mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-19 18:41:58 +00:00
5b4bd91476
为什么:问数据的答案要和已治理口径接起来(next-round 02 做法 2、3、9、10;总方案 X1、X2、X3、X12,拍板 L2、L4、L8)。
- 比对(ledger/match.py):探索运行(chat、canvas、trial、personal_app)成功结束、封存完、有查询快照时,在 runner 封存之后和
notify_run 同一个位置放到后台;候选口径是每个工作流最新的已发布未退役版本里查库的调用工具节点(受管、已发布都算),按数据源
和表预筛;逐个用 match_template 比,schema 用查询当时冻结的表结构、列名取同一份快照里冻结的数据目录。每次运行限时 8 秒、在
线程里、同时最多 2 个;超时、快照被改过或读不出、SQL 解析不了照实记 skipped。启动时(run_manager.setup,和调度循环同一个开关)
补比最近 7 天封存了却没比过的运行,只读工件和版本表。
- 记法(ledger/store.py,拍板 L8):内容寻址工件 kind=caliber_match(agentlab.caliber_match/1)+ run_append 的 sealed 方式追加
ledger.matched(core/events.py 登记);工件里不放时间和工作流名,同一份封存再比一次哈希一样;meta 记 calibers / related / counts。
- 标签(ledger/labels.py):单元格按查询的比对状态;指标要每个输入都等价于同一份口径同一版、绑定一致,式子按「哪个取数节点、
第几行、哪一列」换完和口径卡上的一字不差。caliber(how:independent 等价已证明;composed 助手选用留给第 3 波)/ related /
reasoning / unknown;整题判断附还缺的入口字段。
- 读(ledger/view.py):GET /api/runs/{id}/ledger(ledger_api.runs_router,和证据同一道门),没比对过、能比的按需排上;证据
片段接口加 ledger 键(没比对过的运行没有这个键,返回一字不差);报告导出的面板最上面多「口径」一步并写明「封存后比对,不在
封存范围内」。遮罩按读的那一刻重滤:绑定的值是这次运行查到过的遮罩值时不给值,差异条目和逐行对照走 _Shown。
- 来源链:POST /api/runs 加 origin_ref(只在 origin=chat 时收,服务端核对对话访问权和轮次归属,写进 Run.origin 和封存的
run.origin);WorkflowIn 加 origin(run_id、conversation_id、turn_id,核对规则同 from-run);报告导出和可验证报告下载各记一条
审计 report.exported。可验证报告列封存后追加的事件时给 ledger.matched 一句只有条数的说明。
- 统计(ledger/stats.py):E1–E4 从工件、运行、查询快照、审计现算,GET /api/ledger/stats(搭建者);E3 用 app/chat/questions
的聚类加按(数据源、normalize_sql 文本)的确定性重复;评测直接调 compute_stats(可点名 run_ids)。对账反馈、认证转化、按发布
的召回在后面几波,照实写没有数据。
测试:test_ledger_match.py(真跑探索运行:等价、相关、没有相关口径、写死的码值、非探索运行、超时、快照被改、解析失败、预筛、
指标两种、级别、只认最新版、补比)、test_ledger_api.py(账本接口、按需比、遮罩、证据片段、导出、审计、统计)、
test_run_origin_ref.py(origin_ref 与画布来源的核对)。
354 lines
19 KiB
Python
354 lines
19 KiB
Python
"""逐个数标口径(docs/design/next-round/02-口径账本.md 做法 3):报告里每个引用了查询格、口径卡指标的数,来自哪一份口径。
|
||
|
||
标签四种(KINDS):
|
||
- caliber:来自口径「X」vN。how 是怎么认的:independent「等价已证明」(这次 Agent 自己写的 SQL 被证明和口径的取数等价)、
|
||
composed「助手选用」(助手按口径组合出来的;组合在第 3 波,这一波只留这个值)。
|
||
- related:相关口径——同一个数据源、查了同样的表的已发布口径,但不等价;附差在哪(explain_sql_diff 的条目)。
|
||
- reasoning:这次的推理——没有相关的已发布口径。
|
||
- unknown:没比对、比对没做完(超时、解析失败)、快照读不出来。
|
||
|
||
怎么认:
|
||
- 单元格([[v:]]、[[table:]] 里的格):按证据目录找到查询快照,取那条查询的比对状态。等价时按列映射写出口径那边的列名。
|
||
- 指标([[m:]]):指标集工件里这个指标的每个输入都要来自查询快照,而且这些快照都被证明等价于同一份口径的同一版、绑定一致;
|
||
指标式还要和那一版里某张口径卡上的某个指标相同——两边的式子把每个输入换成「口径的哪个取数节点、第几行、哪一列」之后
|
||
一字不差(caliber_from 转抄来的式子本来就是那一张卡的,同样按这个比)。输入有不来自查询的(运行输入、整形、代码)不算来自口径。
|
||
|
||
整题(whole_of):主报告里所有带标签的数都来自同一口径的同一版、绑定一致,就记 equivalent,并写出还缺的入口字段(口径入口
|
||
节点声明了、这次的绑定里没有的;required 且没有缺省值的另列)。
|
||
|
||
这里只是纯函数:标签在比对时算好、存进比对工件(ledger/match.py);接口读的时候按这次运行查到过的遮罩值遮一遍(view_label):
|
||
绑定的值是遮罩值的不给值,差异条目里的 SQL 片段走证据面板同一套遮罩(api/evidence._Shown)。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
from collections.abc import Callable, Mapping
|
||
from typing import Any
|
||
|
||
#: 标签的四种、认法的两种
|
||
KINDS = ("caliber", "related", "reasoning", "unknown")
|
||
HOWS = ("independent", "composed")
|
||
#: 界面上标签要写明的两句:比对在封存之后做,不在封存范围内;事后补比的是按现在的口径回看
|
||
POST_SEAL_NOTE = "封存后比对,不在封存范围内"
|
||
RETRO_NOTE = "按现在的口径回看"
|
||
#: 发布级别的叫法(同 blueprints/publish_plan.LEVEL_TEXT、前端 lib/terms.ts)
|
||
LEVEL_TEXT = {"published": "已发布", "governed": "受管"}
|
||
MASKED = "已遮罩"
|
||
|
||
#: unknown 的几种原因
|
||
NO_RECORD = "这一格出自的查询没有比对记录"
|
||
CARD_UNREADABLE = "指标集读不出来,没有比对"
|
||
METRIC_GONE = "指标集里没有这个指标,没有比对"
|
||
UNFINISHED = "比对没有做完"
|
||
#: related 的几种说明
|
||
FORMULA_DIFFERS = "取数和口径等价,但指标的算法和口径卡上的不同"
|
||
INPUTS_ELSEWHERE = "指标有的输入不来自查询(运行输入、整形或代码),不能算来自口径"
|
||
MIXED_CALIBERS = "指标的几个输入分别对上了不同的口径"
|
||
BINDING_CONFLICT = "指标的几个输入对上了同一份口径,但参数不一致"
|
||
|
||
#: 能追到查询快照里那一格的输入(同 engine/lineage.SNAPSHOT_VIAS)
|
||
_SNAPSHOT_VIAS = ("tool_cell", "agent_field")
|
||
|
||
|
||
def caliber_key(workflow_id: str, version: int) -> str:
|
||
"""口径版本的写法「工作流 id@版本」:工件 meta、召回反查、统计都按它认。"""
|
||
return f"{workflow_id}@{version}"
|
||
|
||
|
||
def _unknown(reason: str) -> dict[str, Any]:
|
||
return {"kind": "unknown", "reason": reason}
|
||
|
||
|
||
def _related(query: Mapping[str, Any], *, reason: str | None = None) -> dict[str, Any]:
|
||
"""相关口径:这条查询比过、没等价的那几份(附差在哪)。一份相关的都没有时是这次的推理。"""
|
||
related = [{k: cand[k] for k in ("caliber", "workflow_id", "version", "level", "node_id", "comparable",
|
||
"reason", "diff", "binding", "via") if k in cand}
|
||
for cand in query.get("candidates") or [] if cand.get("status") == "related"]
|
||
if not related:
|
||
return {"kind": "reasoning"}
|
||
out: dict[str, Any] = {"kind": "related", "related": related}
|
||
if reason:
|
||
out["reason"] = reason
|
||
return out
|
||
|
||
|
||
def _caliber(cand: Mapping[str, Any], **extra: Any) -> dict[str, Any]:
|
||
return {"kind": "caliber", "how": "independent", "caliber": cand.get("caliber"),
|
||
"workflow_id": cand.get("workflow_id"), "version": cand.get("version"), "level": cand.get("level"),
|
||
"binding": dict(cand.get("binding") or {}), **extra}
|
||
|
||
|
||
def _proven(query: Mapping[str, Any]) -> list[Mapping[str, Any]]:
|
||
return [c for c in query.get("candidates") or [] if c.get("status") == "proven"]
|
||
|
||
|
||
def _caliber_column(cand: Mapping[str, Any], column: Any) -> str | None:
|
||
"""这次结果里的列 → 口径那边同一位置的列名(列映射,按位置);没有映射、对不上时为 None。"""
|
||
for item in cand.get("columns") or []:
|
||
if item.get("ours") == column:
|
||
return item.get("caliber")
|
||
return None
|
||
|
||
|
||
def cell_label(artifact: str, column: Any, queries: Mapping[str, Mapping[str, Any]]) -> dict[str, Any]:
|
||
"""单元格的标签:它所在的那条查询的比对状态。"""
|
||
query = queries.get(artifact)
|
||
if query is None:
|
||
return _unknown(NO_RECORD)
|
||
status = query.get("status")
|
||
if status in ("proven", "composed"):
|
||
cand = _proven(query)[0] if _proven(query) else {}
|
||
label = _caliber(cand, query=artifact, node_id=cand.get("node_id"),
|
||
column=_caliber_column(cand, column) if column is not None else None)
|
||
if status == "composed":
|
||
label["how"] = "composed"
|
||
return label
|
||
if status == "related":
|
||
return {**_related(query), "query": artifact}
|
||
if status == "none":
|
||
return {"kind": "reasoning", "query": artifact}
|
||
return {**_unknown(str(query.get("reason") or UNFINISHED)), "query": artifact}
|
||
|
||
|
||
def _leaf_token(node_id: Any, row: Any, column: Any) -> str:
|
||
"""把指标式里的一个输入换成「口径的哪个取数节点、第几行、哪一列」:两边的式子按这个换完再逐字比。"""
|
||
return f"⟦{node_id}|{row}|{column}⟧"
|
||
|
||
|
||
def _mapped_expression(expression: str, tokens: Mapping[str, str]) -> str | None:
|
||
"""指标式里的输入按 tokens(输入路径 → 记号)换掉之后的写法;解析不了、有输入没换到的为 None。"""
|
||
from app.engine.expressions import ExpressionError, leaf_refs, parse_expression, substitute
|
||
|
||
try:
|
||
tree = parse_expression(expression)[0]
|
||
except ExpressionError:
|
||
return None
|
||
leaves = leaf_refs(tree)
|
||
if not leaves or any(leaf not in tokens for leaf in leaves):
|
||
return None
|
||
return substitute(tree, dict(tokens))
|
||
|
||
|
||
def card_expressions(cards: list[Mapping[str, Any]]) -> list[tuple[str, str, str]]:
|
||
"""口径版本里每张源头口径卡的每个指标:(卡的节点 id, 指标 id, 换成记号后的式子)。卡上的输入要是 cell(nodes.X, 行, '列')
|
||
(look/calibers.card_inputs 认的那种),别的写法的指标不收。"""
|
||
import ast
|
||
|
||
from app.engine.expressions import ExpressionError, cell_parts, leaf_refs, parse_expression
|
||
|
||
out: list[tuple[str, str, str]] = []
|
||
for card in cards:
|
||
for metric in card.get("metrics") or []:
|
||
expression = str((metric or {}).get("expression") or "")
|
||
try:
|
||
tree = parse_expression(expression)[0]
|
||
except ExpressionError:
|
||
continue
|
||
tokens: dict[str, str] = {}
|
||
for leaf in leaf_refs(tree):
|
||
parts = cell_parts(leaf)
|
||
if parts is None or not parts[0].startswith("nodes."):
|
||
break
|
||
try:
|
||
row, column = ast.literal_eval(parts[1]), ast.literal_eval(parts[2])
|
||
except (ValueError, SyntaxError):
|
||
break
|
||
tokens[leaf] = _leaf_token(parts[0][len("nodes."):], row, column)
|
||
else:
|
||
if (mapped := _mapped_expression(expression, tokens)) is not None:
|
||
out.append((str(card.get("node_id")), str(metric.get("id")), mapped))
|
||
return out
|
||
|
||
|
||
def metric_label(card: Mapping[str, Any] | None, metric_id: Any, queries: Mapping[str, Mapping[str, Any]],
|
||
calibers: Mapping[str, Mapping[str, Any]]) -> dict[str, Any]:
|
||
"""指标的标签(见模块说明)。card 是封存的指标集工件(读不出为 None);calibers 是 {「工作流@版本」: {level, cards}},
|
||
cards 是那一版里源头口径卡的指标式(match 从版本快照取)。"""
|
||
if card is None:
|
||
return _unknown(CARD_UNREADABLE)
|
||
metric = next((m for m in card.get("metrics") or [] if isinstance(m, dict) and m.get("id") == metric_id), None)
|
||
if metric is None:
|
||
return _unknown(METRIC_GONE)
|
||
inputs = [i for i in metric.get("inputs") or [] if isinstance(i, dict)]
|
||
from_queries = [i for i in inputs if i.get("artifact") and i.get("via") in _SNAPSHOT_VIAS]
|
||
if not from_queries:
|
||
return {"kind": "reasoning"}
|
||
found = [queries.get(str(i["artifact"])) for i in from_queries]
|
||
if any(q is None for q in found):
|
||
return _unknown(NO_RECORD)
|
||
if any(q.get("status") == "skipped" for q in found): # type: ignore[union-attr]
|
||
return _unknown(next(str(q.get("reason") or UNFINISHED) for q in found
|
||
if q.get("status") == "skipped")) # type: ignore[union-attr]
|
||
queries_found: list[Mapping[str, Any]] = [q for q in found if q is not None]
|
||
proven_keys = [{str(c.get("caliber")) for c in _proven(q)} for q in queries_found]
|
||
common = set.intersection(*proven_keys) if proven_keys else set()
|
||
nearby = _brief([c for q in queries_found for c in q.get("candidates") or []
|
||
if c.get("status") in ("proven", "related")])
|
||
if len(from_queries) < len(inputs):
|
||
return {"kind": "related", "related": nearby, "reason": INPUTS_ELSEWHERE} if nearby else {"kind": "reasoning"}
|
||
if not common:
|
||
if all(proven_keys):
|
||
return {"kind": "related", "related": nearby, "reason": MIXED_CALIBERS}
|
||
related = next((q for q in queries_found if q.get("status") == "related"), None)
|
||
return _related(related) if related is not None else {"kind": "reasoning"}
|
||
reason = FORMULA_DIFFERS
|
||
for key in sorted(common):
|
||
chosen = [next(c for c in _proven(q) if c.get("caliber") == key) for q in queries_found]
|
||
binding: dict[str, Any] = {}
|
||
conflict = False
|
||
for cand in chosen:
|
||
for name, value in (cand.get("binding") or {}).items():
|
||
if name in binding and binding[name] != value:
|
||
conflict = True
|
||
binding.setdefault(name, value)
|
||
if conflict:
|
||
reason = BINDING_CONFLICT
|
||
continue
|
||
tokens: dict[str, str] = {}
|
||
for item, cand in zip(from_queries, chosen):
|
||
locator = item.get("locator") if isinstance(item.get("locator"), dict) else {}
|
||
mapped = _caliber_column(cand, locator.get("column"))
|
||
if mapped is None or item.get("path") is None:
|
||
tokens = {}
|
||
break
|
||
tokens[str(item["path"])] = _leaf_token(cand.get("node_id"), locator.get("row"), mapped)
|
||
mine = _mapped_expression(str(metric.get("expression") or ""), tokens) if tokens else None
|
||
if mine is None:
|
||
continue
|
||
hit = next(((node, mid) for node, mid, theirs in (calibers.get(key) or {}).get("expressions") or []
|
||
if theirs == mine), None)
|
||
if hit is None:
|
||
continue
|
||
return {**_caliber(chosen[0], card=hit[0], metric=hit[1]), "binding": binding}
|
||
return {"kind": "related", "related": _brief([c for q in queries_found for c in _proven(q)
|
||
if c.get("caliber") in common]), "reason": reason}
|
||
|
||
|
||
def _brief(cands: list[Mapping[str, Any]]) -> list[dict[str, Any]]:
|
||
"""指标上的相关口径:只写是哪一版、哪个取数节点(差在哪在各个输入的查询上),同一版只写一次,最多 5 个。"""
|
||
out: list[dict[str, Any]] = []
|
||
for cand in cands:
|
||
if any(o.get("caliber") == cand.get("caliber") for o in out):
|
||
continue
|
||
out.append({k: cand[k] for k in ("caliber", "workflow_id", "version", "level", "node_id") if k in cand})
|
||
return out[:5]
|
||
|
||
|
||
def label_doc(doc: Mapping[str, Any], queries: Mapping[str, Mapping[str, Any]],
|
||
calibers: Mapping[str, Mapping[str, Any]],
|
||
load_card: Callable[[str], Mapping[str, Any] | None]) -> dict[str, dict[str, Any]]:
|
||
"""一份报告文档里每个数的标签:{片段 id: 标签}。只标引用解析了的数字、取值片段(单元格、指标);运行输入、表名、引文、
|
||
没有引用的数不标。"""
|
||
from app.engine.evidence import iter_units
|
||
|
||
catalog = doc.get("catalog") if isinstance(doc.get("catalog"), dict) else {}
|
||
cards: dict[str, Mapping[str, Any] | None] = {}
|
||
out: dict[str, dict[str, Any]] = {}
|
||
for _, unit in iter_units(doc):
|
||
for seg in unit.get("segments") or []:
|
||
if not isinstance(seg, dict) or seg.get("kind") not in ("number", "value") or not seg.get("id"):
|
||
continue
|
||
cite = seg.get("cite") if isinstance(seg.get("cite"), dict) else {}
|
||
if cite.get("status") != "resolved":
|
||
continue
|
||
entry = catalog.get(cite.get("alias")) if isinstance(catalog.get(cite.get("alias")), dict) else {}
|
||
locator = cite.get("locator") if isinstance(cite.get("locator"), dict) else {}
|
||
artifact = str(entry.get("artifact") or "")
|
||
if cite.get("kind") == "cell" and artifact:
|
||
out[str(seg["id"])] = cell_label(artifact, locator.get("column"), queries)
|
||
elif cite.get("kind") == "metric" and artifact:
|
||
if artifact not in cards:
|
||
cards[artifact] = load_card(artifact)
|
||
out[str(seg["id"])] = metric_label(cards[artifact], locator.get("metric"), queries, calibers)
|
||
return out
|
||
|
||
|
||
def whole_of(labels: Mapping[str, Mapping[str, Any]], calibers: Mapping[str, Mapping[str, Any]]) -> dict[str, Any]:
|
||
"""整题判断(主报告):所有带标签的数都来自同一口径的同一版、绑定一致时 equivalent,附还缺的入口字段。"""
|
||
counts = {kind: 0 for kind in KINDS}
|
||
for label in labels.values():
|
||
counts[str(label.get("kind"))] = counts.get(str(label.get("kind")), 0) + 1
|
||
if not labels:
|
||
return {"equivalent": False, "reason": "报告里没有可标口径的数", "counts": counts}
|
||
keys = {label.get("caliber") for label in labels.values() if label.get("kind") == "caliber"}
|
||
if counts["caliber"] != len(labels):
|
||
reason = ("有的数没有比对" if counts["unknown"] else "有的数是这次的推理" if counts["caliber"]
|
||
else "没有数来自口径")
|
||
return {"equivalent": False, "reason": reason, "counts": counts}
|
||
if len(keys) != 1:
|
||
return {"equivalent": False, "reason": "数来自不同的口径", "counts": counts}
|
||
binding: dict[str, Any] = {}
|
||
for label in labels.values():
|
||
for name, value in (label.get("binding") or {}).items():
|
||
if name in binding and binding[name] != value:
|
||
return {"equivalent": False, "reason": "各个数的参数不一致", "counts": counts}
|
||
binding.setdefault(name, value)
|
||
key = str(next(iter(keys)))
|
||
info = calibers.get(key) or {}
|
||
fields = [f for f in info.get("fields") or [] if isinstance(f, dict) and f.get("name")]
|
||
first = next(iter(labels.values()))
|
||
return {
|
||
"equivalent": True, "caliber": key, "workflow_id": first.get("workflow_id"), "version": first.get("version"),
|
||
"level": first.get("level"),
|
||
"how": "composed" if any(label.get("how") == "composed" for label in labels.values()) else "independent",
|
||
"binding": binding,
|
||
"missing": [f["name"] for f in fields if f["name"] not in binding],
|
||
"missing_required": [f["name"] for f in fields
|
||
if f["name"] not in binding and f.get("required") and not f.get("has_default")],
|
||
"counts": counts,
|
||
}
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 读的时候遮:绑定的值、差异条目里的 SQL 片段
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def binding_view(binding: Mapping[str, Any] | None, hide: Callable[[Any], bool]) -> list[dict[str, Any]]:
|
||
"""绑定 → [{field, value} 或 {field, masked: true}]:值是这次运行查到过的遮罩值的不给值(也就不能拿它代填出具)。"""
|
||
out = []
|
||
for name, value in (binding or {}).items():
|
||
out.append({"field": name, "masked": True, "value": MASKED} if hide(value) else {"field": name, "value": value})
|
||
return out
|
||
|
||
|
||
def diff_view(diff: Mapping[str, Any] | None, shown: Callable[[Any], Any]) -> dict[str, Any] | None:
|
||
"""差在哪:条目的两边写法、说明和逐行对照里的 SQL 片段按遮罩值换掉。"""
|
||
if not isinstance(diff, Mapping):
|
||
return None
|
||
items = [{**item, "a": shown(item.get("a")), "b": shown(item.get("b")), "text": shown(item.get("text"))}
|
||
for item in diff.get("items") or [] if isinstance(item, Mapping)]
|
||
return {"items": items, "fallback": bool(diff.get("fallback")),
|
||
"text_diff": [shown(line) for line in diff.get("text_diff") or []]}
|
||
|
||
|
||
def view_label(label: Mapping[str, Any], *, hide: Callable[[Any], bool], shown: Callable[[Any], Any],
|
||
names: Mapping[str, str]) -> dict[str, Any]:
|
||
"""一个标签交给界面的样子:绑定按遮罩给、差在哪按遮罩换,口径写上工作流名和级别的叫法(名字按现在的)。"""
|
||
out = {k: v for k, v in label.items() if k not in ("binding", "related")}
|
||
if label.get("kind") == "caliber":
|
||
out["binding"] = binding_view(label.get("binding"), hide)
|
||
out["workflow_name"] = names.get(str(label.get("workflow_id") or ""))
|
||
out["level_text"] = LEVEL_TEXT.get(str(label.get("level") or ""), "")
|
||
if label.get("kind") == "related":
|
||
out["related"] = [related_view(item, hide=hide, shown=shown, names=names)
|
||
for item in label.get("related") or []]
|
||
return out
|
||
|
||
|
||
def related_view(item: Mapping[str, Any], *, hide: Callable[[Any], bool], shown: Callable[[Any], Any],
|
||
names: Mapping[str, str]) -> dict[str, Any]:
|
||
out = {k: item[k] for k in ("caliber", "workflow_id", "version", "level", "node_id", "comparable", "reason", "via")
|
||
if k in item}
|
||
out["workflow_name"] = names.get(str(item.get("workflow_id") or ""))
|
||
out["level_text"] = LEVEL_TEXT.get(str(item.get("level") or ""), "")
|
||
if "binding" in item:
|
||
out["binding"] = binding_view(item.get("binding"), hide)
|
||
if "diff" in item:
|
||
out["diff"] = diff_view(item.get("diff"), shown)
|
||
return out
|
||
|
||
|
||
__all__ = ["HOWS", "KINDS", "LEVEL_TEXT", "POST_SEAL_NOTE", "RETRO_NOTE", "binding_view", "caliber_key",
|
||
"card_expressions", "cell_label", "diff_view", "label_doc", "metric_label", "related_view", "view_label",
|
||
"whole_of"]
|