码层按动作分开渲染:写入把**记了什么**原样显示
(只报「写入 1 条」等于没说,用户没法判断它记的是不是他想让它记的)、召回说
「想起 N 条」、清空标成警告。
真模型验过(deepseek-v4-pro):
「我平时叫小蒋,以后报表金额用万元」
→ content='用户平时叫小蒋,他要求以后报表里的金额统一用万元作单位'
scope 未写、跑完落库、事件里带着内容
→ 换个说法召回得到:「该怎么称呼我」「金额用什么单位」都命中
三条对照确认不滥记:
shop 里有几家店 → 没记
重来 → 走 reply,连图都没建
把上一轮的结果按月份排一下 → 没记
182 lines
7.5 KiB
Python
182 lines
7.5 KiB
Python
from __future__ import annotations
|
||
|
||
import json
|
||
from typing import Any
|
||
|
||
from app.core.events import EventType
|
||
from app.db.base import SessionLocal
|
||
from app.engine.context import NodeContext, NodeError
|
||
from app.engine.state import GraphState, template_context
|
||
from app.memory import kb, store
|
||
|
||
|
||
async def run_memory(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||
"""长期记忆读写节点。
|
||
|
||
把记忆做成画布上的一个节点,而不是藏在 agent 内部,是为了让"什么时候记、
|
||
记什么、记到哪个 scope"变成可见、可调的编排决策。
|
||
"""
|
||
action = ctx.cfg("action", "recall") # recall | write | clear
|
||
scope = ctx.render_str(ctx.cfg("scope", "") or ctx.run.memory_scope, state)
|
||
|
||
async with SessionLocal() as session:
|
||
if action == "recall":
|
||
query = ctx.render_str(ctx.cfg("query", "{{ last_message }}"), state)
|
||
limit = int(ctx.cfg("limit", 5) or 5)
|
||
hits = await store.recall(
|
||
session, scope=scope, query=query, limit=limit, kind=ctx.cfg("kind") or None
|
||
)
|
||
text = "\n".join(f"- {h['content']}" for h in hits)
|
||
result: dict[str, Any] = {
|
||
"memories": hits,
|
||
"count": len(hits),
|
||
"text": text,
|
||
"scope": scope,
|
||
}
|
||
elif action == "write":
|
||
content = ctx.render_str(ctx.cfg("content", ""), state)
|
||
if not content.strip():
|
||
raise NodeError(ctx.node.id, "记忆写入节点没有内容")
|
||
item = await store.remember(
|
||
session,
|
||
scope=scope,
|
||
content=content,
|
||
kind=ctx.cfg("kind", "fact") or "fact",
|
||
importance=float(ctx.cfg("importance", 0.5) or 0.5),
|
||
meta={"run_id": ctx.run.run_id, "node_id": ctx.node.id},
|
||
)
|
||
result = {"id": item.id, "saved": True, "scope": scope, "content": item.content}
|
||
elif action == "clear":
|
||
removed = await store.clear_scope(session, scope)
|
||
result = {"cleared": removed, "scope": scope}
|
||
else:
|
||
raise NodeError(ctx.node.id, f"未知的记忆操作:{action}")
|
||
|
||
# 写入必须看得见:这是往长期记忆里存东西,会影响以后每一次对话。
|
||
# 以前发的是 info 日志,而解码层丢弃所有 info,等于系统背着用户记了东西
|
||
ctx.emit(
|
||
EventType.MEMORY_END,
|
||
action=action,
|
||
scope=scope,
|
||
count=result.get("count", result.get("cleared", 1 if action == "write" else 0)),
|
||
# 记了什么要原样说出来,只报"写了 1 条"等于没说
|
||
content=result.get("content", "")[:300],
|
||
)
|
||
|
||
updates: dict[str, Any] = {"nodes": {ctx.node.id: result}}
|
||
var_name = ctx.cfg("assign_to", "")
|
||
if var_name:
|
||
updates["vars"] = {var_name: result.get("text", result)}
|
||
return updates
|
||
|
||
|
||
async def run_retrieve(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||
"""知识库检索节点。输出既有结构化命中列表,也有拼好的上下文文本。"""
|
||
query = ctx.render_str(ctx.cfg("query", "{{ last_message }}"), state)
|
||
if not query.strip():
|
||
raise NodeError(ctx.node.id, "检索节点没有查询内容")
|
||
|
||
collection = ctx.render_str(ctx.cfg("collection", "") or ctx.run.collection, state)
|
||
limit = int(ctx.cfg("limit", 5) or 5)
|
||
# 不写死 0.5:没配置就跟着 embedder 的能力走(见 embeddings.default_alpha)
|
||
raw_alpha = ctx.cfg("alpha")
|
||
alpha = float(raw_alpha) if raw_alpha not in (None, "") else None
|
||
|
||
# 要重排就先多捞一些:重排只能在初筛给出的候选里挑,初筛只给 5 条的话
|
||
# 它最多把这 5 条换个顺序,第 6 条是对的也救不回来
|
||
mode = ctx.cfg("rerank", "off") or "off"
|
||
fetch = max(limit, 20) if mode == "model" else limit
|
||
|
||
degraded: list[str] = []
|
||
async with SessionLocal() as session:
|
||
hits = await kb.search(
|
||
session, collection=collection, query=query, limit=fetch, alpha=alpha,
|
||
on_degrade=degraded.append,
|
||
)
|
||
|
||
if mode == "model" and hits:
|
||
from app.engine.nodes.llm import _model_spec
|
||
from app.memory.rerank import rerank
|
||
from app.providers.factory import ProviderNotConfigured, get_chat_model
|
||
|
||
# 重排有自己的预算,不跟节点上那个走:
|
||
#
|
||
# 它的输出只有一行几十字的 JSON,但模型为了读完 20 条候选会花掉上千个
|
||
# 输出 token——实测 deepseek-v4-pro 单次用 800~1600 个,一旦顶到上限,
|
||
# 正文就什么都不剩(失败那次 out_tokens 正好等于 max_tokens)。
|
||
#
|
||
# thinking=off 对 Anthropic 有效;openai_compatible 那条路径不读这个字段
|
||
# (factory.py 只在 anthropic 分支里设 thinking),所以真正兜住的是
|
||
# 上面那个预算。打分任务本来也不需要思考模式。
|
||
spec = _model_spec(ctx).model_copy(update={
|
||
"max_tokens": max(int(ctx.cfg("max_tokens") or 0), 4096),
|
||
"thinking": "off",
|
||
})
|
||
try:
|
||
async with SessionLocal() as session:
|
||
model, _ = await get_chat_model(session, spec)
|
||
except ProviderNotConfigured as e:
|
||
degraded.append(f"重排用不了:{e}")
|
||
model = None
|
||
hits = await rerank(model, query, hits, top_n=limit, on_note=degraded.append)
|
||
else:
|
||
hits = hits[:limit]
|
||
for note in degraded:
|
||
# 检索悄悄少了一半能力,不说的话用户只会觉得"最近搜得不准"
|
||
ctx.emit(EventType.LOG, level="warn", message=note)
|
||
|
||
min_score = float(ctx.cfg("min_score", 0.0) or 0.0)
|
||
hits = [h for h in hits if h["score"] >= min_score]
|
||
|
||
# 拼成可以直接塞进 prompt 的上下文,带编号方便模型引用出处
|
||
context_text = "\n\n".join(
|
||
f"[{i + 1}] {h['title']}(片段 {h['ordinal']})\n{h['content']}"
|
||
for i, h in enumerate(hits)
|
||
)
|
||
# 命中落工件库。SQL 取数、工具调用、agent 工具调用都有快照,检索一直没有,
|
||
# 于是结论只要来自知识库,"这个数是哪来的"这条链就断在这儿
|
||
from app.core.artifact_store import put_json
|
||
from app.memory.embeddings import default_alpha, embedder_id
|
||
|
||
snapshot: str | None = None
|
||
try:
|
||
snapshot = await put_json(
|
||
{
|
||
"query": query, "collection": collection,
|
||
"alpha": alpha if alpha is not None else default_alpha(),
|
||
"min_score": min_score,
|
||
"embedder": embedder_id(), "rerank": mode, "degraded": degraded,
|
||
"hits": hits,
|
||
},
|
||
kind="retrieval_snapshot",
|
||
run_id=ctx.run.run_id, node_id=ctx.node.id,
|
||
)
|
||
except Exception: # noqa: BLE001 - 存不下不影响这次检索本身
|
||
snapshot = None
|
||
|
||
ctx.emit(
|
||
EventType.RETRIEVE_END,
|
||
collection=collection,
|
||
query=query[:200],
|
||
count=len(hits),
|
||
top_score=hits[0]["score"] if hits else 0,
|
||
reranked=mode == "model",
|
||
degraded=bool(degraded),
|
||
artifact=snapshot,
|
||
)
|
||
|
||
result = {
|
||
"hits": hits,
|
||
"count": len(hits),
|
||
"text": context_text,
|
||
"query": query,
|
||
"collection": collection,
|
||
"artifact": snapshot,
|
||
}
|
||
|
||
updates: dict[str, Any] = {"nodes": {ctx.node.id: result}}
|
||
var_name = ctx.cfg("assign_to", "")
|
||
if var_name:
|
||
updates["vars"] = {var_name: context_text}
|
||
return updates
|