可验证报告样本页加异动周报(analysis,分析能力一轮):Agent 在虚构零售库 shop 上查两周的总数和按地区两周的销售额,做按地区的变化拆解(带总数核对)、波动基线(两周历史不够,算不了)、两期可比性检查,结论按快照真算(app.analysis.run,目录事实按虚构的数据目录给),Agent 节点的 tool.end 照分析工具真实的写法带结论工件和判定那一段;报告引用拆解的判定语、具名的数、结果表的格和整张结果表,正文后面是 engine/analysis_scope 按虚构的数据目录算出的分析范围(没拆的是渠道)。check-bundle:样本页 19 份;异动周报原样导出全部通过——③ 三份结论和两份输入快照随文件附带,④ 输入快照按「已核对的分析结论的输入」认,⑤ 分开数判定语,点开判定语看得到判定标准、方法与版本、附带说明、Agent 写的假设、输入和导出时的复算,具名的数、结果表的格写对出处,分析范围的编号点得开(正文没引用的 A2 也随文件附带)、范围里的编号不核正文显示,390 宽不溢出;只改判定语、具名的数的显示并重算全部哈希各自变红并点名。另:「查询快照以外的工件标成遮罩」的原因照第 1 波改过的说法(只有查询快照、口径卡、分析结论和报告文档……),原来的断言没跟着改。test_bundle_build 的样本页清单加上它、核对它的形状

This commit is contained in:
ThearchyHelios
2026-10-21 19:09:21 +08:00
parent 019676670b
commit 017cbee5aa
3 changed files with 216 additions and 4 deletions
+124
View File
@@ -38,6 +38,10 @@
原样导出照样全部通过
- prose:同经营月报的口径卡和流水,正文是写作者自己写在引用旁边的普通文字(「运费统一[[m:…]]」「合计——[[m:…]]」
「SKU-[[v:…]]」「共[[m:…]]家」这类):原样导出全部通过、总览是绿的,紧挨着核对的规则不误判正常写法
- analysis:异动周报(分析能力一轮)。Agent 在虚构零售库 shop 上查了两周的总数(Q1)和按地区两周的销售额(Q2),做了按地区的
变化拆解(A1,带总数核对)、波动基线(A2,两周历史不够,算不了)、两期可比性检查(A3),结论都是 app.analysis 真算的;
报告引用 A1 的判定语、具名的数(占比)、结果表的格,正文后面是系统写的「分析范围」(engine/analysis_scope 按虚构的数据
目录算出没拆的渠道)。结论和输入快照随文件附带、导出时复算过,原样导出全部通过;分析范围里的编号点得开(A2、A3 正文没引用)
"""
from __future__ import annotations
@@ -313,6 +317,125 @@ def formats_bundle(*, card: dict[str, Any] = FORMATS_CARD, table: dict[str, Any]
return build(f, events, bound)
#: 分析样本(analysis)的两份查询:虚构零售库 shop 两周的总数(Q1)、四个地区两周的销售额(Q2)
AN_TOTALS = {"columns": ["week", "gmv"], "column_types": {"week": "text", "gmv": "number"},
"rows": [["2026-W41", 2220], ["2026-W42", 2026]], "source": "shop",
"sql": "SELECT week, SUM(amount) AS gmv FROM orders GROUP BY week ORDER BY week", "truncated": False,
"row_count": 2}
AN_REGIONS = {"columns": ["region", "week", "gmv"], "column_types": {"region": "text", "week": "text", "gmv": "number"},
"rows": [["华东", "2026-W41", 820], ["华东", "2026-W42", 700], ["华北", "2026-W41", 600],
["华北", "2026-W42", 560], ["华南", "2026-W41", 700], ["华南", "2026-W42", 666],
["西南", "2026-W41", 100], ["西南", "2026-W42", 100]], "source": "shop",
"sql": "SELECT region, week, SUM(amount) AS gmv FROM orders GROUP BY region, week ORDER BY region, week",
"truncated": False, "row_count": 8}
#: Agent 的三次分析调用(工具名、请求),照分析工具交给 app.analysis.run 的写法
AN_CALLS = [
("analysis_decompose", {"op": "decompose", "mode": "additive", "hypothesis": "下降是不是集中在某个地区",
"params": {"query": "Q2", "by": ["region"], "period": "week", "base": "2026-W41",
"target": "2026-W42", "measure": "gmv", "label": "销售额",
"headline": {"base": "Q1.r0.gmv", "target": "Q1.r1.gmv"}}}),
("analysis_baseline", {"op": "baseline", "mode": None, "hypothesis": "这周的下降算不算异常",
"params": {"query": "Q1", "period": "week", "measure": "gmv", "label": "销售额"}}),
("analysis_comparable", {"op": "comparable", "mode": None, "hypothesis": "两周能不能直接比",
"params": {"query": "Q2", "period": "week", "base": "2026-W41", "target": "2026-W42",
"measure": "gmv"}}),
]
#: 异动周报:引用 A1 的判定语、具名的数(占比)、结果表的格;A2、A3 只在系统写的分析范围里
AN_REPORT = ("## 异动周报\n\n[[a:A1]]。华东占总变化的 [[a:A1.top_share|pct]],华东的变化是 [[v:A1.r0.delta]]。[[see:A1]]\n\n"
"各地区的变化见下表。\n\n[[table:A1 cols=region,base,target,delta rows=0-3]]")
def _shop_notes() -> dict[str, Any]:
"""虚构零售库 shop 的数据目录(orders 是事实表;地区、渠道是属性列,码值表列全):分析范围按它算没拆的候选维度。"""
from app.data.catalog import make_item
def column(label: str, measure: str, codes: tuple[str, ...] = ()) -> dict[str, Any]:
out = {"label": make_item(label, "human"), "measure": make_item(measure, "human")}
if codes:
out["codes"] = make_item({c: "" for c in codes}, "profile", "verified")
return out
return {"orders": {"kind": make_item("fact", "human"), "columns": {
"region": column("地区", "attribute", ("华东", "华北", "华南", "西南")),
"channel": column("渠道", "attribute", ("线上", "线下")),
"week": column("周", "attribute"), "amount": column("实收金额", "flow")}}}
def analysis_bundle() -> dict[str, Any]:
"""异动周报(analysis)的页面数据:结论按两份快照真算(app.analysis.run),Agent 节点的事件照分析工具真实的写法
(tool.end 带结论工件和判定那一段),报告文档按写作目录真实组装,分析范围按虚构的数据目录真实算出。"""
import asyncio
from datetime import date
from app import analysis
from app.analysis.contract import effective_settings, ledger_entry
from app.engine import analysis_scope
from app.tools.analysis import event_view, input_snapshot
run = {**RUN, "id": "9c2e4a6b8d0f4e1a3c5b7d9f1e3a5c7b", "workflow_id": "wf-sample-anomaly", "workflow_name": "异动周报",
"started_at": "2026-10-19T01:00:00+00:00", "finished_at": "2026-10-19T01:01:30+00:00"}
graph = {"nodes": [{"id": n, "type": t, "data": {"label": label}} for n, t, label in [
("start", "input", "开始"), ("bot", "agent", "问数据"), ("write", "report", "写周报"), ("out", "output", "成果")]]}
files: dict[str, bytes] = {}
def put(obj: Any) -> str:
artifact, raw = stored(obj)
files[artifact] = raw
return artifact
def load(artifact: str) -> Any:
raw = files.get(artifact)
return json.loads(raw) if raw is not None else None
queries = [(put(snap), snap) for snap in (AN_TOTALS, AN_REGIONS)]
# 照分析工具交给算子的写法:查询快照原样加上出处;目录事实(列的叫法、度量类型)按虚构的数据目录给,结论里记下
snapshots = {f"Q{i}": input_snapshot(snap, ref=f"Q{i}", artifact=sha, node_id="bot")
for i, (sha, snap) in enumerate(queries, 1)}
context = {"today": date(2026, 10, 19), "catalogs": {"shop": _shop_notes()}}
ledger: list[dict[str, Any]] = [
{"kind": "query", "node_id": "bot", "exec": 1, "artifact": sha, "call_id": f"call_{i}", "tool": "db_query__shop",
"source": "shop", "columns": snap["columns"], "rows": len(snap["rows"]), "truncated": False}
for i, (sha, snap) in enumerate(queries)]
results = []
for n, (tool, request) in enumerate(AN_CALLS, len(queries)):
result = dict(asyncio.run(analysis.run(request, snapshots.get, settings=effective_settings(None), context=context)))
artifact = put(result)
results.append((tool, f"call_{n}", artifact, result))
ledger.append(ledger_entry(result, node_id="bot", exec_no=1, artifact=artifact, call_id=f"call_{n}"))
catalog = build_catalog(nodes={}, ledger=ledger, loader=load)
doc = compose_doc(AN_REPORT, catalog, node_id="write", run_id=run["id"], loader=load)
doc["analysis_scope"] = analysis_scope.build(catalog, sources={"shop": {"tables": _shop_notes(), "masked": []}},
load=load)
doc_id = put(doc)
events: list[tuple[int, str, str | None, Any]] = []
def add(etype: str, node_id: str | None, data: dict[str, Any]) -> None:
events.append((len(events) + 1, etype, node_id, data))
add("run.started", None, {"nodes": len(graph["nodes"]), "resumed": False, "memory_scope": "default",
"collection": "default", "approval_default": None, "tool_trust": "formal",
"agent_limits": {}, "data_versions": None, "replay_protocol": 2})
add("node.started", "bot", {"node_type": "agent", "label": "问数据"})
for i, (sha, snap) in enumerate(queries):
add("tool.end", "bot", {"tool": "db_query__shop", "call_id": f"call_{i}", "duration_ms": 9, "query_artifact": sha,
"preview": json.dumps({"columns": snap["columns"], "rows": snap["rows"][:2]},
ensure_ascii=False)})
for k, (tool, call_id, artifact, result) in enumerate(results, 1):
add("tool.end", "bot", {"tool": tool, "call_id": call_id, "duration_ms": 6, "analysis_artifact": artifact,
"preview": json.dumps({"analysis_artifact": artifact}), "analysis": event_view(result, f"A{k}")})
add("node.finished", "bot", {"duration_ms": 240, "attempt": 1, "preview": {"text": "销售额下降,主要来自华东。"},
"evidence": ledger})
add("llm.end", "write", {"model": "mock", "output": "(写作草稿)"})
add("report.checked", "write", {"doc_artifact": doc_id, "ok": True})
add("node.finished", "write", {"duration_ms": 20, "attempt": 1, "preview": {"text": "(报告正文)", "doc_artifact": doc_id}})
add("issuance", "out", {"tier": "formal", "mode": "citations"})
add("node.finished", "out", {"duration_ms": 2, "attempt": 1, "preview": {"周报": "(报告正文)"}})
add("run.finished", None, {"output": {"周报": "(报告正文)"}})
return build_bundle(run, events, manifest_seq=len(events), manifest_hash=manifest_hash(events), read=files.get,
exported_at=EXPORTED_AT, graph=graph, source_masks={})
def bundles() -> dict[str, dict[str, Any]]:
"""{页面名: 页面数据}"""
out: dict[str, dict[str, Any]] = {}
@@ -368,6 +491,7 @@ def bundles() -> dict[str, dict[str, Any]]:
out["formats"] = formats_bundle()
out["formats-unit"] = formats_bundle(card=FORMATS_UNIT_CARD)
out["prose"] = formats_bundle(text=PROSE)
out["analysis"] = analysis_bundle()
return out