mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 16:52:00 +00:00
ThearchyHelios
7739340607
- 问数据和个人应用的图放行绑定只读查询等合规工具的调用工具节点，只读和遮罩在这条路径上生效；被治理成写入的工具在只读运行里直接拒；知识库范围在工具层固定（图级默认值、参数名纠错、Agent 都改不了） - 个人应用的刷新和个人蓝图的试运行只有拥有者能读：运行、事件、证据、工件、文件、运行与审批列表、对比、可验证报告、质询、假设分析、提取工作流一并收紧；个人蓝图的试运行不论谁发起都按个人应用约束 - 共享蓝图只给拥有者、对应应用的管理者、管理员，看不到的一律 404；审核副本归审核人；管理员看别人的个人应用只有元数据、可以归档 - 记录表跨版本汇总（只读挂载旧版本、按当前列 UNION ALL、各版本各自固定高水位，缺的版本照实写进印章）；停用前发起的运行可以继续写旧版本；更正沿用原提交人、更正人记审计 - 约定收口：待办种类 review、事件 file.extracted、表单错误 file、模板定义带示例与预置；前端类型一次补齐、去掉临时的局部类型；审核待办卡不再给打不开的链接；读附件事件有中文说法；检查脚本补断言
192 lines
8.3 KiB
Python
192 lines
8.3 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.evidence import ledger_enabled, next_exec
|
||
from app.engine.state import GraphState, template_context
|
||
from app.memory import kb, store
|
||
from app.tools.datasource import run_scope_fixed
|
||
|
||
|
||
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, "「长期记忆」节点没有填写要记住的内容")
|
||
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"不支持的记忆操作：{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, "「知识检索」节点没有填写检索问题")
|
||
|
||
# 遮罩、只读的运行：知识库只认运行定下的（tools/registry.ToolContext.fixed_scope，同调用工具节点、Agent）
|
||
collection = ctx.run.collection if run_scope_fixed(ctx.run) \
|
||
else 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:
|
||
from app.core.errors import not_configured
|
||
|
||
degraded.append(f"重排不可用：{not_configured(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, code="retrieve_degraded")
|
||
|
||
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}}
|
||
if snapshot and ledger_enabled(ctx.run):
|
||
# 证据台账：报告目录按它编 K1…，[[see:K1]] 点开追得到这次检索命中了哪几段
|
||
updates["evidence"] = [{"kind": "retrieval", "node_id": ctx.node.id, "exec": next_exec(state, ctx.node.id),
|
||
"artifact": snapshot, "source": collection, "rows": len(hits)}]
|
||
var_name = ctx.cfg("assign_to", "")
|
||
if var_name:
|
||
updates["vars"] = {var_name: context_text}
|
||
return updates
|