mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 09:11:59 +00:00
ThearchyHelios
8d9a712eae
- 新增标准业务实体（CanonicalEntity）模型、确定性跨源编码翻译、后台候选实体扫描与文档驱动实体智能助手 - 新增指标本体图谱（factor / reconciles_with / upstream_driver）、DAG 防环校验以及 ontology_neighbors / source_health 内置工具 - 新增跨源对账（analysis_reconcile）与滞后共变（analysis_cochange）分析算子，打通 [[h:H1]] 数据源健康证据链、报告导出与本体消融评测 - 数据目录新增「标准业务实体」视图与候选审阅工作台，口径页集成指标本体拓扑面板与端到端浏览器检查
576 lines
34 KiB
Python
576 lines
34 KiB
Python
from __future__ import annotations
|
||
|
||
import logging
|
||
import os
|
||
from typing import Any
|
||
|
||
from sqlalchemy import func, select
|
||
|
||
from app.core.crypto import encrypt
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import Provider, Skill, Workflow, WorkflowVersion
|
||
from app.engine.schema import GraphSpec
|
||
from app.providers import catalog
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
def _n(node_id: str, node_type: str, label: str, **config: Any) -> dict[str, Any]:
|
||
return {
|
||
"id": node_id,
|
||
"type": node_type,
|
||
"position": {"x": 0, "y": 0},
|
||
"data": {"label": label, "config": config},
|
||
}
|
||
|
||
|
||
def _e(source: str, target: str, handle: str | None = None) -> dict[str, Any]:
|
||
edge: dict[str, Any] = {"source": source, "target": target}
|
||
if handle:
|
||
edge["sourceHandle"] = handle
|
||
return edge
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 内置模板：打开就有得玩，也是各种节点用法的活文档
|
||
# --------------------------------------------------------------------------
|
||
|
||
TEMPLATES: list[dict[str, Any]] = [
|
||
{
|
||
"name": "① 最小问答",
|
||
"description": "用一个「模型调用」节点完成从输入到成果的完整流程，适合作为起点。",
|
||
"tags": ["入门"],
|
||
"nodes": [
|
||
_n("start", "input", "提问", fields=[{"name": "question", "required": True,
|
||
"description": "你想问什么"}]),
|
||
_n("answer", "llm", "回答", system="你是一个耐心、准确的助手。",
|
||
prompt="{{ input.question }}", assign_to="answer"),
|
||
_n("done", "output", "成果", fields=[{"name": "回答", "value": "{{ vars.answer }}"}]),
|
||
],
|
||
"edges": [_e("start", "answer"), _e("answer", "done")],
|
||
},
|
||
{
|
||
"name": "② 研究助手（Agent + 工具）",
|
||
"description": "Agent 自主决定搜索和抓取哪些网页，再汇总成带出处的简报。",
|
||
"tags": ["Agent", "工具"],
|
||
"nodes": [
|
||
_n("start", "input", "研究主题", fields=[{"name": "topic", "required": True}]),
|
||
_n("research", "agent", "调研 Agent",
|
||
system="你是严谨的研究员。先搜索再抓取原文，只写有依据的内容。",
|
||
prompt="围绕这个主题做调研：{{ input.topic }}",
|
||
tools=["web_search", "web_fetch", "current_time"],
|
||
max_steps=8, approval="never", assign_to="findings"),
|
||
_n("brief", "llm", "写成简报",
|
||
system="你是一位把复杂信息讲清楚的编辑。",
|
||
prompt="把下面的调研结果整理成中文简报，分「结论 / 关键发现 / 待确认」三段，"
|
||
"每条结论后面标注来源链接：\n\n{{ vars.findings }}",
|
||
assign_to="brief"),
|
||
_n("done", "output", "简报", fields=[
|
||
{"name": "简报", "value": "{{ vars.brief }}"},
|
||
{"name": "调研过程", "value": "{{ nodes.research.tool_calls | json }}"}]),
|
||
],
|
||
"edges": [_e("start", "research"), _e("research", "brief"), _e("brief", "done")],
|
||
},
|
||
{
|
||
"name": "③ 带人工审批的发布流",
|
||
"description": "起草 → 人工审批 → 批准即发布、驳回则改写。演示运行中断与断点恢复。",
|
||
"tags": ["人工审批", "分支"],
|
||
"nodes": [
|
||
_n("start", "input", "主题", fields=[{"name": "topic", "required": True}]),
|
||
_n("draft", "llm", "起草", system="你是品牌文案。",
|
||
prompt="为「{{ input.topic }}」写一条 100 字以内的对外公告。",
|
||
assign_to="draft"),
|
||
_n("review", "human", "人工审批", mode="approve",
|
||
title="是否批准发布这条公告？",
|
||
message="草稿：\n\n{{ vars.draft }}"),
|
||
_n("publish", "transform", "标记发布", mode="template",
|
||
template="✅ 已批准发布\n\n{{ vars.draft }}", assign_to="final"),
|
||
_n("rewrite", "llm", "按意见改写",
|
||
prompt="下面这条公告被驳回了，理由是「{{ nodes.review.note }}」。请据此改写：\n\n{{ vars.draft }}",
|
||
assign_to="final"),
|
||
_n("done", "output", "成果", fields=[
|
||
{"name": "结果", "value": "{{ vars.final }}"},
|
||
{"name": "是否通过", "value": "{{ nodes.review.approved }}"}]),
|
||
],
|
||
"edges": [
|
||
_e("start", "draft"), _e("draft", "review"),
|
||
_e("review", "publish", "approved"), _e("review", "rewrite", "rejected"),
|
||
_e("publish", "done"), _e("rewrite", "done"),
|
||
],
|
||
},
|
||
{
|
||
"name": "④ 知识库问答（RAG）",
|
||
"description": "先检索再回答，检索不到时明确告知。使用前请在知识库页上传文档。",
|
||
"tags": ["RAG", "知识库"],
|
||
"nodes": [
|
||
_n("start", "input", "提问", fields=[{"name": "question", "required": True},
|
||
{"name": "collection", "default": "default"}]),
|
||
_n("search", "retrieve", "检索知识库", query="{{ input.question }}",
|
||
collection="{{ input.collection }}", limit=5, alpha=0.5, assign_to="context"),
|
||
_n("gate", "branch", "是否检索到内容", mode="expression",
|
||
cases=[{"key": "hit", "condition": "nodes.search.count > 0", "label": "有命中"}]),
|
||
_n("answer", "llm", "基于资料回答",
|
||
system="只根据提供的资料回答，每句话后面用 [编号] 标注依据。资料不足就直接说不确定。",
|
||
prompt="资料：\n{{ vars.context }}\n\n问题：{{ input.question }}",
|
||
assign_to="answer"),
|
||
_n("miss", "transform", "没有资料", mode="template",
|
||
template="知识库里没有找到与「{{ input.question }}」相关的内容，请先上传相关文档。",
|
||
assign_to="answer"),
|
||
_n("done", "output", "答案", fields=[
|
||
{"name": "答案", "value": "{{ vars.answer }}"},
|
||
{"name": "引用片段", "value": "{{ nodes.search.hits | json }}"}]),
|
||
],
|
||
"edges": [
|
||
_e("start", "search"), _e("search", "gate"),
|
||
_e("gate", "answer", "hit"), _e("gate", "miss", "default"),
|
||
_e("answer", "done"), _e("miss", "done"),
|
||
],
|
||
},
|
||
{
|
||
"name": "⑤ 代码数据处理（沙箱 + 校验）",
|
||
"description": "模型编写 Python 代码并在沙箱中执行，输出经 JSON Schema 校验，不合格时自动重试。",
|
||
"tags": ["沙箱", "结构化"],
|
||
"nodes": [
|
||
_n("start", "input", "分析需求", fields=[
|
||
{"name": "task", "required": True, "default": "统计 1 到 100 里所有质数的个数与总和"}]),
|
||
_n("write", "llm", "写代码",
|
||
system="你是数据工程师。只输出可直接运行的 Python 代码，不要解释、不要 markdown 围栏。"
|
||
"结果必须用 print(json.dumps(...)) 输出一个 JSON 对象。",
|
||
prompt="任务：{{ input.task }}", assign_to="code"),
|
||
_n("run", "code", "沙箱执行", language="python", code="{{ vars.code }}",
|
||
timeout=30, network=False, fail_fast=False, assign_to="result"),
|
||
_n("check", "validate", "校验结果", source="{{ nodes.run.stdout }}",
|
||
max_retries=2, repair_with_llm=True, fail_fast=False,
|
||
schema={"type": "object", "minProperties": 1}),
|
||
_n("done", "output", "成果", fields=[
|
||
{"name": "结果", "value": "{{ nodes.check.data | json }}"},
|
||
{"name": "代码", "value": "{{ vars.code }}"},
|
||
{"name": "校验通过", "value": "{{ nodes.check.valid }}"}]),
|
||
],
|
||
"edges": [_e("start", "write"), _e("write", "run"), _e("run", "check"), _e("check", "done")],
|
||
},
|
||
{
|
||
"name": "⑥ 多 Agent 协作",
|
||
"description": "调度者按进展分派任务给研究员、批评家和写手，直到目标达成。",
|
||
"tags": ["多 Agent"],
|
||
"nodes": [
|
||
_n("start", "input", "目标", fields=[{"name": "goal", "required": True}]),
|
||
_n("team", "supervisor", "协作团队", goal="{{ input.goal }}", max_rounds=6,
|
||
agents=[
|
||
{"name": "researcher", "description": "查资料、找事实",
|
||
"system": "你是研究员，负责收集事实和数据，给出来源。",
|
||
"tools": ["web_search", "web_fetch"], "max_steps": 4},
|
||
{"name": "critic", "description": "挑毛病、找漏洞",
|
||
"system": "你是批评家。指出前面工作里的事实错误、逻辑漏洞和遗漏，要具体。",
|
||
"tools": [], "max_steps": 2},
|
||
{"name": "writer", "description": "整合成最终文稿",
|
||
"system": "你是写手，把团队成果整合成结构清晰的中文文稿。",
|
||
"tools": [], "max_steps": 2},
|
||
], assign_to="result"),
|
||
_n("done", "output", "成果", fields=[
|
||
{"name": "结果", "value": "{{ vars.result }}"},
|
||
{"name": "协作记录", "value": "{{ nodes.team.transcript | json }}"}]),
|
||
],
|
||
"edges": [_e("start", "team"), _e("team", "done")],
|
||
},
|
||
{
|
||
"name": "⑧ 周报（口径卡 + 三档出具）",
|
||
"description": "用口径卡计算指标、按出具契约三档出具的示例：取数固定，报告只能引用指标清单中的数字，"
|
||
"出具时逐一核对，无法核对的数字会被标出并降档出具。",
|
||
"tags": ["出具", "口径卡", "周报"],
|
||
"nodes": [
|
||
_n("start", "input", "周期", fields=[
|
||
{"name": "week", "default": "2026-W37", "description": "报告周期"}]),
|
||
# 取数节点：产出本身就是源数据（evidence_role=source），口径卡读它记得下出处
|
||
_n("fetch", "code", "取数（数据底座）", language="python", network=False,
|
||
assign_to="agg", evidence_role="source",
|
||
code=(
|
||
"# 固定层：取数与汇总。真实场景换成对数据库/API 的查询，\n"
|
||
"# 结果会作为快照落工件库，出具溯源指回这里。\n"
|
||
"import json\n"
|
||
"print(json.dumps({'orders': 1234, 'amount': 45678.5,\n"
|
||
" 'amount_prev': 42010.0, 'returns': 29}))"
|
||
)),
|
||
_n("caliber", "metrics", "周报口径卡",
|
||
caliber="周报口径", caliber_version="v1", assign_to="m",
|
||
metrics=[
|
||
{"id": "orders", "name": "订单数", "unit": "单",
|
||
"expression": "vars.agg.orders"},
|
||
{"id": "amount", "name": "销售额", "unit": "元",
|
||
"expression": "vars.agg.amount"},
|
||
{"id": "wow", "name": "环比增幅", "unit": "%",
|
||
"expression": "round((vars.agg.amount - vars.agg.amount_prev) / vars.agg.amount_prev * 100, 1)"},
|
||
{"id": "return_rate", "name": "退货率", "unit": "%",
|
||
"expression": "round(vars.agg.returns / vars.agg.orders * 100, 2)"},
|
||
]),
|
||
_n("narrate", "llm", "叙述（关节层）",
|
||
system="你是周报撰写人。只允许引用下面指标清单里出现的数值，"
|
||
"一个清单外的数字都不要写——包括字数统计、序号补充说明等元内容。"
|
||
"只输出周报正文本身，不要任何撰写说明。"
|
||
"你的自由度在于：挑哪些变化值得说、怎么归因。",
|
||
prompt="{{ nodes.caliber.text }}\n\n请为 {{ input.week }} 写一段简短的周报正文。",
|
||
assign_to="report"),
|
||
_n("done", "output", "出具",
|
||
fields=[
|
||
{"name": "周报", "value": "{{ vars.report }}"},
|
||
{"name": "指标清单", "value": "{{ nodes.caliber.text }}"}],
|
||
contract={
|
||
"metrics_from": ["caliber"],
|
||
"narrative": "{{ vars.report }}",
|
||
"required": ["amount", "orders"],
|
||
"expected": ["wow", "return_rate"],
|
||
"allow_numbers": ["37", "120"],
|
||
"strict": False,
|
||
}),
|
||
],
|
||
"edges": [
|
||
_e("start", "fetch"), _e("fetch", "caliber"),
|
||
_e("caliber", "narrate"), _e("narrate", "done"),
|
||
],
|
||
},
|
||
{
|
||
"name": "⑨ 可追溯周报",
|
||
"description": "报告中的每个数字都可追溯出处：「调用工具」节点查询数据，口径卡登记指标，报告撰写节点只写"
|
||
"引用标记，数字由系统填入，成果节点按引用逐段核对后三档出具。使用前请接入名为 shop 的数据源（含 orders "
|
||
"表的 week、amount、refunded 列），或上传同名表格，再按你的口径修改 SQL。报告撰写节点已按受管级别设置核对规则。",
|
||
"tags": ["出具", "口径卡", "周报", "证据"],
|
||
"nodes": [
|
||
_n("start", "input", "周期", fields=[
|
||
{"name": "week", "default": "2026-W37", "description": "报告周期（ISO 周）"},
|
||
{"name": "prev_week", "default": "2026-W36", "description": "对比的上一周期"}]),
|
||
# 聚合放在 SQL 里做完，口径卡只做标量运算
|
||
_n("fetch", "tool", "取数（查库）", tool="db_query__shop", args={"sql": (
|
||
"SELECT\n"
|
||
" SUM(CASE WHEN week = '{{ input.week }}' THEN amount ELSE 0 END) AS gmv,\n"
|
||
" SUM(CASE WHEN week = '{{ input.prev_week }}' THEN amount ELSE 0 END) AS gmv_prev,\n"
|
||
" SUM(CASE WHEN week = '{{ input.week }}' THEN 1 ELSE 0 END) AS order_cnt,\n"
|
||
" SUM(CASE WHEN week = '{{ input.week }}' AND refunded = 1 THEN 1 ELSE 0 END) AS refund_cnt\n"
|
||
"FROM orders")}),
|
||
# 口径卡用 cell(查询节点, 行, 列名) 直接取查询结果的一格：每个数点开都追到这次查询的那一格（查询快照），
|
||
# 列的顺序改了也取不错。不必先经数据整形解析
|
||
_n("caliber", "metrics", "周报口径卡", caliber="周报口径", caliber_version="v1",
|
||
metrics=[
|
||
{"id": "gmv", "name": "销售额", "unit": "元", "decimals": 2, "format": "thousands",
|
||
"expression": "cell(nodes.fetch, 0, 'gmv')"},
|
||
{"id": "orders", "name": "订单数", "unit": "单", "format": "thousands",
|
||
"expression": "cell(nodes.fetch, 0, 'order_cnt')"},
|
||
# 分母为 0 时记为空值（缺输入），交给契约按 expected 降档，而不是整张卡失败
|
||
{"id": "wow", "name": "环比增幅", "unit": "%", "decimals": 1, "format": "plain",
|
||
"expression": "round((cell(nodes.fetch, 0, 'gmv') - cell(nodes.fetch, 0, 'gmv_prev'))"
|
||
" / cell(nodes.fetch, 0, 'gmv_prev') * 100, 1)"
|
||
" if cell(nodes.fetch, 0, 'gmv_prev') else None"},
|
||
{"id": "aov", "name": "客单价", "unit": "元", "decimals": 2, "format": "thousands",
|
||
"expression": "round(cell(nodes.fetch, 0, 'gmv') / cell(nodes.fetch, 0, 'order_cnt'), 2)"
|
||
" if cell(nodes.fetch, 0, 'order_cnt') else None"},
|
||
{"id": "refund_rate", "name": "退款率", "decimals": 4, "format": "percent_of_ratio",
|
||
"expression": "cell(nodes.fetch, 0, 'refund_cnt') / cell(nodes.fetch, 0, 'order_cnt')"
|
||
" if cell(nodes.fetch, 0, 'order_cnt') else None"},
|
||
]),
|
||
# 受管出具要求报告撰写节点写明最严的几项（发布门禁 G3，含分析说法核对设为严格），模板照着写，拿来就能发布成受管
|
||
_n("write", "report", "报告撰写", metrics_from=["caliber"],
|
||
numbers="strict", on_violation="fail", claims="require_citation", analysis_claims="strict",
|
||
system="你是周报撰写人：结论先行，只写有数据支撑的判断。",
|
||
instructions="为 {{ input.week }} 写一份简短的周报：先用一两句话总结本周，"
|
||
"再分点说销售额、订单、客单价和退款的变化。"),
|
||
_n("done", "output", "出具",
|
||
fields=[{"name": "周报", "value": "{{ nodes.write.text }}"}],
|
||
contract={
|
||
"report_from": "write",
|
||
"metrics_from": ["caliber"],
|
||
"required": ["gmv", "orders"],
|
||
"expected": ["wow", "aov", "refund_rate"],
|
||
"strict": True,
|
||
}),
|
||
],
|
||
"edges": [
|
||
_e("start", "fetch"), _e("fetch", "caliber"),
|
||
_e("caliber", "write"), _e("write", "done"),
|
||
],
|
||
},
|
||
{
|
||
"name": "⑦ 批量处理（循环）",
|
||
"description": "对列表逐项处理后汇总，演示「循环」节点「循环体」和「结束」两个出口的连法。",
|
||
"tags": ["循环"],
|
||
"nodes": [
|
||
_n("start", "input", "条目列表", fields=[
|
||
{"name": "items", "required": True,
|
||
"default": ["降低获客成本", "提升续费率", "缩短上线周期"],
|
||
"description": "JSON 数组或每行一条"}]),
|
||
_n("each", "loop", "逐条循环", mode="foreach", items="{{ input.items }}",
|
||
item_var="item", max_iterations=10),
|
||
_n("handle", "llm", "处理单条",
|
||
prompt="针对「{{ vars.item }}」给出一条可执行的建议，不超过 40 字。",
|
||
assign_to="one"),
|
||
_n("collect", "transform", "累积结果", mode="expression",
|
||
expression="(vars.collected or '') + '· ' + str(vars.one) + '\\n'",
|
||
assign_to="collected"),
|
||
_n("done", "output", "汇总", fields=[
|
||
{"name": "逐条建议", "value": "{{ vars.collected }}"},
|
||
{"name": "处理条数", "value": "{{ nodes.each.iteration }}"}]),
|
||
],
|
||
"edges": [
|
||
_e("start", "each"),
|
||
_e("each", "handle", "body"), _e("handle", "collect"), _e("collect", "each"),
|
||
_e("each", "done", "done"),
|
||
],
|
||
},
|
||
{
|
||
# 流程模板（应用平台 §3.7）：图里只写岗位名，不写具体的人——做成应用时在应用里把「财务」「总监」绑到人或组。
|
||
# 不依赖任何外部端点：发布、做成应用、会签、超时升级都能直接演示
|
||
"name": "⑩ 报销审批（会签）",
|
||
"description": "员工发起报销，财务岗位全体同意才通过，1 小时内没处理完升级给总监；任何一人驳回即退回。",
|
||
"tags": ["流程", "人工审批", "会签"],
|
||
"nodes": [
|
||
_n("start", "input", "发起报销", fields=[
|
||
{"name": "applicant", "required": True, "description": "报销人"},
|
||
{"name": "amount", "type": "number", "minimum": 0.01, "required": True, "description": "金额（元）"},
|
||
{"name": "reason", "required": True, "description": "事由"}]),
|
||
# 标题带上金额：同一个人的两笔报销，待办列表里一眼分得开
|
||
_n("review", "human", "财务会签", mode="approve",
|
||
title="报销审批：{{ input.applicant }} · {{ input.amount }} 元",
|
||
message="报销人：{{ input.applicant }}\n金额（元）：{{ input.amount }}\n事由：{{ input.reason }}",
|
||
assignee={"kind": "slot", "slot": "finance"}, policy="all", due="PT1H",
|
||
on_timeout={"action": "escalate", "to_slot": "director"}),
|
||
_n("approved", "output", "已通过", fields=[
|
||
{"name": "结论", "value": "已通过"}, {"name": "会签备注", "value": "{{ nodes.review.note }}"}]),
|
||
_n("rejected", "output", "已退回", fields=[
|
||
{"name": "结论", "value": "已退回"}, {"name": "退回原因", "value": "{{ nodes.review.note }}"}]),
|
||
],
|
||
"edges": [_e("start", "review"), _e("review", "approved", "approved"), _e("review", "rejected", "rejected")],
|
||
},
|
||
{
|
||
# 分析能力一轮（docs/design/analysis-round/00-总方案.md §3.7）：模板定骨架——分析卡固定做哪几项分析写在图里，判定由
|
||
# 系统按平台的方法和阈值算、系统写判定语，报告撰写节点只引用。照模板 ⑨ 的写法，拿来就能按受管级别发布
|
||
"name": "⑪ 异动分析（可追溯）",
|
||
"description": "周销售额的异动分析，每个判断都可追溯：两个「调用工具」节点取数（整体按周、按地区两周），分析卡固定做"
|
||
"两项分析——这一周的变化在不在常规波动范围内、变化主要来自哪个地区（拆开后各段之和和总数核对闭合），报告撰写节点只引用"
|
||
"判定语和数，判断词逐句核对。使用前请接入名为 shop 的数据源（含 orders 表的 order_date、region、amount 列），或上传"
|
||
"同名表格。SQL 按 SQLite 的写法，按周汇总时期次取每周第一天（周一）的日期；接别的库时换成那个库的写法。报告撰写节点已按"
|
||
"受管级别设置核对规则。",
|
||
"tags": ["分析", "出具", "周报", "证据"],
|
||
"nodes": [
|
||
_n("start", "input", "周期", fields=[
|
||
{"name": "week", "default": "2026-10-12", "description": "本周的第一天（周一）"},
|
||
{"name": "prev_week", "default": "2026-10-05", "description": "上周的第一天（周一）"}]),
|
||
# 期次在 SQL 里由日期列算出来（每周第一天的日期）：分析卡据此认出每一期覆盖哪几天，本期过没过完核对得了。
|
||
# 倒序排：第 0 行是本周、第 1 行是上周，变化拆解按这两格核对闭合
|
||
_n("weekly", "tool", "整体按周（26 周）", tool="db_query__shop", args={"sql": (
|
||
"SELECT date(order_date, '-6 days', 'weekday 1') AS \"周\", SUM(amount) AS \"销售额\"\n"
|
||
"FROM orders\n"
|
||
"WHERE order_date >= date('{{ input.week }}', '-175 days')"
|
||
" AND order_date < date('{{ input.week }}', '+7 days')\n"
|
||
"GROUP BY date(order_date, '-6 days', 'weekday 1')\n"
|
||
"ORDER BY \"周\" DESC")}),
|
||
_n("by_region", "tool", "按地区两周", tool="db_query__shop", args={"sql": (
|
||
"SELECT region AS \"地区\", date(order_date, '-6 days', 'weekday 1') AS \"周\", SUM(amount) AS \"销售额\"\n"
|
||
"FROM orders\n"
|
||
"WHERE order_date >= '{{ input.prev_week }}' AND order_date < date('{{ input.week }}', '+7 days')\n"
|
||
"GROUP BY region, date(order_date, '-6 days', 'weekday 1')\n"
|
||
"ORDER BY \"地区\", \"周\"")}),
|
||
# 阈值不写在图里：取设置里的分析一组（发起时快照）；要固定阈值时在这张卡上写「阈值覆盖」
|
||
_n("check", "analysis", "异动分析卡", items=[
|
||
{"id": "wow", "op": "baseline", "from": "weekly",
|
||
"params": {"period": "周", "measure": "销售额", "target": "{{ input.week }}", "grain": "week",
|
||
"label": "销售额"},
|
||
"hypothesis": "本周销售额的变化在不在常规波动范围内"},
|
||
{"id": "where", "op": "decompose", "mode": "additive", "from": "by_region",
|
||
"params": {"by": ["地区"], "period": "周", "base": "{{ input.prev_week }}", "target": "{{ input.week }}",
|
||
"measure": "销售额", "headline": {"base": "weekly.r1.销售额", "target": "weekly.r0.销售额"},
|
||
"label": "销售额"},
|
||
"hypothesis": "本周销售额的变化主要来自哪个地区"},
|
||
]),
|
||
# 受管出具要求报告撰写节点写明最严的几项（发布门禁 G3，含分析说法核对设为严格），模板照着写
|
||
_n("write", "report", "报告撰写", numbers="strict", on_violation="fail", claims="require_citation",
|
||
analysis_claims="strict",
|
||
system="你是经营分析撰写人：结论先行，判断照分析结论写，只写有数据支撑的内容。",
|
||
instructions="为 {{ input.week }} 开始的这一周写一份销售额异动分析：先用一句话说本周销售额和上周比变了多少，"
|
||
"再按分析卡的两项结论说这个变化在不在常规波动范围内、主要来自哪个地区，最后用一张表列出各地区的"
|
||
"变化。判断照分析结论写、直接引用判定语，没有分析结论的判断不下；有一项没有做成的，照实写原因。"),
|
||
# 图里没有口径卡：报告直接引用查询和分析结论的格，契约开「单元格引用」（受管级别要写明）
|
||
_n("done", "output", "出具",
|
||
fields=[{"name": "异动分析", "value": "{{ nodes.write.text }}"}],
|
||
contract={"report_from": "write", "cells": True, "strict": True}),
|
||
],
|
||
"edges": [
|
||
_e("start", "weekly"), _e("weekly", "by_region"), _e("by_region", "check"),
|
||
_e("check", "write"), _e("write", "done"),
|
||
],
|
||
},
|
||
{
|
||
"name": "⑫ 跨源指标异动与对账诊断",
|
||
"description": "先核查数据源健康与跨源对账口径，再做同源维度定位与已确认本体穿透，出具带引用与非因果约束的诊断报告。",
|
||
"tags": ["分析", "本体", "对账", "证据"],
|
||
"nodes": [
|
||
_n("start", "input", "诊断问题", fields=[
|
||
{"name": "question", "required": True,
|
||
"default": "星河智造昨日设备有效运转率为何异常异动？请先核查数据源健康与跨源对账，再定位重点产线。",
|
||
"description": "要排查的指标异常或跨源对账问题"},
|
||
{"name": "period", "default": "2026-10-21", "description": "目标统计周期"}]),
|
||
_n("investigate", "agent", "跨源本体排查 Agent",
|
||
system="你是跨源指标诊断研究员：严格遵循先核数据源健康与对账、再同源拆解切片、最后沿已确认本体穿透的三段式纪律；"
|
||
"一旦发现源表断更、断档或对账失配即停止业务下钻。",
|
||
prompt="围绕统计周期 {{ input.period }} 排查问题：{{ input.question }}\n"
|
||
"1. 先调用本体邻居查询查看目标口径已确认的构成因子、跨源对账镜像与上游驱动关系，并调用数据源健康核查检查"
|
||
"涉及的数据源是否存在断更、断档、行数骤降或新增未识别编码；若存在对账镜像，调用口径查询取数后做跨源对账。\n"
|
||
"2. 若数据源健康正常且主指标超出波动基线，先在同源做变化拆解锁定贡献最大的业务实体切片。\n"
|
||
"3. 将异常切片传入本体邻居查询获取对齐后的调用参数，调用口径查询拉取已确认邻居指标并检验伴随异动（最多穿透 2 跳）。",
|
||
tools=[
|
||
"caliber_query",
|
||
"ontology_neighbors",
|
||
"source_health",
|
||
"analysis_comparable",
|
||
"analysis_baseline",
|
||
"analysis_decompose",
|
||
"analysis_reconcile",
|
||
"analysis_cochange",
|
||
],
|
||
approval="dangerous",
|
||
assign_to="investigation_notes"),
|
||
_n("write", "report", "诊断报告撰写", numbers="strict", on_violation="fail", claims="require_citation",
|
||
analysis_claims="strict",
|
||
system="你是经营诊断撰写人：结论先行，判断严格引用分析结论与数据源健康判定语，统计伴随异动绝不写成因果定论。",
|
||
instructions="围绕统计周期 [[i:period]] 回答：{{ input.question }}（排查摘要：{{ vars.investigation_notes }}）。"
|
||
"按「核查结论 / 数据源健康与跨源对账 / 切片定位与伴随异动」三段组织正文：数据源健康结论引用 [[h:H1]]，"
|
||
"分析判定直接引用 [[a:A1]] 等判定语，口径与单元格按证据引用；伴随异动一律表述为统计同步异动，"
|
||
"没有明确规则依据时写成待核实的推测。"),
|
||
_n("done", "output", "出具",
|
||
fields=[{"name": "跨源诊断报告", "value": "{{ nodes.write.text }}"}],
|
||
contract={"report_from": "write", "cells": True, "strict": True}),
|
||
],
|
||
"edges": [
|
||
_e("start", "investigate"), _e("investigate", "write"), _e("write", "done"),
|
||
],
|
||
},
|
||
]
|
||
|
||
|
||
SKILLS: list[dict[str, Any]] = [
|
||
{
|
||
"name": "结构化分析",
|
||
"description": "把开放问题拆成结论先行的分析",
|
||
"instructions": (
|
||
"分析任何问题时遵循：\n"
|
||
"1. 先给结论，再给理由，不要铺垫\n"
|
||
"2. 区分「事实」「推断」「假设」，并明确标注\n"
|
||
"3. 给出至少一个反面视角\n"
|
||
"4. 指出当前信息不足以判断的部分，不要硬答"
|
||
),
|
||
"tags": ["通用"],
|
||
},
|
||
{
|
||
"name": "批判性审阅",
|
||
"description": "只指出问题，不复述优点",
|
||
"instructions": (
|
||
"审阅内容时：\n"
|
||
"1. 只指出问题，不要复述优点\n"
|
||
"2. 每条问题写清楚「问题是什么 / 为什么是问题 / 怎么改」\n"
|
||
"3. 按严重程度排序，最致命的放最前面\n"
|
||
"4. 没有问题就直说没有，不要为凑数硬挑"
|
||
),
|
||
"tags": ["审阅"],
|
||
},
|
||
{
|
||
"name": "中文写作",
|
||
"description": "简洁自然、避免翻译腔的中文",
|
||
"instructions": (
|
||
"写中文时：\n"
|
||
"1. 不用「进行」「作出」这类空动词，直接用实义动词\n"
|
||
"2. 少用被动句和长定语从句\n"
|
||
"3. 术语首次出现时用括号给出英文原文\n"
|
||
"4. 能用短句就不写长句"
|
||
),
|
||
"tags": ["写作"],
|
||
},
|
||
]
|
||
|
||
|
||
async def seed_defaults() -> None:
|
||
"""首次启动时灌入默认 provider、模板工作流和 Skill。
|
||
|
||
已经存在的不会被覆盖 —— 这个函数每次启动都会跑，必须是幂等的。
|
||
"""
|
||
async with SessionLocal() as session:
|
||
# --- provider ---
|
||
count = (await session.execute(select(func.count(Provider.id)))).scalar() or 0
|
||
if count == 0:
|
||
session.add(
|
||
Provider(
|
||
name="演示模型（无需 API Key）",
|
||
kind="mock",
|
||
models=catalog.MOCK_MODELS,
|
||
default_model="mock-fast",
|
||
enabled=True,
|
||
)
|
||
)
|
||
# 环境里已有 Anthropic 凭据就自动配好，省得再手填一遍
|
||
token = os.environ.get("ANTHROPIC_API_KEY") or os.environ.get("ANTHROPIC_AUTH_TOKEN")
|
||
if token:
|
||
base_url = os.environ.get("ANTHROPIC_BASE_URL")
|
||
is_gateway = bool(base_url) and not os.environ.get("ANTHROPIC_API_KEY")
|
||
session.add(
|
||
Provider(
|
||
name="Anthropic" + ("（网关）" if is_gateway else ""),
|
||
kind="anthropic",
|
||
base_url=base_url,
|
||
api_key=encrypt(token),
|
||
models=catalog.ANTHROPIC_MODELS,
|
||
default_model="claude-opus-5",
|
||
enabled=True,
|
||
# 中转网关通常认 Authorization: Bearer 而不是 x-api-key
|
||
extra={"auth_style": "bearer"} if is_gateway else {},
|
||
)
|
||
)
|
||
logger.info("已从环境变量自动创建 Anthropic provider")
|
||
if os.environ.get("OPENAI_API_KEY"):
|
||
session.add(
|
||
Provider(
|
||
name="OpenAI",
|
||
kind="openai",
|
||
api_key=encrypt(os.environ["OPENAI_API_KEY"]),
|
||
models=catalog.OPENAI_MODELS,
|
||
default_model="gpt-5.2",
|
||
enabled=True,
|
||
)
|
||
)
|
||
await session.commit()
|
||
|
||
# --- Skill ---
|
||
existing_skills = {
|
||
s for (s,) in await session.execute(select(Skill.name))
|
||
}
|
||
for skill in SKILLS:
|
||
if skill["name"] not in existing_skills:
|
||
session.add(Skill(**skill))
|
||
await session.commit()
|
||
|
||
# --- 模板工作流 ---
|
||
from app.engine.layout import auto_layout
|
||
|
||
existing = {n for (n,) in await session.execute(select(Workflow.name))}
|
||
added = 0
|
||
for template in TEMPLATES:
|
||
if template["name"] in existing:
|
||
continue
|
||
spec = auto_layout(
|
||
GraphSpec.model_validate(
|
||
{"nodes": template["nodes"], "edges": template["edges"]}
|
||
)
|
||
)
|
||
graph = spec.model_dump(mode="json")
|
||
workflow = Workflow(
|
||
name=template["name"],
|
||
description=template["description"],
|
||
graph=graph,
|
||
tags=template["tags"],
|
||
is_template=True,
|
||
)
|
||
session.add(workflow)
|
||
await session.flush()
|
||
session.add(
|
||
WorkflowVersion(workflow_id=workflow.id, version=1, graph=graph, note="内置模板")
|
||
)
|
||
added += 1
|
||
if added:
|
||
await session.commit()
|
||
logger.info("已导入 %d 个内置模板", added)
|