mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:31:58 +00:00
ThearchyHelios
1363 lines
78 KiB
Python
1363 lines
78 KiB
Python
from __future__ import annotations
|
||
|
||
import json
|
||
import re
|
||
from typing import Any
|
||
|
||
from app.analysis.contract import LEDGER_KIND
|
||
from app.core.events import EventType
|
||
from app.engine.context import NodeContext, NodeError
|
||
from app.engine.expressions import ExpressionError, eval_expression
|
||
from app.engine.schema import GraphNode, NodeType, _ancestors
|
||
from app.engine.state import GraphState, template_context
|
||
|
||
|
||
async def run_input(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||
"""入口节点。把运行输入按声明的字段整理进变量池，补上默认值。"""
|
||
raw = dict(state.get("input") or {})
|
||
fields = ctx.cfg("fields", []) or []
|
||
resolved: dict[str, Any] = dict(raw)
|
||
|
||
for field in fields:
|
||
name = field.get("name")
|
||
if not name:
|
||
continue
|
||
if name not in resolved or resolved[name] in (None, ""):
|
||
if field.get("required") and field.get("default") in (None, ""):
|
||
raise NodeError(ctx.node.id, f"缺少必填输入：{name}")
|
||
resolved[name] = field.get("default")
|
||
|
||
return {
|
||
"input": resolved,
|
||
"vars": resolved,
|
||
"nodes": {ctx.node.id: resolved},
|
||
}
|
||
|
||
|
||
async def run_output(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||
"""出口节点。按配置把散落在各节点的结果收成一个结构化成果。
|
||
|
||
配了 contract 就升级成出具契约：核对指标齐不齐、叙述里的数字
|
||
能不能回指到指标集，然后给出三档判定（formal / degraded / withheld）。
|
||
判定和缺口清单一起印在成果上——降档不是悄悄地降。
|
||
"""
|
||
mapping = ctx.cfg("fields", None)
|
||
tctx = template_context(state)
|
||
|
||
if mapping:
|
||
result: dict[str, Any] = {}
|
||
for field in mapping:
|
||
name = field.get("name")
|
||
if not name:
|
||
continue
|
||
result[name] = ctx.render(field.get("value", ""), state)
|
||
else:
|
||
# 没配映射就把最后一条消息当作结果，让新建的图开箱即用
|
||
result = {"result": tctx.get("last_message", "")}
|
||
|
||
reports = _report_fields(result, mapping, state, ctx)
|
||
contract = ctx.cfg("contract")
|
||
if contract:
|
||
cells, governed = True, False
|
||
if isinstance(contract, dict) and contract.get("report_from") not in (None, ""):
|
||
from app.engine.governance import cells_allowed, governed_formal
|
||
|
||
governed = await governed_formal(ctx.run.run_id)
|
||
cells = cells_allowed(contract, governed=governed)
|
||
result["_issuance"] = _apply_contract(contract, state, ctx, reports=reports, cells=cells, governed=governed,
|
||
fields=result)
|
||
evidence = _evidence_of(reports, contract)
|
||
if evidence:
|
||
result["_evidence"] = evidence
|
||
|
||
return {"output": result, "nodes": {ctx.node.id: result}}
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 成果字段和上游报告的对应
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def _reference(node: GraphNode) -> re.Pattern[str]:
|
||
"""字段模板里「带出了这个报告的正文」的写法。
|
||
|
||
nodes.write.text、nodes['write']['text']、整个 nodes.write（过 json 也带着正文），以及
|
||
存的就是正文的 vars.<assign_to>。nodes.write.repairs / stats / doc_artifact 只是把报告的
|
||
统计或文档 id 一起放进成果，没碰正文——算成改动的话，一份全部带引用的报告会被一个
|
||
「改写次数」字段拖成降档。模板只认路径加过滤器、没有方法调用，所以这几种写法就是全部。
|
||
"""
|
||
nid = re.escape(node.id)
|
||
head = rf"nodes\s*(?:\.\s*{nid}(?![\w-])|\[\s*['\"]{nid}['\"]\s*\])"
|
||
alternatives = [rf"{head}(?:\s*\.\s*text(?![\w-])|\s*\[\s*['\"]text['\"]\s*\]|(?!\s*[.\[]))"]
|
||
var = str(node.config.get("assign_to") or "").strip()
|
||
if var:
|
||
alternatives += [rf"vars\s*\.\s*{re.escape(var)}(?![\w-])",
|
||
rf"vars\s*\[\s*['\"]{re.escape(var)}['\"]\s*\]"]
|
||
return re.compile("|".join(alternatives))
|
||
|
||
|
||
def _report_fields(
|
||
result: dict[str, Any], mapping: Any, state: GraphState, ctx: NodeContext,
|
||
) -> list[dict[str, Any]]:
|
||
"""上游每个写出了文档的报告节点：哪些成果字段逐字就是它的原文，哪些引用了它却改动过。
|
||
|
||
逐字相等的字段，界面能按文档逐段画、逐个数点开证据；模板在报告前后拼了别的字、
|
||
或者过了过滤器，偏移就对不上了——这种字段不标证据，有契约时记一条 gap。
|
||
首尾空白不算改动：Markdown 渲染出来一样，文档的 markdown 本来就去掉了首尾空白。
|
||
"""
|
||
spec = ctx.run.spec
|
||
upstream = _ancestors(spec, ctx.node.id)
|
||
nodes = state.get("nodes") or {}
|
||
sources = {f.get("name"): f.get("value", "") for f in mapping or [] if isinstance(f, dict) and f.get("name")}
|
||
out = []
|
||
for node in spec.nodes:
|
||
payload = nodes.get(node.id)
|
||
if node.type != NodeType.REPORT or node.id not in upstream or not isinstance(payload, dict):
|
||
continue
|
||
text = payload.get("text")
|
||
if not payload.get("doc_artifact") or not isinstance(text, str) or not text.strip():
|
||
continue
|
||
verbatim = [name for name, value in result.items()
|
||
if not name.startswith("_") and isinstance(value, str) and value.strip() == text]
|
||
pattern = _reference(node)
|
||
edited = [name for name, src in sources.items() if name not in verbatim and pattern.search(
|
||
src if isinstance(src, str) else json.dumps(src, ensure_ascii=False, default=str))]
|
||
out.append({"node": node, "doc_artifact": payload["doc_artifact"], "verbatim": verbatim,
|
||
"edited": edited})
|
||
return out
|
||
|
||
|
||
def _evidence_of(reports: list[dict[str, Any]], contract: Any) -> dict[str, Any] | None:
|
||
"""成果上的 _evidence：{report_node, doc_artifact, fields}。
|
||
|
||
不管有没有契约、是不是正式运行都标——探索运行里的报告一样能点开看证据。契约的
|
||
report_from 指着的那份优先；另有报告也被原样放进了成果的，列在 others 里。
|
||
"""
|
||
shown = [r for r in reports if r["verbatim"]]
|
||
if not shown:
|
||
return None
|
||
preferred = contract.get("report_from") if isinstance(contract, dict) else None
|
||
shown.sort(key=lambda r: r["node"].id != preferred)
|
||
first, *rest = [{"report_node": r["node"].id, "doc_artifact": r["doc_artifact"], "fields": r["verbatim"]}
|
||
for r in shown]
|
||
return {**first, "others": rest} if rest else first
|
||
|
||
|
||
def _apply_contract(
|
||
contract: dict[str, Any], state: GraphState, ctx: NodeContext,
|
||
*, reports: list[dict[str, Any]] | None = None, cells: bool = True, governed: bool = False,
|
||
fields: dict[str, Any] | None = None,
|
||
) -> dict[str, Any]:
|
||
from datetime import datetime, timezone
|
||
|
||
from app.engine.governance import board_only
|
||
from app.engine.issuance import decide_tier, incomplete_gaps, no_data_gaps, sql_check_findings, trace_numbers
|
||
|
||
nodes = state.get("nodes") or {}
|
||
# 写了 report_from 就是引用模式：按报告文档逐段复核，不再按数值回指叙述
|
||
report_from = contract.get("report_from")
|
||
citations = report_from not in (None, "")
|
||
|
||
# 1) 收指标集：contract.metrics_from 指向一个或多个口径卡节点
|
||
metrics: list[dict[str, Any]] = []
|
||
calibers: list[dict[str, str]] = []
|
||
sources = contract.get("metrics_from") or []
|
||
if isinstance(sources, str):
|
||
sources = [sources]
|
||
# gaps 记的是"校验本身哪里没跑起来"，和"校验跑了但没通过"是两回事。
|
||
# 没有它，一个写错的节点名会让所有列表都是空的，看起来和"全部通过"一模一样。
|
||
gaps: list[str] = []
|
||
unresolved: list[str] = []
|
||
#: 指标 id → 它来自哪张口径卡（「销售口径 @ v2」），逐个数字回指时标出处
|
||
caliber_of: dict[str, str] = {}
|
||
for source in sources:
|
||
payload = nodes.get(source)
|
||
if isinstance(payload, dict) and payload.get("kind") == "metric_set":
|
||
metrics.extend(payload.get("metrics") or [])
|
||
calibers.append(
|
||
{
|
||
"node": source,
|
||
"caliber": payload.get("caliber", ""),
|
||
"version": payload.get("caliber_version", ""),
|
||
}
|
||
)
|
||
label = " @ ".join(str(v) for v in (payload.get("caliber"), payload.get("caliber_version")) if v)
|
||
for m in payload.get("metrics") or []:
|
||
caliber_of.setdefault(str(m.get("id")), label)
|
||
else:
|
||
# 节点不存在、还没跑到、或者根本不是口径卡——指标集就是缺的，
|
||
# 不能当作"这次没有指标要查"
|
||
unresolved.append(source)
|
||
if unresolved:
|
||
gaps.append(f"指标来源未解析：{'、'.join(unresolved[:5])}")
|
||
# 拿截断的查询结果整组算出来的指标：数回指得上，但算的只是取回的那一截，不能完整出具
|
||
gaps.extend(incomplete_gaps(metrics))
|
||
# 输入取自「查询区间里这张表没有数据」的列的指标：数回指得上、复算一致，但那个 0 不代表真的是 0，不能完整出具
|
||
gaps.extend(no_data_gaps(metrics))
|
||
# 来源查询没通过 SQL 检查（error 级：重复计算、存量跨期加总）：口径卡复算得一致，算的却是错数。
|
||
# 报告直接引用的查询要等引用核对之后才知道，缺口放在这个位置，下面一起算
|
||
sql_at = len(gaps)
|
||
|
||
# 引用模式下，缺输入记成空值的指标（口径卡 on_missing=null）就是缺：报告里引用不了它，
|
||
# 也不能因为 id 在清单里就算「齐了」。旧模式照旧只看 id，行为不变
|
||
present = {m["id"] for m in metrics if not citations or m.get("value") is not None}
|
||
missing_required = [m for m in (contract.get("required") or []) if m not in present]
|
||
missing_expected = [m for m in (contract.get("expected") or []) if m not in present]
|
||
|
||
trace = None
|
||
cited: dict[str, Any] = {}
|
||
by_metrics = not citations and board_only(contract, ctx.run.spec, ctx.node.id)
|
||
if citations:
|
||
cited = _check_citations(str(report_from), contract, state, ctx, reports or [], cells=cells,
|
||
governed=governed)
|
||
gaps.extend(cited["gaps"])
|
||
unmatched = cited["unmatched"]
|
||
elif by_metrics:
|
||
# 指标出具：没有叙述，给人看的是成果字段——字段里的数回指指标集，和叙述同一套规则；每个指标的输入要追到原始
|
||
# 数据，追不到的是「这个数有算式、没出处」，照实记成缺口
|
||
shown = _shown_text(fields or {})
|
||
trace = trace_numbers(shown, metrics, allow=contract.get("allow_numbers")) if shown.strip() else None
|
||
unmatched = trace.unmatched if trace else []
|
||
gaps.extend(_untraced_gaps(metrics, ctx))
|
||
else:
|
||
# 2) 叙述数字回指
|
||
declared_narrative = str(contract.get("narrative", ""))
|
||
narrative = ctx.render_str(declared_narrative, state)
|
||
# 模板取不到路径时 render 出来是空串，所以"声明了叙述但渲染成空"必须单独识别：
|
||
# 节点改名、vars 写错都会让回指校验静默变成空操作
|
||
if declared_narrative.strip() and not narrative.strip():
|
||
gaps.append("叙述模板渲染为空（路径可能有误）")
|
||
elif not declared_narrative.strip():
|
||
gaps.append("出具契约未设置「叙述」，未执行数字追溯")
|
||
|
||
trace = (
|
||
trace_numbers(narrative, metrics, allow=contract.get("allow_numbers"))
|
||
if narrative.strip()
|
||
else None
|
||
)
|
||
unmatched = trace.unmatched if trace else []
|
||
|
||
# 口径卡指标上的可比性（口径卡跨两份同形查询算环比、同比时自动查的）：报告引用了的（引用模式按文档，别的按数字回指上
|
||
# 的指标）、契约必需的指标两期不可比，记缺口降档；有条件可比的写进出具声明作标注
|
||
required = {str(r) for r in contract.get("required") or []}
|
||
matched_ids = {str(h.get("metric")) for h in (trace.matched if trace else []) if h.get("metric")}
|
||
compared_gaps, comparability = _comparability(
|
||
[*(cited.get("cited_metrics") or []),
|
||
*(m for m in metrics if str(m.get("id")) in required | (set() if citations else matched_ids))])
|
||
gaps.extend(compared_gaps)
|
||
|
||
# 同一条查询算出几个指标、报告又引用了它的格，只记一句（按出问题的那条查询归并，列出受影响的引用）
|
||
titles = {n.id: n.title for n in ctx.run.spec.nodes}
|
||
aliases: dict[str, str] = cited.get("aliases") or {}
|
||
|
||
def query_label(artifact: str, node_id: str | None) -> str:
|
||
from app.engine.sql_problems import node_of
|
||
|
||
title = titles.get(node_id or node_of(artifact, state.get("evidence")) or "")
|
||
alias = aliases.get(artifact)
|
||
if title and alias:
|
||
return f"查询「{title}」（{alias}）"
|
||
return f"查询「{title}」" if title else (f"查询 {alias}" if alias else "")
|
||
|
||
findings = sql_check_findings(metrics, cited.get("queries"), label=query_label)
|
||
gaps[sql_at:sql_at] = [f["text"] for f in findings]
|
||
# 出具声明另存一份结构化的：报告页在显眼处说「因为 SQL 检查没通过而降档」，不从缺口的中文里认。没有就不加这个键
|
||
from app.engine.sql_problems import node_of
|
||
|
||
for f in findings:
|
||
if not f.get("node_id") and f.get("artifact"):
|
||
f["node_id"] = node_of(f["artifact"], state.get("evidence"))
|
||
# gap 是 gaps 里对应的那一句：界面单列这几条时，据它不在「其余缺口」里重复
|
||
sql_checks = [{**{k: f[k] for k in ("query", "node_id", "problems", "refs") if f.get(k)}, "gap": f["text"]}
|
||
for f in findings]
|
||
|
||
# 声明了 metrics_from 却一个指标都没收到，同样是"没查成"
|
||
if sources and not metrics:
|
||
gaps.append("指标集为空，叙述中的数字无法追溯")
|
||
|
||
# 上游有协作团队用完轮数、按降档交付的：叙述里的话不是调度者认可的结论，
|
||
# 数字全都对得上也不能盖「完整出具」
|
||
for node_id, payload in nodes.items():
|
||
if isinstance(payload, dict) and payload.get("exhausted"):
|
||
gaps.append(f"协作团队「{titles.get(node_id, node_id)}」用完 {payload.get('rounds', '?')} "
|
||
"轮仍未完成，交付的是成员最后的回复")
|
||
# 上游节点自己报上来的缺口（双盲复算比不上、反方审稿的反驳没回应……）：只会降档，不会升档。报告节点报的分析说法那一条，
|
||
# 引用模式下出口自己复核过、记了自己的缺口的不再重复收
|
||
gaps.extend(_upstream_issuance_gaps(state, ctx, skip=cited.get("analysis_skip")))
|
||
|
||
tier = decide_tier(
|
||
missing_required=missing_required,
|
||
missing_expected=missing_expected,
|
||
unmatched=unmatched,
|
||
strict=bool(contract.get("strict")),
|
||
gaps=gaps,
|
||
unresolved=cited.get("unresolved"),
|
||
unsupported=cited.get("unsupported"),
|
||
uncited_claims=cited.get("uncited"),
|
||
claims_policy=cited.get("claims_policy"),
|
||
)
|
||
|
||
if citations:
|
||
return _declare_citations(tier, cited, calibers=calibers, metrics=metrics, gaps=gaps,
|
||
missing_required=missing_required, missing_expected=missing_expected,
|
||
report_from=str(report_from), declared_at=datetime.now(timezone.utc),
|
||
ctx=ctx, sql_checks=sql_checks, comparability=comparability)
|
||
|
||
# 逐个数字的出处。以前只有一个计数，界面标不出「这个数来自哪个指标」
|
||
matched = [
|
||
{**hit, **({"caliber": caliber_of[hit["metric"]]} if caliber_of.get(hit.get("metric")) else {})}
|
||
for hit in (trace.matched if trace else [])
|
||
][:50]
|
||
issuance = {
|
||
"tier": tier,
|
||
**({"mode": "metrics"} if by_metrics else {}),
|
||
"calibers": calibers, # 口径版本照常印
|
||
"metrics_checked": len(metrics),
|
||
"missing_required": missing_required,
|
||
"missing_expected": missing_expected, # 缺数据声明照常印
|
||
"unmatched_numbers": unmatched[:20],
|
||
"matched_numbers": len(trace.matched) if trace else 0,
|
||
"matched": matched,
|
||
# 校验没跑全的原因照实印出来——读的人要能分辨"查过都对"和"根本没查"
|
||
"gaps": gaps,
|
||
"declared_at": datetime.now(timezone.utc).isoformat(),
|
||
**({"sql_checks": sql_checks} if sql_checks else {}),
|
||
# 指标的可比性（有条件可比的标注、不可比的同时在缺口里）：只在有的时候带这个键
|
||
**({"comparability": comparability} if comparability else {}),
|
||
}
|
||
ctx.emit(
|
||
EventType.ISSUANCE,
|
||
tier=tier,
|
||
**({"mode": "metrics"} if by_metrics else {}),
|
||
missing_required=missing_required,
|
||
missing_expected=missing_expected,
|
||
unmatched=len(unmatched),
|
||
calibers=calibers,
|
||
# 降档的原因要跟着事件走：只有 gaps 的时候，时间线上的出具步骤以前只能写
|
||
# 一个光秃秃的「降档出具」，横幅还说「请对照下方声明」，下方却什么都没有
|
||
gaps=gaps,
|
||
metrics_checked=len(metrics),
|
||
matched_numbers=len(trace.matched) if trace else 0,
|
||
matched=matched,
|
||
**({"sql_checks": sql_checks} if sql_checks else {}),
|
||
)
|
||
return issuance
|
||
|
||
|
||
def _shown_text(fields: dict[str, Any]) -> str:
|
||
"""指标出具时给人看的文字：成果字段里的文字和数（成果自己的 _issuance、_evidence 不算）。表格、对象这类结构化的值不在
|
||
这里核：它们是查询结果，出处是自己的快照，不是口径卡的指标。"""
|
||
parts = []
|
||
for name, value in fields.items():
|
||
if str(name).startswith("_") or isinstance(value, bool):
|
||
continue
|
||
if isinstance(value, (str, int, float)):
|
||
parts.append(str(value))
|
||
return "\n".join(parts)
|
||
|
||
|
||
def _untraced(metrics: list[dict[str, Any]], ctx: NodeContext) -> list[dict[str, Any]]:
|
||
"""输入没追到原始数据的指标（engine/lineage 的认法）：[{metric, name, stops: [{path, node_id}], gap}]。gap 是缺口那一句，
|
||
说清停在哪个节点。metrics 可以是契约收来的全部指标（指标出具），也可以只是报告引用的那几个（引用模式，D15）。"""
|
||
from app.engine.lineage import untraced
|
||
|
||
titles = {n.id: n.title for n in ctx.run.spec.nodes}
|
||
out = []
|
||
for m in metrics:
|
||
stops = untraced(m.get("inputs") or [])
|
||
if not stops:
|
||
continue
|
||
# 停在子工作流里的（节点 id「子工作流节点 id/…」）说停在那个子工作流：这张图上看得到的是它，静态检查也这么说
|
||
where = "、".join(dict.fromkeys(
|
||
f"{s.get('path')}（停在「{titles.get(at, at or '未知来源')}」）"
|
||
for s in stops[:3] for at in [str(s.get('node_id') or '').split('/', 1)[0]]))
|
||
out.append({"metric": m.get("id"), "name": m.get("name") or m.get("id"),
|
||
"stops": [{"path": s.get("path"), "node_id": s.get("node_id")} for s in stops[:3]],
|
||
"gap": f"指标「{m.get('name') or m.get('id')}」的输入没追到原始数据：{where}"})
|
||
return out
|
||
|
||
|
||
def _untraced_gaps(metrics: list[dict[str, Any]], ctx: NodeContext) -> list[str]:
|
||
"""指标出具的缺口：输入没追到原始数据的指标，每个一句（_untraced）。"""
|
||
return [u["gap"] for u in _untraced(metrics, ctx)]
|
||
|
||
|
||
def _comparability(metrics: list[dict[str, Any]]) -> tuple[list[str], list[dict[str, Any]]]:
|
||
"""口径卡指标上的可比性（contract.MetricComparability，口径卡跨两份同形查询算环比、同比时自动做的检查）：
|
||
(缺口, 出具声明里的那一项)。metrics 是要看的指标（报告引用了的、契约必需的，调用方挑好）：两期不可比的记缺口（降档，
|
||
同 sql_check_failed），有条件可比的只标注。没有可比性记录的指标不在里面；同一条指标只看一次。"""
|
||
gaps: list[str] = []
|
||
declared: list[dict[str, Any]] = []
|
||
seen: set[int] = set()
|
||
for m in metrics:
|
||
mid = str(m.get("id") or "") if isinstance(m, dict) else ""
|
||
found = m.get("comparability") if mid else None
|
||
if not isinstance(found, dict) or id(m) in seen:
|
||
continue
|
||
seen.add(id(m))
|
||
verdict = found.get("verdict")
|
||
if verdict not in ("conditional", "incomparable"):
|
||
continue
|
||
name = m.get("name") or mid
|
||
phrase = str(found.get("phrase") or "").strip()
|
||
declared.append({"metric": mid, "name": name, "verdict": verdict, "phrase": phrase,
|
||
"reasons": [str(r) for r in found.get("reasons") or [] if r],
|
||
**({"artifact": found["artifact"]} if found.get("artifact") else {})})
|
||
if verdict == "incomparable":
|
||
gaps.append(f"指标「{name}」{phrase or '两期不可比'}")
|
||
return gaps, declared
|
||
|
||
|
||
def _upstream_issuance_gaps(state: GraphState, ctx: NodeContext,
|
||
skip: dict[str, set[str]] | None = None) -> list[str]:
|
||
"""上游节点产出里的 issuance_gaps（字符串列表），每条前面加上是哪个节点说的。
|
||
|
||
这是给「把关类」节点留的通用口子：节点觉得这次的结论不该完整出具（双盲复算两路对不上、
|
||
反方审稿的反驳没有回应），在自己的产出里写 issuance_gaps，成果节点的出具契约照单收进缺口，
|
||
出具随之降档。写的时候不用带节点名，这里统一加上「双盲复算「复核销售额」：」这样的前缀，
|
||
和别的缺口一样读得出是谁报的。节点名就是类型的叫法（拖进画布没改名）时只写一次：「反方审稿：」，
|
||
不写成「反方审稿「反方审稿」：」。缺口正文也不必再提节点类型，前缀已经说了。
|
||
|
||
只认这个成果节点的上游：另一条支路上的节点管不到这份成果。不是字符串、空白的条目跳过——
|
||
这个键谁都能写（数据整形的 JSON 模板也行），写坏了只当没写，不能让出具崩掉。
|
||
|
||
skip：{节点 id: 不再收的那几条}。引用模式的出口自己复核过报告文档的分析说法、记了自己的缺口时，报告节点报上来的同一件事
|
||
不再重复收一条（_analysis_claims）。
|
||
"""
|
||
from app.engine.labels import type_label
|
||
|
||
spec = ctx.run.spec
|
||
upstream = _ancestors(spec, ctx.node.id)
|
||
nodes = state.get("nodes") or {}
|
||
out: list[str] = []
|
||
for node in spec.nodes:
|
||
payload = nodes.get(node.id)
|
||
if node.id not in upstream or not isinstance(payload, dict):
|
||
continue
|
||
reported = payload.get("issuance_gaps")
|
||
if not isinstance(reported, list):
|
||
continue
|
||
kind = type_label(node.type)
|
||
prefix = f"{kind}：" if node.title == kind else f"{kind}「{node.title}」："
|
||
dropped = (skip or {}).get(node.id) or set()
|
||
out.extend(prefix + gap.strip() for gap in reported
|
||
if isinstance(gap, str) and gap.strip() and gap not in dropped)
|
||
return out
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 引用模式：按报告文档逐段复核
|
||
# --------------------------------------------------------------------------
|
||
|
||
#: 文档本身对不上（不是某个数没出处）：这次复核没法完成，记 gap
|
||
_INTEGRITY = {"bad_schema", "segment_mismatch", "structural_text", "render_mismatch", "eid_mismatch",
|
||
"state_mismatch"}
|
||
#: 出具声明里列多少个可疑名字、多少句没挂依据的结论（计数照实，列表只是给人看的样本）
|
||
_LISTED = 20
|
||
|
||
|
||
def _check_citations(
|
||
report_from: str, contract: dict[str, Any], state: GraphState, ctx: NodeContext,
|
||
reports: list[dict[str, Any]], *, cells: bool = True, governed: bool = False,
|
||
) -> dict[str, Any]:
|
||
"""独立复核报告文档：{matched, unmatched, unresolved, gaps, doc_artifact, stats, claims_policy, uncited,
|
||
unsupported, claims, entities}。
|
||
|
||
不信报告节点自己的统计：文档按 id 从工件库取回（取回时复验哈希），目录按报告节点
|
||
同一套参数从状态里重建（不用文档里存的那份），每个带引用的片段重新解析、重新渲染、
|
||
逐字比对。一个配错的报告节点绕不过这里。
|
||
|
||
cells：受管级别的正式运行、契约没声明 cells 时为 False，单元格引用一律判解析不了。
|
||
governed：受管级别的正式运行。可疑实体（名字哪里都找不到）只在这时按缺口降档，别处只标注。
|
||
"""
|
||
from app.core.artifact_store import load
|
||
from app.engine import analysis_claims as said
|
||
from app.engine.evidence import iter_units, ledger_enabled, resolve_ref, verify_doc
|
||
from app.engine.nodes.report import report_catalog, report_entities
|
||
|
||
out: dict[str, Any] = {"matched": [], "unmatched": [], "unresolved": [], "gaps": [],
|
||
"doc_artifact": None, "stats": None, "claims_policy": None, "uncited": None,
|
||
"unsupported": None, "claims": None, "entities": None, "queries": {}, "aliases": {},
|
||
"analysis": None, "analysis_skip": {}, "cited_metrics": [], "untraced": None}
|
||
gaps = out["gaps"]
|
||
spec = ctx.run.spec
|
||
node = spec.node_map().get(report_from)
|
||
if node is None or node.type != NodeType.REPORT:
|
||
gaps.append(f"出具契约的「报告来自」指向的「{report_from}」不是报告撰写节点，未执行引用核对")
|
||
return out
|
||
payload = (state.get("nodes") or {}).get(report_from)
|
||
doc_artifact = payload.get("doc_artifact") if isinstance(payload, dict) else None
|
||
if not doc_artifact:
|
||
gaps.append(f"报告撰写节点「{node.title}」没有产出报告文档（未执行、被跳过或失败），"
|
||
"未执行引用核对")
|
||
return out
|
||
out["doc_artifact"] = doc_artifact
|
||
try:
|
||
doc = load(doc_artifact)
|
||
except ValueError:
|
||
gaps.append(f"报告「{node.title}」的文档与其哈希不一致，可能已被修改，未执行引用核对")
|
||
return out
|
||
if not isinstance(doc, dict):
|
||
gaps.append(f"无法读取报告「{node.title}」的文档，未执行引用核对")
|
||
return out
|
||
# 文档记的是这次运行里的节点 id（子工作流里的报告带前缀，nodes/report 的 compose）
|
||
if doc.get("node_id") != ctx.run.global_id(str(report_from)) or doc.get("run_id") != ctx.run.run_id \
|
||
or doc.get("markdown") != payload.get("text"):
|
||
gaps.append(f"报告「{node.title}」的文档与本次运行中的产出不一致，未执行引用核对")
|
||
return out
|
||
|
||
catalog = report_catalog(state, spec, node)
|
||
checked = verify_doc(doc, catalog, allow_numbers=contract.get("allow_numbers"), cells_allowed=cells,
|
||
entities=report_entities(spec, node) == "link")
|
||
out["stats"] = checked["stats"]
|
||
broken: list[str] = []
|
||
flagged: set[str] = set()
|
||
suspicious: dict[str, list[dict[str, Any]]] = {"unknown_entity": [], "unverified_entity": []}
|
||
said_found: list[dict[str, Any]] = [] # 分析说法的违规（只有记了核对这一项的文档才有，_analysis_claims）
|
||
for v in checked["violations"]:
|
||
if v["code"] in said.CODES:
|
||
said_found.append(v)
|
||
continue
|
||
where = {k: v[k] for k in ("segment", "unit") if v.get(k)}
|
||
if v.get("span"):
|
||
where.update(start=v["span"][0], end=v["span"][1])
|
||
if v.get("segment"):
|
||
flagged.add(v["segment"])
|
||
if v["code"] == "uncited_number":
|
||
out["unmatched"].append({"token": v.get("text", ""), "context": v.get("context", ""), **where})
|
||
elif v["code"] == "unresolved_ref":
|
||
out["unresolved"].append({"ref": v.get("ref", ""), "message": v["message"], **where})
|
||
elif v["code"] in suspicious:
|
||
# 写 [[see:t:x]] 时违规上没有正文里的字，名字取自引用
|
||
name = v.get("text") or str(v.get("ref") or "").partition(":")[2]
|
||
suspicious[v["code"]].append({"name": name, **({"ref": v["ref"]} if v.get("ref") else {}), **where})
|
||
elif v["code"] in _INTEGRITY:
|
||
broken.append(v["message"])
|
||
if broken:
|
||
more = f"等 {len(broken)} 处" if len(broken) > 1 else ""
|
||
gaps.append(f"报告「{node.title}」的文档未通过复核：{broken[0]}{more}")
|
||
unknown, unverified = suspicious["unknown_entity"], suspicious["unverified_entity"]
|
||
if unknown and governed:
|
||
# 用户拍板：可疑实体在受管级别的正式出具里按「有缺口」降档，不拦截——反引号里的一个业务词
|
||
# 也可能被当成名字，为它不予出具太重；探索运行、已发布级别只标注
|
||
names = "".join(f"「{n}」" for n in dict.fromkeys(u["name"] for u in unknown[:5]))
|
||
more = "等" if len(unknown) > 5 else ""
|
||
gaps.append(f"报告「{node.title}」中有 {len(unknown)} 处可疑实体（{names}{more}）：本次运行的表结构、查询和"
|
||
"结果列中都没有这些名字，可能是编造的")
|
||
if unknown or unverified:
|
||
# 表结构快照不全时核对不了的名字（unverified）在哪个级别都只标注
|
||
out["entities"] = {"unknown": unknown[:_LISTED], "unverified": unverified[:_LISTED],
|
||
"counted": governed and bool(unknown)}
|
||
|
||
out["queries"], out["aliases"] = _cited_queries(doc, catalog)
|
||
# 报告直接引用了「查询区间里这张表没有数据」的列的格（查询快照的 no_data）：这些 0 不代表真的是 0，按缺口降档
|
||
gaps.extend(_cited_no_data(doc, catalog))
|
||
|
||
# 老前端要用的逐个数字出处，从复核通过的数字片段拼出来
|
||
for _, unit in iter_units(doc):
|
||
for seg in unit.get("segments") or []:
|
||
if seg.get("kind") != "number" or not seg.get("ref") or seg.get("id") in flagged:
|
||
continue
|
||
cite = resolve_ref(str(seg["ref"]), catalog, cells_allowed=cells)
|
||
if cite["status"] != "resolved" or cite["rendered"] != seg.get("text"):
|
||
continue
|
||
entry = catalog.get(cite["alias"]) or {}
|
||
start, end = seg["span"]
|
||
hit: dict[str, Any] = {"token": seg["text"], "metric": cite["locator"].get("metric")}
|
||
if cite["kind"] == "metric":
|
||
label = " @ ".join(str(v) for v in (entry.get("caliber"), entry.get("version")) if v)
|
||
if label:
|
||
hit["caliber"] = label
|
||
elif cite["kind"] == "cell":
|
||
# 查询单元格：Q1.r0.gmv（全局编号，和 doc.catalog 一致）；分析结论结果表的一格同一种写法（A3.r0.share）
|
||
hit["cell"] = f"{cite['alias']}.r{cite['locator']['row']}.{cite['locator']['column']}"
|
||
elif cite["kind"] == LEDGER_KIND:
|
||
# 分析结论的具名标量（[[a:A3.top_share]]）：A3.top_share
|
||
key = (cite.get("locator") or {}).get("key")
|
||
hit["analysis"] = f"{cite['alias']}.{key}" if key else str(cite["alias"])
|
||
else:
|
||
hit["input"] = cite["locator"].get("field")
|
||
out["matched"].append({**hit, "segment": seg["id"], "unit": unit["id"], "start": start, "end": end,
|
||
"span": [start, end], "eid": cite["eid"]})
|
||
|
||
_claims(out, contract, payload, evidence_on=ledger_enabled(ctx.run), governed=governed,
|
||
uncited=checked.get("uncited") or [], doc=doc, title=node.title)
|
||
_analysis_claims(out, doc, payload, said_found, title=node.title, node_id=str(report_from))
|
||
|
||
# 报告引用的指标（[[m:]] 片段和句末依据里的 m:）：可比性按它们看（_apply_contract）；输入追不到原始数据的，受管级别的
|
||
# 正式运行记缺口（降档，和证据面板「追不到原始数据」的说法一致：代码节点自称取数也不能完整出具），其余写进出具声明
|
||
out["cited_metrics"] = _cited_metrics(doc, catalog, state)
|
||
if stray := _untraced(out["cited_metrics"], ctx):
|
||
if governed:
|
||
gaps.extend(u["gap"] for u in stray)
|
||
out["untraced"] = {"metrics": [{k: u[k] for k in ("metric", "name", "stops")} for u in stray[:_LISTED]],
|
||
"counted": governed}
|
||
|
||
mine = next((r for r in reports if r["node"].id == report_from), None)
|
||
edited = mine["edited"] if mine else []
|
||
for name in edited:
|
||
gaps.append(f"成果字段「{name}」改动了报告，无法逐段对应")
|
||
if not (mine and mine["verbatim"]) and not edited:
|
||
gaps.append(f"成果中没有任何字段是报告「{node.title}」的原文：读者看到的内容未经过引用核对")
|
||
return out
|
||
|
||
|
||
#: 没挂依据的结论句怎么处置，从松到严。认不出的写法按 degrade 算：写错一个词不能让缺口悄悄消失
|
||
_ON_UNCITED = ("ignore", "degrade", "withhold")
|
||
#: 证据相矛盾的结论句怎么处置（claims: judge 的 on_unsupported），从松到严。认不出的写法同样按 degrade 算
|
||
_ON_UNSUPPORTED = ("degrade", "withhold")
|
||
#: 要求结论句挂依据的策略：judge 在这一点上和 require_citation 一样严，另外再按裁判的判定判档
|
||
_CITING = ("require_citation", "judge")
|
||
|
||
#: 报告的正文写在代码块里、一句结论都没切出来时的缺口（_fenced_body）。不带核查节点的前缀：界面归在「校验未全部完成」下
|
||
GAP_FENCED_BODY = "报告「{title}」的正文写在代码块里，结论句未能核对"
|
||
#: 写的是文字的代码块：信息串为空、markdown、md（报告节点去整篇外层围栏时认的那几个，report._PROSE_INFO），另加 text、
|
||
#: plaintext、txt、plain 这几种纯文字的写法。sql、python、json 这类是模型真要贴的代码
|
||
_PROSE_FENCES = frozenset({"", "markdown", "md", "text", "plaintext", "txt", "plain"})
|
||
_CJK = re.compile(r"[\u4e00-\u9fff]")
|
||
#: 中文的句读：一行里有汉字、又有它们的，是在写句子（SQL 里的中文列名、别名不带这些）
|
||
_CJK_PUNCT = re.compile(r"[，。！？；：]")
|
||
#: 一行里连着六个以上汉字：信息串写明是文字的块里，这就是在写中文。表格的列名、地区名、门店名这类标签多在五个字以内
|
||
_CJK_RUN = re.compile(r"[\u4e00-\u9fff]{6,}")
|
||
#: 英文的词串：单词之间只隔空格、逗号。五个以上的单词紧跟句末标点算一句（_en_sentence）
|
||
_EN_RUN = re.compile(r"[A-Za-z]+(?:[ ,]+[A-Za-z]+)*")
|
||
_EN_WORD = re.compile(r"[A-Za-z]+")
|
||
#: 整行是注释（-- // # /* */ 开头），和行尾隔着空白的注释：写着中文句子也是代码的一部分。# 开头的 Markdown 标题本来
|
||
#: 也不算结论句（evidence._scan_blocks 把它切成 heading）
|
||
_COMMENT_LINE = re.compile(r"\s*(?:--|//|#|/\*|\*/)")
|
||
_TRAILING_COMMENT = re.compile(r"\s(?:--|//|#).*")
|
||
#: 看得出是代码的行里，-- // 紧挨着前面的字也是注释（SELECT SUM(amount)--按区域汇总）：字符串已经先去掉了
|
||
_CODE_COMMENT = re.compile(r"(?:--|//|\s#).*")
|
||
#: 写明是文字的块里顶格写的 Markdown 标题：标题前后是两段（_code_lines）
|
||
_MD_HEADING = re.compile(r"#{1,6}\s")
|
||
#: 代码里的字符串字面量（'…'、"…"，可带 f、r、b、u、N 这类前缀）：里面的中文标点是数据。引号前面紧挨着字母、数字的
|
||
#: 不算开头（英文的 it's、don't）
|
||
_STRING_LITERAL = re.compile(r"(?<![A-Za-z0-9_])[rRbBuUfFnN]{0,2}(?:'[^'\n]*'|\"[^\"\n]*\")")
|
||
#: 可以跨行的块注释、多行字符串：开头的写法，和它对应的收尾
|
||
_SPAN_OPEN = re.compile(r"/\*|\"\"\"|'''")
|
||
_SPAN_CLOSE = {"/*": "*/", '"""': '"""', "'''": "'''"}
|
||
#: SQL、Python 语句的开头。SQL 的关键字只认全小写、全大写的写法：「From 数据看……」「Update ……」这类首字母大写的是
|
||
#: 正文里的英文词；def、class 要写成定义的样子（def 名字(、class 名字: ），「class A 门店……」说的是门店的分类
|
||
_CODE_LINE = re.compile(r"\s*(?:(?:select|with|from|where|join|group\s+by|order\s+by|having|union|insert|update|delete"
|
||
r"|create|import|return|SELECT|WITH|FROM|WHERE|JOIN|GROUP\s+BY|ORDER\s+BY|HAVING|UNION|INSERT"
|
||
r"|UPDATE|DELETE|CREATE)\b|(?:def|class)\s+[A-Za-z_]\w*\s*[(:])")
|
||
#: YAML 的键（「columns:」「- name: 退款占比」），只认小写开头的：正文里的「Note:」「Q1:」是标签。值为空、下面缩进着
|
||
#: 列表或子键的另记一种（_YAML_NEST）
|
||
_YAML_KEY = re.compile(r"\s*(?:-\s+)?[a-z_][\w.-]*:(?:\s|$)")
|
||
_YAML_NEST = re.compile(r"\s*(?:-\s+)?[a-z_][\w.-]*:\s*$")
|
||
#: 块里再嵌的代码块：~~~ 或三个以上反引号开头。compose_doc 只认整行三个反引号的围栏，那样的一行外层就收了，块里剩下的
|
||
#: 嵌套写法是 ~~~sql、````sql 这些
|
||
_NESTED_FENCE = re.compile(r"\s*(`{3,}|~{3,})(.*)")
|
||
#: 像数据、公式的一行：前面有字的等号（公式、赋值、比较）和箭头（连起的流程），画目录树的制表符，JSON 的对象、数组开头
|
||
#: （{"、["）和「"键":」，函数调用（print(、groupby(），带扩展名的文件名（周报.xlsx）。行首的「→ 」「=> 」是列表的
|
||
#: 记号，不算
|
||
_DATA_LINE = re.compile(r"\S\s*(?:[=＝→⟶⇒]|->)|[\u2500-\u257f]|^\s*[\[{]\s*[\"'\[{]|[\"']\s*:"
|
||
r"|(?<![\w.])[A-Za-z_][\w.]*\(|(?<=[^\s.])\.[A-Za-z][A-Za-z0-9]{1,4}(?![\w.])")
|
||
|
||
|
||
def _en_sentence(line: str) -> bool:
|
||
"""一句英文：五个以上的单词（之间只隔空格、逗号）紧跟句末标点，后面是空白或行尾。按词串逐个数，不靠正则回溯：
|
||
回溯的写法遇上很长、不带句末标点的一行，耗时随行长平方增长，出口复核又是在事件循环上同步执行的。"""
|
||
for run in _EN_RUN.finditer(line):
|
||
end = run.end()
|
||
if line[end:end + 1] in (".", "!", "?") and (end + 1 == len(line) or line[end + 1].isspace()) \
|
||
and len(_EN_WORD.findall(run.group())) >= 5:
|
||
return True
|
||
return False
|
||
|
||
|
||
def _sentence(line: str) -> bool:
|
||
"""成句：两个以上汉字又有中文句读，或者是一句英文。"""
|
||
return (len(_CJK.findall(line)) >= 2 and _CJK_PUNCT.search(line) is not None) or _en_sentence(line)
|
||
|
||
|
||
def _concluding(line: str) -> bool:
|
||
"""按正文的规矩这一行会算一句结论（evidence._classify 认文字结论的那条）：带方向词、因果词，又不只是在下定义。
|
||
只认有两个以上汉字的行：drop、rise 这类英文词在代码里也常见。数字不看，代码块里的数字报告节点按裸数字查。"""
|
||
from app.engine.evidence import _CAUSAL, _DEFINITION, _DIRECTIONAL
|
||
|
||
return len(_CJK.findall(line)) >= 2 and _DIRECTIONAL.search(line) is not None \
|
||
and not (_DEFINITION.search(line) is not None and _CAUSAL.search(line) is None)
|
||
|
||
|
||
def _written(line: str) -> bool:
|
||
"""信息串写明是文字（markdown、text 这类）的块里，不是代码的这一行在写正文：成句、按正文的规矩会算结论，或者一行里
|
||
连着六个以上汉字。不看列表标记：标记的写法太多（- + 1. 1、 一、 （1） · ●，也常不带标记），补一种漏一种；也不去
|
||
字符串、行尾注释，# -- 和半角引号都是正文里的符号。整行是注释、标题的不算。"""
|
||
if _COMMENT_LINE.match(line):
|
||
return False
|
||
return _sentence(line) or _CJK_RUN.search(line) is not None or _concluding(line)
|
||
|
||
|
||
def _prose_line(line: str, *, code: bool) -> bool:
|
||
"""代码的行（code），或者信息串为空的块里的行，在写正文：去掉字符串字面量和行尾注释之后成句；不是代码的行另认按正文
|
||
的规矩会算结论的（不带句读的列表项、不带标记的行）。不带句读、也不像结论的不算：YAML 里的列名、diff 里改的口径和
|
||
列表项没有分别。整行是注释的不算。"""
|
||
if _COMMENT_LINE.match(line):
|
||
return False
|
||
line = (_CODE_COMMENT if code else _TRAILING_COMMENT).sub("", _STRING_LITERAL.sub(" ", line), count=1)
|
||
return _sentence(line) or (not code and _concluding(line))
|
||
|
||
|
||
def _blank_spans(text: str) -> str:
|
||
"""块注释（/* … */）、三个引号括起的多行字符串涂成空格，换行留着，块里还是原来那几行：里面写着中文句子也是代码的
|
||
一部分。没收尾的不动，开头那一行照注释行跳过。"""
|
||
out: list[str] = []
|
||
at = 0
|
||
while (start := _SPAN_OPEN.search(text, at)) is not None:
|
||
end = text.find(_SPAN_CLOSE[start.group()], start.end())
|
||
if end == -1:
|
||
break
|
||
stop = end + len(_SPAN_CLOSE[start.group()])
|
||
out += [text[at:start.start()], "\n".join(" " * len(part) for part in text[start.start():stop].split("\n"))]
|
||
at = stop
|
||
return "".join(out) + text[at:]
|
||
|
||
|
||
def _diff(lines: list[str], *, labelled: bool) -> bool:
|
||
"""整块是 diff：有 - 开头的行也有 + 开头的行，其余是空格开头的上下文行、@@ 开头的块头。diff 里一处改动先列删掉的行、
|
||
再列加上的行，第一个改动行就是 + 的，是 + / - 开头的正负面要点（有 @@ 块头的照旧算 diff）。信息串写明是文字的块里，
|
||
+ - 后面空一格的是 Markdown 的列表项，只有紧挨着字写的（-按下单日期统计）才是 diff。"""
|
||
heads = [line[:1] for line in lines if line.strip()]
|
||
if not {"+", "-"} <= set(heads) <= {"+", "-", " ", "@"}:
|
||
return False
|
||
if labelled and any(line[:2] in ("+ ", "- ") for line in lines):
|
||
return False
|
||
return "@" in heads or next(h for h in heads if h in ("+", "-")) == "-"
|
||
|
||
|
||
def _data_like(line: str) -> bool:
|
||
"""像数据、公式的一行（_DATA_LINE），或者逗号隔开的一行表头、数据（两个以上半角逗号，逗号后面都不空格）。"""
|
||
return _DATA_LINE.search(line) is not None or (line.count(",") >= 2 and ", " not in line)
|
||
|
||
|
||
def _yaml(lines: list[str], *, labelled: bool) -> bool:
|
||
"""一段是 YAML：第一行是小写开头的键，其余每一行都是键或者缩进的行（键下面的列表项、子键），注释行不看。写明是文字的
|
||
块里另要有值为空、往下缩进的键：一行一个「键: 一句话」的（「- note: 华南销售额超过华北」）多半是正文里的标签。"""
|
||
body = [line for line in lines if line.strip() and not _COMMENT_LINE.match(line)]
|
||
if not body or not _YAML_KEY.match(body[0]):
|
||
return False
|
||
if not all(_YAML_KEY.match(line) or line[:1].isspace() for line in body[1:]):
|
||
return False
|
||
return not labelled or any(_YAML_NEST.match(line) for line in body)
|
||
|
||
|
||
def _bare_fence(line: str) -> str:
|
||
"""整行只写着围栏（``` 或 ~~~，不带信息串）的，返回这串围栏，能给块里再嵌的代码块收尾；别的行返回空串。"""
|
||
fence = _NESTED_FENCE.match(line)
|
||
return fence.group(1) if fence and not fence.group(2).strip() else ""
|
||
|
||
|
||
def _code_lines(lines: list[str], raw: list[str], *, labelled: bool) -> list[bool]:
|
||
"""逐行标出哪些是代码：块里再嵌的代码块（~~~sql），SQL、Python 的语句（从开头那一行到这一段结束，续行也算；前面的
|
||
注释行本来就不算正文），整段是 YAML 的（_yaml），像数据、公式的行（_data_like）。
|
||
|
||
lines 是涂掉块注释、多行字符串之后的行，raw 是原样的行：一段到空行为止，只写着注释的行不算空行；写明是文字的块里，
|
||
顶格的标题也另起一段。不把整块当代码：正文里另贴一段 SQL、嵌一个 ~~~sql，或者有一行像 YAML 的键，别的行照样按正文认。
|
||
再嵌的代码块要收得了才算（后面有同一种符号、不比开头短的收尾），开了没收的那一行照普通的行看。"""
|
||
code = [False] * len(lines)
|
||
paragraph: list[int] = []
|
||
fence = ""
|
||
# 每一行后面最长的收尾围栏（整行只有 ``` 或 ~~~），按符号分开记：开了收不了的围栏不算，照普通的行看，免得后面的
|
||
# 正文都被当成代码
|
||
reach = {"`": [0] * (len(lines) + 1), "~": [0] * (len(lines) + 1)}
|
||
for k in range(len(lines) - 1, -1, -1):
|
||
for after in reach.values():
|
||
after[k] = after[k + 1]
|
||
if bare := _bare_fence(lines[k]):
|
||
reach[bare[0]][k] = max(reach[bare[0]][k], len(bare))
|
||
|
||
def close() -> None:
|
||
statement = _yaml([lines[k] for k in paragraph], labelled=labelled)
|
||
for k in paragraph:
|
||
statement = statement or _CODE_LINE.match(lines[k]) is not None
|
||
code[k] = statement or _data_like(lines[k])
|
||
paragraph.clear()
|
||
|
||
for k, line in enumerate(lines):
|
||
if fence:
|
||
code[k] = True
|
||
closing = _bare_fence(line)
|
||
if closing[:1] == fence[0] and len(closing) >= len(fence):
|
||
fence = ""
|
||
continue
|
||
if (opening := _NESTED_FENCE.match(line)) and reach[opening.group(1)[0]][k + 1] >= len(opening.group(1)):
|
||
close()
|
||
fence, code[k] = opening.group(1), True
|
||
continue
|
||
if not raw[k].strip() or (labelled and _MD_HEADING.match(line)):
|
||
close()
|
||
continue
|
||
paragraph.append(k)
|
||
close()
|
||
return code
|
||
|
||
|
||
def _fenced_body(doc: dict[str, Any]) -> bool:
|
||
"""报告的正文写在代码块里：文档里一句结论都没有，却有写文字的代码块（_PROSE_FENCES），里面写着引用（片段里的
|
||
[[m:…]] 这类标记、句子上挂的 [[see:…]] 依据，代码里不会这样写）或者在写正文。
|
||
|
||
报告节点只在整篇正好是一个代码块时去掉外层围栏（report._unfenced），围栏外有前言、只有开头没有收尾、信息串写成
|
||
text 的原样交给 compose_doc：整段正文落在代码块里，切不出结论句，裁判没有句子可判、反方审稿记 no_claims，没挂依据
|
||
的、证据相矛盾的结论句都数出 0 句，出具就成了完整出具。有一句结论的不算：正文没被整个藏起来，结论句照常核对。
|
||
|
||
在不在写正文逐行看。先标出哪些行是代码（_code_lines；信息串为空、整块是 diff 的全是代码），代码的行只认去掉注释、
|
||
字符串之后成句的（_prose_line），中文别名、YAML 的值、公式里的「最高」「占比」这类词都不算。其余的行：信息串写明
|
||
是文字的按 _written，信息串为空的块里也常是代码，去掉注释、字符串之后再看（_prose_line）。一块里只要有一行在写正文
|
||
就算，所以宁可把拿不准的行当代码：藏起来的正文总有别的行认得出，贴的一段代码、公式里却一行认错就会记错缺口。"""
|
||
from app.engine.evidence import iter_units
|
||
|
||
units = list(iter_units(doc))
|
||
if any(unit.get("kind") == "claim" for _, unit in units):
|
||
return False
|
||
for block, unit in units:
|
||
lang = str(block.get("lang") or "").lower()
|
||
if block.get("type") != "code" or lang not in _PROSE_FENCES:
|
||
continue
|
||
segments = [s for s in unit.get("segments") or [] if isinstance(s, dict)]
|
||
if unit.get("see") or any(s.get("ref") for s in segments):
|
||
return True
|
||
# 围栏那两行是挂在块上的结构片段，不是块里写的内容：算进来的话，开头那行 ```markdown 会被当成块里再嵌的代码块，
|
||
# 整块是不是 diff 也看不准
|
||
text = "".join(str(s.get("text") or "") for s in segments if s.get("kind") != "structural")
|
||
raw, lines = text.split("\n"), _blank_spans(text).split("\n")
|
||
code = [True] * len(lines) if _diff(raw, labelled=bool(lang)) else _code_lines(lines, raw, labelled=bool(lang))
|
||
if any(_written(line) if lang and not is_code else _prose_line(line, code=is_code)
|
||
for line, is_code in zip(lines, code)):
|
||
return True
|
||
return False
|
||
|
||
|
||
def _claims(out: dict[str, Any], contract: dict[str, Any], payload: dict[str, Any], *, evidence_on: bool,
|
||
governed: bool = False, uncited: list[dict[str, Any]], doc: dict[str, Any] | None = None,
|
||
title: str = "") -> None:
|
||
"""结论句策略：没挂依据的、裁判判为证据不支持的结论句按 claims_policy 交给 decide_tier 判档。
|
||
|
||
两处可以要求挂依据：报告节点写作时记下的 claims（节点产出里的，写作提示按它要求过），和契约自己写的
|
||
require_citation / judge（对象形式可以另写 on_uncited）。两处都写了取更严的那个——契约只能收紧、不能放松：
|
||
报告节点要求了，契约写 ignore 或 off 也照旧计入缺口，写 withhold 就收紧成不予出具。受管级别的正式运行
|
||
里 ignore 一律当 degrade：没有哪道门禁查契约里的 claims，不能让它把用户要的底线拉低。
|
||
|
||
报告节点写的是 judge 时，判定在文档里（按哈希取回的那份：判定是模型给的，出口复算不了，只能信封存的
|
||
文档）。节点里判过的（正式运行，以及交给人工审批的探索运行）：证据相矛盾的（contradicted，旧文档里的 unsupported 同样算）按 on_unsupported
|
||
判档（报告节点和契约取更严的）；证据不足的（insufficient）记一条缺口——最多降档，不会不予出具，也不会
|
||
完整出具；没判完的（裁判没跑成、触顶）把裁判摘要里的缺口照抄进来——不能判完整出具，也不当成「矛盾」。
|
||
其余探索运行按需裁判，这里只标注。契约要 judge、报告节点却没开的，照实记缺口：出口替不了报告节点去裁判。
|
||
|
||
数的是这里重新核对出来的 uncited，不信文档里记的 cites。升级前写的报告（节点产出里没有 claims 这个键，
|
||
包括跨着升级还没跑完的运行）一律不管，契约写了也只记一条缺口：报告生成于升级前，结论句未核对。
|
||
|
||
要核对结论句、正文却写在代码块里（一句结论都没切出来，_fenced_body）的，另记一条缺口：数出 0 句不是「核对过、
|
||
没问题」。
|
||
"""
|
||
written = payload.get("claims") if isinstance(payload, dict) else None
|
||
# 报告节点升级后才在产出里记 claims（off 也记）：没有这个键就是升级前写的报告
|
||
upgraded = evidence_on and isinstance(payload, dict) and "claims" in payload
|
||
declared = contract.get("claims")
|
||
name = declared.get("policy") if isinstance(declared, dict) else declared
|
||
asks: list[str] = []
|
||
if upgraded and written in _CITING:
|
||
asks.append("degrade")
|
||
if upgraded and name in _CITING:
|
||
wanted = declared.get("on_uncited") if isinstance(declared, dict) else None
|
||
asks.append(wanted if wanted in _ON_UNCITED else "degrade")
|
||
elif name not in (None, "", "off"):
|
||
out["gaps"].append("出具契约要求核对结论句，但这份报告生成于系统升级前，没有结论句核对记录，结论句未核对")
|
||
judged = upgraded and written == "judge"
|
||
if upgraded and name == "judge" and not judged:
|
||
out["gaps"].append(f"出具契约要求结论句由模型裁判，但报告「{title}」未开启模型裁判，结论句未裁判")
|
||
if not asks:
|
||
return
|
||
if doc is not None and _fenced_body(doc):
|
||
# 正文在代码块里，一句结论都没切出来：没挂依据的、证据相矛盾的都是 0 句，不能当成「核对过、没问题」
|
||
out["gaps"].append(GAP_FENCED_BODY.format(title=title))
|
||
on_uncited = max(asks, key=_ON_UNCITED.index)
|
||
if governed and on_uncited == "ignore":
|
||
on_uncited = "degrade"
|
||
extra = {"on_uncited": on_uncited} if on_uncited != "degrade" else {}
|
||
if not judged:
|
||
policy: Any = "require_citation" if on_uncited == "degrade" else \
|
||
{"policy": "require_citation", "on_uncited": on_uncited}
|
||
out["claims_policy"], out["uncited"] = policy, uncited
|
||
out["claims"] = {"policy": "require_citation", "uncited_claims": len(uncited), "uncited": uncited[:_LISTED],
|
||
**extra}
|
||
return
|
||
|
||
from app.engine.evidence import iter_units
|
||
from app.engine.judge import COUNT_KEYS, UNSUPPORTED
|
||
|
||
summary = (doc or {}).get("judge")
|
||
if not isinstance(summary, dict):
|
||
# 报告节点开了 judge、文档里却没有裁判摘要（不该发生）：当作没判，不能判完整出具
|
||
summary = {"mode": "inline", "counts": {}, "unjudged": {}, "limits_hit": [], "complete": False,
|
||
"gaps": ["已开启模型裁判，但文档中没有裁判结果，结论句未裁判"]}
|
||
asked = [summary.get("on_unsupported")]
|
||
if name == "judge" and isinstance(declared, dict):
|
||
asked.append(declared.get("on_unsupported"))
|
||
on_unsupported = max((a if a in _ON_UNSUPPORTED else "degrade" for a in asked), key=_ON_UNSUPPORTED.index)
|
||
inline = summary.get("mode") == "inline"
|
||
flagged: dict[str, list[dict[str, Any]]] = {s: [] for s in ("contradicted", "insufficient", UNSUPPORTED, "partial")}
|
||
if inline:
|
||
# 按需裁判的判定不在文档里（在封存之后追加的 evidence.judged 事件里），这里只数节点里判过的
|
||
markdown = str((doc or {}).get("markdown") or "")
|
||
for _, unit in iter_units(doc or {}):
|
||
verdict = unit.get("verdict")
|
||
status = verdict.get("status") if isinstance(verdict, dict) else None
|
||
if status in flagged:
|
||
span = unit.get("span") or [0, 0]
|
||
flagged[status].append({"unit": unit.get("id"), "span": span, "text": markdown[span[0]:span[1]],
|
||
"rationale": verdict.get("rationale") or "",
|
||
**({"missing": verdict["missing"]} if status == "insufficient"
|
||
and verdict.get("missing") else {})})
|
||
out["gaps"].extend(f"报告「{title}」{gap}" for gap in summary.get("gaps") or [])
|
||
if lacking := flagged["insufficient"]:
|
||
# 证据不足按缺口处理：最多降档，不因 on_unsupported 不予出具，也不能完整出具
|
||
missing = list(dict.fromkeys(i["missing"] for i in lacking if i.get("missing")))
|
||
what = f"（缺少：{'、'.join(missing[:5])}{'等' if len(missing) > 5 else ''}）" if missing else ""
|
||
out["gaps"].append(f"报告「{title}」有 {len(lacking)} 句结论证据不足{what}：裁判未能在证据中找到判断所需的"
|
||
"信息，这些结论未经核实")
|
||
counts = {k: int((summary.get("counts") or {}).get(k) or 0) for k in COUNT_KEYS}
|
||
# 证据相矛盾的按 on_unsupported 判档；旧文档里的 unsupported（当时没分矛盾和不足）照旧当 contradicted
|
||
contradicted = [*flagged["contradicted"], *flagged[UNSUPPORTED]]
|
||
out["claims_policy"] = {"policy": "judge", "on_unsupported": on_unsupported, "on_uncited": on_uncited}
|
||
out["uncited"], out["unsupported"] = uncited, contradicted
|
||
out["claims"] = {
|
||
"policy": "judge", "uncited_claims": len(uncited), "uncited": uncited[:_LISTED], **extra,
|
||
"on_unsupported": on_unsupported, "mode": "inline" if inline else "on_demand", "counts": counts,
|
||
"contradicted": flagged["contradicted"][:_LISTED], "insufficient": flagged["insufficient"][:_LISTED],
|
||
"unsupported": flagged[UNSUPPORTED][:_LISTED], "partial": flagged["partial"][:_LISTED],
|
||
"unjudged": dict(summary.get("unjudged") or {}), "limits_hit": list(summary.get("limits_hit") or []),
|
||
"complete": bool(summary.get("complete")), "model": summary.get("model"),
|
||
}
|
||
|
||
|
||
#: 报告节点 analysis_claims 的三种处置，从松到严
|
||
_ANALYSIS_ORDER = ("off", "check", "strict")
|
||
|
||
|
||
def _analysis_claims(out: dict[str, Any], doc: dict[str, Any], payload: Any, found: list[dict[str, Any]], *,
|
||
title: str, node_id: str) -> None:
|
||
"""分析说法核对的出口独立复核（00-总方案 §3.5、D9）：found 是 verify_doc 这次重新核出的说法违规（只有文档记了
|
||
analysis_claims 那一项才核，老文档复核结果不变）。
|
||
|
||
处置取报告文档记下的和报告节点产出里写的更严的那个：strict 有违规就记一条缺口（降档），报告节点报上来的同一条
|
||
（产出的 analysis_gap）不再重复收；check 只写进出具声明作标注、不影响档位。要求 strict、文档里却没有核对记录的
|
||
（不该发生），照实记一条缺口：出口替不了报告节点去核。结果放进 out["analysis"]（出具声明的 analysis_claims），
|
||
strict 和有违规时才有。"""
|
||
from app.engine import analysis_claims as said
|
||
|
||
found_mode = said.recorded(doc)
|
||
written = payload.get("analysis_claims") if isinstance(payload, dict) else None
|
||
modes = [m for m in (found_mode[0] if found_mode else None, written) if m in _ANALYSIS_ORDER]
|
||
mode = max(modes, key=_ANALYSIS_ORDER.index) if modes else None
|
||
if mode in (None, "off"):
|
||
return
|
||
if found_mode is None:
|
||
if mode == "strict":
|
||
out["gaps"].append(f"报告「{title}」要求严格核对分析说法，但文档中没有分析说法的核对记录，分析说法未核对")
|
||
return
|
||
counted = mode == "strict"
|
||
if found and counted:
|
||
out["gaps"].append(f"报告「{title}」有 {len(found)} 处分析说法未通过核对（{said.summary(found)}）："
|
||
"判断要挂对应的分析结论并照判定写，原因要有知识库原话或运行输入作依据")
|
||
if isinstance(payload, dict) and isinstance(payload.get("analysis_gap"), str):
|
||
out["analysis_skip"] = {node_id: {payload["analysis_gap"]}}
|
||
if found or counted:
|
||
keep = ("code", "text", "claim", "op", "analyses", "unit", "segment", "span", "message", "context")
|
||
out["analysis"] = {"mode": mode, "count": len(found), "counts": said.counts(found), "counted": counted,
|
||
"violations": [{k: v[k] for k in keep if v.get(k) not in (None, "", [])}
|
||
for v in found[:_LISTED]]}
|
||
|
||
|
||
def _cited_metrics(doc: dict[str, Any], catalog: dict[str, Any], state: GraphState) -> list[dict[str, Any]]:
|
||
"""报告引用了的指标（[[m:]] 片段和句末 [[see:m:…]]）：口径卡产出里的那一条（带 inputs、comparability），按引用的先后、
|
||
同一张卡的同一个指标只算一次。目录里认不出、卡的产出里找不到的跳过。"""
|
||
from app.engine.evidence import iter_units
|
||
|
||
nodes = state.get("nodes") or {}
|
||
out: list[dict[str, Any]] = []
|
||
seen: set[tuple[str, str]] = set()
|
||
for _, unit in iter_units(doc):
|
||
cites = [s.get("cite") for s in unit.get("segments") or [] if s.get("ref") and not s.get("auto")]
|
||
for cite in [*cites, *(unit.get("see") or [])]:
|
||
if not isinstance(cite, dict) or cite.get("kind") != "metric" or cite.get("status") != "resolved":
|
||
continue
|
||
entry = catalog.get(cite.get("alias")) or {}
|
||
card, mid = str(entry.get("node_id") or ""), str((entry.get("locator") or {}).get("metric") or "")
|
||
if entry.get("kind") != "metric" or not card or not mid or (card, mid) in seen:
|
||
continue
|
||
seen.add((card, mid))
|
||
payload = nodes.get(card)
|
||
metric = next((m for m in (payload.get("metrics") or []) if isinstance(m, dict) and str(m.get("id")) == mid),
|
||
None) if isinstance(payload, dict) else None
|
||
if metric is not None:
|
||
out.append(metric)
|
||
return out
|
||
|
||
|
||
def _cited_no_data(doc: dict[str, Any], catalog: dict[str, Any]) -> list[str]:
|
||
"""报告直接引用的格里，落在「查询区间里这张表没有数据」的列上的（查询快照的 no_data，data/coverage.py）：按列一句缺口
|
||
「报告引用的 Q1「上周到访人数」：所依据的查询区间里这张表没有数据（…）」。读不出来的快照不算。"""
|
||
from app.core.artifact_store import load
|
||
from app.data.coverage import NO_DATA_SHORT, no_data_of
|
||
from app.engine.evidence import iter_units
|
||
|
||
snapshots: dict[str, Any] = {}
|
||
out: list[str] = []
|
||
seen: set[tuple[str, str]] = set()
|
||
for _, unit in iter_units(doc):
|
||
for seg in unit.get("segments") or []:
|
||
cite = seg.get("cite") or {}
|
||
if cite.get("kind") != "cell" or cite.get("status") != "resolved" or not cite.get("alias"):
|
||
continue
|
||
alias = str(cite["alias"])
|
||
column = str((cite.get("locator") or {}).get("column") or "")
|
||
artifact = (catalog.get(alias) or {}).get("artifact")
|
||
if not column or not isinstance(artifact, str) or (alias, column) in seen:
|
||
continue
|
||
seen.add((alias, column))
|
||
if artifact not in snapshots:
|
||
try:
|
||
snapshots[artifact] = load(artifact)
|
||
except (ValueError, OSError):
|
||
snapshots[artifact] = None
|
||
if (found := no_data_of(snapshots[artifact], column)) is not None:
|
||
out.append(f"报告引用的 {alias}「{column}」：{NO_DATA_SHORT}（{found.get('message') or ''}）")
|
||
return out
|
||
|
||
|
||
def _cited_queries(doc: dict[str, Any], catalog: dict[str, Any]) -> tuple[dict[str, list[str]], dict[str, str]]:
|
||
"""报告直接引用的查询：({查询快照 id: [引用的写法…]}, {目录里每个查询的快照 id: 编号})。
|
||
|
||
单元格（[[v:Q1.r0.gmv]]，[[table:]] 展开后也是一格一格的引用）按查询数：一格写「Q1 第 1 行「gmv」」，
|
||
几格写「Q1 的 3 格」；只拿它当依据（[[see:Q1]]）的写「Q1（结论依据）」。出具契约据此判这些查询有没有通过
|
||
SQL 检查（issuance.sql_check_gaps）。
|
||
|
||
分析结论按它的工件记（sql_problems.query_problems 顺着结论的输入找到查询）：结果表的格同上（A3 第 1 行「share」）；
|
||
判定语、具名标量写「A3（判定语）」「A3「top_share」」，当依据（[[see:A3]]）的写「A3（结论依据）」——引用了一份分析
|
||
结论，它的输入查询上的问题照样计入缺口。
|
||
"""
|
||
from app.engine.evidence import iter_units
|
||
|
||
cells: dict[str, list[tuple[Any, Any]]] = {}
|
||
said: dict[str, list[str]] = {} # 分析结论的判定语、具名标量：{A 编号: [键或 None…]}
|
||
support: list[str] = []
|
||
for _, unit in iter_units(doc):
|
||
for seg in unit.get("segments") or []:
|
||
cite = seg.get("cite") or {}
|
||
if cite.get("status") != "resolved" or not cite.get("alias"):
|
||
continue
|
||
if cite.get("kind") == "cell":
|
||
loc = cite.get("locator") or {}
|
||
cells.setdefault(str(cite["alias"]), [])
|
||
if (loc.get("row"), loc.get("column")) not in cells[str(cite["alias"])]:
|
||
cells[str(cite["alias"])].append((loc.get("row"), loc.get("column")))
|
||
elif cite.get("kind") == LEDGER_KIND:
|
||
keys = said.setdefault(str(cite["alias"]), [])
|
||
key = (cite.get("locator") or {}).get("key")
|
||
if key not in keys:
|
||
keys.append(key)
|
||
for cite in unit.get("see") or []:
|
||
alias = str((cite or {}).get("alias") or "")
|
||
if (catalog.get(alias) or {}).get("kind") in ("query", LEDGER_KIND) and alias not in support:
|
||
support.append(alias)
|
||
queries: dict[str, list[str]] = {}
|
||
# 目录里全部查询的编号：合并查询的输入没被直接引用，缺口里也要写得出它是 Q 几
|
||
aliases = {str(e["artifact"]): str(e["alias"]) for e in catalog.values()
|
||
if isinstance(e, dict) and e.get("kind") == "query" and e.get("artifact") and e.get("alias")}
|
||
for alias in dict.fromkeys([*cells, *said, *support]):
|
||
artifact = (catalog.get(alias) or {}).get("artifact")
|
||
if not isinstance(artifact, str) or not artifact:
|
||
continue
|
||
hit = cells.get(alias) or []
|
||
refs = []
|
||
if len(hit) == 1 and isinstance(hit[0][0], int):
|
||
refs.append(f"{alias} 第 {hit[0][0] + 1} 行「{hit[0][1]}」")
|
||
elif hit:
|
||
refs.append(f"{alias} 的 {len(hit)} 格")
|
||
refs.extend(f"{alias}「{key}」" if key else f"{alias}（判定语）" for key in said.get(alias) or [])
|
||
queries.setdefault(artifact, []).extend(refs or [f"{alias}（结论依据）"])
|
||
return queries, aliases
|
||
|
||
|
||
def _declare_citations(
|
||
tier: str, cited: dict[str, Any], *, calibers: list[dict[str, str]], metrics: list[dict[str, Any]],
|
||
gaps: list[str], missing_required: list[str], missing_expected: list[str], report_from: str,
|
||
declared_at: Any, ctx: NodeContext, sql_checks: list[dict[str, Any]] | None = None,
|
||
comparability: list[dict[str, Any]] | None = None,
|
||
) -> dict[str, Any]:
|
||
"""引用模式的 _issuance。旧字段一个不少（老前端照常显示），另加 mode、report、unresolved。"""
|
||
# 节点 id 和报告文档、report.checked 事件上记的同一个（这次运行里的，子工作流里的带前缀）
|
||
report = {"node_id": ctx.run.global_id(report_from), "doc_artifact": cited["doc_artifact"]}
|
||
issuance = {
|
||
"mode": "citations",
|
||
"tier": tier,
|
||
"report": report,
|
||
"calibers": calibers,
|
||
"metrics_checked": len(metrics),
|
||
"missing_required": missing_required,
|
||
"missing_expected": missing_expected,
|
||
"unmatched_numbers": cited["unmatched"][:20],
|
||
"unresolved": cited["unresolved"][:20],
|
||
"matched_numbers": len(cited["matched"]),
|
||
"matched": cited["matched"][:50],
|
||
"stats": cited["stats"],
|
||
"gaps": gaps,
|
||
"declared_at": declared_at.isoformat(),
|
||
# 结论句策略和可疑实体：只在用上时才有这两个键，升级前的运行和没用上的契约形状不变
|
||
**({"claims": cited["claims"]} if cited.get("claims") else {}),
|
||
**({"entities": cited["entities"]} if cited.get("entities") else {}),
|
||
# 没通过 SQL 检查的查询（结构化）：报告页据此在显眼处说明降档原因。没有就不加这个键
|
||
**({"sql_checks": sql_checks} if sql_checks else {}),
|
||
# 分析说法核对（strict，或者有违规时）、指标的可比性、报告引用的指标里追不到原始数据的：同样只在有的时候带
|
||
**({"analysis_claims": cited["analysis"]} if cited.get("analysis") else {}),
|
||
**({"comparability": comparability} if comparability else {}),
|
||
**({"untraced": cited["untraced"]} if cited.get("untraced") else {}),
|
||
}
|
||
entities = cited.get("entities")
|
||
ctx.emit(
|
||
EventType.ISSUANCE,
|
||
mode="citations",
|
||
tier=tier,
|
||
report=report,
|
||
missing_required=missing_required,
|
||
missing_expected=missing_expected,
|
||
unmatched=len(cited["unmatched"]),
|
||
unresolved=len(cited["unresolved"]),
|
||
calibers=calibers,
|
||
gaps=gaps,
|
||
**({"sql_checks": sql_checks} if sql_checks else {}),
|
||
metrics_checked=len(metrics),
|
||
matched_numbers=len(cited["matched"]),
|
||
matched=cited["matched"][:50],
|
||
# on_uncited 只在不是默认的 degrade 时才有（收紧成 withhold、探索运行的契约写了 ignore）；
|
||
# judge 另带不支持时怎么判档、在哪判的和各判定的句数（名单只在出具声明里）
|
||
**({"claims": {k: cited["claims"][k] for k in ("policy", "uncited_claims", "on_uncited", "on_unsupported",
|
||
"mode", "counts")
|
||
if k in cited["claims"]}} if cited.get("claims") else {}),
|
||
# 计数取复核的统计（声明里的名单最多列 _LISTED 个）
|
||
**({"entities": {"unknown": (cited["stats"] or {}).get("unknown_entities", len(entities["unknown"])),
|
||
"unverified": (cited["stats"] or {}).get("unverified_entities", len(entities["unverified"])),
|
||
"counted": entities["counted"]}} if entities else {}),
|
||
# 分析说法：处置、几处、记没记缺口（名单只在出具声明里）
|
||
**({"analysis_claims": {k: cited["analysis"][k] for k in ("mode", "count", "counted")}}
|
||
if cited.get("analysis") else {}),
|
||
)
|
||
return issuance
|
||
|
||
|
||
async def run_transform(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||
"""数据整形。用受限表达式把上游输出揉成下游要的形状，不用为此专门起一个 LLM。"""
|
||
tctx = template_context(state)
|
||
mode = ctx.cfg("mode", "expression")
|
||
output: Any
|
||
|
||
if mode == "expression":
|
||
expr = ctx.cfg("expression", "")
|
||
if not expr:
|
||
raise NodeError(ctx.node.id, "「数据整形」节点还没有填写表达式")
|
||
try:
|
||
output = eval_expression(expr, tctx)
|
||
except ExpressionError as e:
|
||
raise NodeError(ctx.node.id, f"表达式有误：{e}") from e
|
||
elif mode == "template":
|
||
output = ctx.render_str(ctx.cfg("template", ""), state)
|
||
elif mode == "json":
|
||
template = ctx.cfg("template", "{}")
|
||
rendered = ctx.render_str(template, state)
|
||
try:
|
||
output = json.loads(rendered)
|
||
except json.JSONDecodeError as e:
|
||
raise _json_error(e, str(template or ""), tctx, ctx) from e
|
||
else:
|
||
raise NodeError(ctx.node.id, f"不支持的整形方式：{mode}")
|
||
|
||
updates: dict[str, Any] = {"nodes": {ctx.node.id: output}}
|
||
var_name = ctx.cfg("assign_to", "")
|
||
if var_name:
|
||
updates["vars"] = {var_name: output}
|
||
return updates
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 按 JSON 解析失败：错在模板，还是错在上游模型写的文字
|
||
# --------------------------------------------------------------------------
|
||
|
||
#: 产出是模型写的文字的节点（llm 配了 output_schema 时 text 是系统序列化的，但那样也解析得了）
|
||
_MODEL_TEXT = (NodeType.AGENT, NodeType.LLM, NodeType.SUPERVISOR)
|
||
#: 技术细节里给出出错位置前后各多少个字
|
||
_NEAR = 30
|
||
_NODE_HEAD = re.compile(r"^nodes\s*(?:\.\s*([\w-]+)|\[\s*['\"]([^'\"]+)['\"]\s*\])")
|
||
_VAR_HEAD = re.compile(r"^vars\s*(?:\.\s*([\w-]+)|\[\s*['\"]([^'\"]+)['\"]\s*\])")
|
||
|
||
|
||
TEMPLATE_HINT = "请检查模板中的引号和逗号，字符串值需使用 | json 过滤器输出"
|
||
#: 解析器在这一段的头一个字上要的是这些：前一个值在模板里已经写完了，缺的是模板里的标点
|
||
_DELIMITED = ("Expecting ',' delimiter", "Expecting ':' delimiter", "Extra data")
|
||
|
||
|
||
def _json_error(e: json.JSONDecodeError, template: str, tctx: dict[str, Any], ctx: NodeContext) -> NodeError:
|
||
"""整形节点按 JSON 解析失败时的报错。
|
||
|
||
真实踩过：整个模板就是 {{ nodes.X.text }}，agent 在 JSON 字符串里夹了没转义的英文双引号，
|
||
报错却让人去检查模板、加 | json，把人引到了错的方向。反过来也不行：出错的位置落在模型
|
||
写的那一段里，不等于错在模型——模板把一句话原样放进值的位置、放进引号里，或者紧挨着它
|
||
漏了逗号，都是模板的错，而且改法就是那句老提示。所以分三种说（见 _diagnose）：
|
||
|
||
- 上游确实在写 JSON 而没写对：点名上游、按实情说原因、指向结构化输出；
|
||
- 模板少了 | json：点名是模板里哪一段，给出改好的写法；
|
||
- 其余：原来那句提示。
|
||
|
||
出错位置前后的原文只进技术细节（包在原异常里）：首行是给人看的一句话，不贴模型的原文。
|
||
"""
|
||
doc, pos = e.doc, e.pos
|
||
near = (doc[max(0, pos - _NEAR):pos] + "⟨此处⟩" + doc[pos:pos + _NEAR]).replace("\n", "\\n")
|
||
e.args = (f"{e.args[0]}\n出错位置前后的原文：{near}",)
|
||
verdict = _diagnose(template, tctx, ctx, e)
|
||
if verdict is None or isinstance(verdict, str):
|
||
return NodeError(ctx.node.id, f"模板渲染出来的不是合法 JSON（第 {e.lineno} 行第 {e.colno} 列附近）。"
|
||
f"{verdict or TEMPLATE_HINT}")
|
||
node, text, offset, reason = verdict
|
||
line = text.count("\n", 0, offset) + 1
|
||
column = offset - (text.rfind("\n", 0, offset) + 1) + 1
|
||
where = f"第 {line} 行第 {column} 列" if line > 1 else f"第 {column} 列"
|
||
fix = ("如需模型输出结构化数据，请为其配置「结构化输出 Schema」；数字需要进入口径卡时，"
|
||
"请改用 Agent 并开启「按出处核对字段」"
|
||
if node.type == NodeType.LLM else
|
||
"如需模型输出结构化数据，请配置「结构化输出 Schema」并开启「按出处核对字段」，"
|
||
"不要用「数据整形」节点解析模型写的文字")
|
||
return NodeError(ctx.node.id, f"上游「{node.title}」输出的不是合法 JSON（{where}附近），{reason}；{fix}")
|
||
|
||
|
||
def _diagnose(template: str, tctx: dict[str, Any], ctx: NodeContext,
|
||
e: json.JSONDecodeError) -> tuple[GraphNode, str, int, str] | str | None:
|
||
"""按 JSON 解析失败，错在谁。
|
||
|
||
返回 (上游节点, 它写的那段文字, 出错位置在文字里的偏移, 原因) 表示错在上游；返回一句话
|
||
表示错在模板、而且说得出是哪一段；返回 None 表示照原来那句提示。
|
||
|
||
只有上游确实在写 JSON（去掉开头空白后以 { 或 [ 开头，或者包在 ``` 代码块里）、而且这一段
|
||
不在模板的引号里，才算上游的错；整个模板就是这一段时，模板不可能写错，也算上游的。
|
||
解析器停在这一段末尾时（紧跟着的逗号漏了），只有它自己开了括号、引号没收尾才怪它。
|
||
|
||
每个 {{ }} 单独渲染一次再拼起来，和 render_template 的整段替换结果逐字相同。
|
||
"""
|
||
from app.engine.expressions import _TEMPLATE_RE, render_template, resolve_path
|
||
|
||
doc, pos = e.doc, e.pos
|
||
at, last = 0, 0
|
||
for m in _TEMPLATE_RE.finditer(template):
|
||
at += m.start() - last
|
||
piece = render_template(m.group(0), tctx)
|
||
start, end = at, at + len(piece)
|
||
at, last = end, m.end()
|
||
expr = m.group(1).strip()
|
||
head, *filters = [p.strip() for p in expr.split("|")]
|
||
if not head or pos < start:
|
||
continue
|
||
# 解析器越过空白才报错：停在这一段后面、中间只隔着空白的，也是停在它末尾
|
||
at_end = pos >= end and not doc[end:pos].strip()
|
||
if not piece.strip():
|
||
# 取出来是空的，解析器正好在这里要一个值
|
||
if e.msg == "Expecting value" and not doc[start:pos].strip() and not _in_json_string(doc, start):
|
||
return f"模板里 {{{{ {expr} }}}} 的取值为空，此处缺少一个值：请检查路径是否正确、前面的节点是否有产出"
|
||
continue
|
||
if pos > end and not at_end:
|
||
continue
|
||
quoted = _in_json_string(doc, start)
|
||
if "json" in filters:
|
||
# 过了 | json 就是一个合法的 JSON 值：错不在它的内容里，除非外面又套了一层引号
|
||
return (f"模板里 {{{{ {expr} }}}} 的结果已是合法的 JSON（文字自带引号），外面不要再加引号"
|
||
if quoted and not at_end else None)
|
||
value = resolve_path(tctx, head)
|
||
if not isinstance(value, str) or value != piece:
|
||
return None
|
||
node = _producer(head, ctx)
|
||
model = node is not None and node.type in _MODEL_TEXT
|
||
lead = start + len(piece) - len(piece.lstrip())
|
||
if pos == lead and e.msg.startswith(_DELIMITED):
|
||
return None # 前一个值在模板里就写完了，缺的是它前面的标点
|
||
if template.strip() == m.group(0):
|
||
return (node, piece, pos - start, _why(piece, at_end)) if model else None
|
||
if at_end and not _unclosed(piece):
|
||
return None # 这一段本身完整，缺的是它后面的标点
|
||
if quoted:
|
||
return _needs_json(expr, model)
|
||
if _writes_json(piece):
|
||
return (node, piece, pos - start, _why(piece, at_end)) if model else None
|
||
return _needs_json(expr, model)
|
||
return None
|
||
|
||
|
||
def _needs_json(expr: str, model: bool) -> str:
|
||
what = "模型写的文字" if model else "一段文字"
|
||
return f"模板里 {{{{ {expr} }}}} 是{what}，放入 JSON 时请写成 {{{{ {expr} | json }}}}（外面不要再加引号）"
|
||
|
||
|
||
def _why(text: str, at_end: bool) -> str:
|
||
"""上游写的 JSON 为什么解析不了，按看得出来的实情说。"""
|
||
body = text.lstrip()
|
||
if body.startswith("```"):
|
||
return "模型把 JSON 包在了 ``` 代码块中"
|
||
if at_end and _unclosed(text):
|
||
return "模型输出的 JSON 不完整，可能在中途被截断"
|
||
if not body.startswith(("{", "[")):
|
||
return "模型输出的是一段文字，而不是 JSON"
|
||
return "常见原因是字符串中有未转义的英文引号"
|
||
|
||
|
||
def _writes_json(text: str) -> bool:
|
||
"""这段文字看得出是想写 JSON：以 { 或 [ 开头，或者是包在 ``` 代码块里的 JSON。"""
|
||
body = text.lstrip()
|
||
if body.startswith("```"):
|
||
body = body.split("\n", 1)[1].lstrip() if "\n" in body else ""
|
||
return body.startswith(("{", "["))
|
||
|
||
|
||
def _in_json_string(doc: str, index: int) -> bool:
|
||
"""doc[:index] 结束时是不是在一个 JSON 字符串里。解析器已经越过了这里，这一段前缀是
|
||
合法的 JSON 开头，数引号（跳过转义）就数得准。"""
|
||
inside, i = False, 0
|
||
while i < index:
|
||
ch = doc[i]
|
||
if inside and ch == "\\":
|
||
i += 2
|
||
continue
|
||
if ch == '"':
|
||
inside = not inside
|
||
i += 1
|
||
return inside
|
||
|
||
|
||
def _unclosed(text: str) -> bool:
|
||
"""这段文字自己开了引号、括号却没收尾（单独看它，从字符串外面数起）。"""
|
||
depth, inside, i = 0, False, 0
|
||
while i < len(text):
|
||
ch = text[i]
|
||
if inside:
|
||
if ch == "\\":
|
||
i += 1
|
||
elif ch == '"':
|
||
inside = False
|
||
elif ch == '"':
|
||
inside = True
|
||
elif ch in "{[":
|
||
depth += 1
|
||
elif ch in "}]":
|
||
depth -= 1
|
||
i += 1
|
||
return inside or depth > 0
|
||
|
||
|
||
def _producer(path: str, ctx: NodeContext) -> GraphNode | None:
|
||
"""模板路径取的是哪个节点的产出：nodes.X 就是 X；vars.Y 是 assign_to 为 Y 的节点。"""
|
||
nodes = ctx.run.spec.node_map()
|
||
if m := _NODE_HEAD.match(path):
|
||
return nodes.get(m.group(1) or m.group(2))
|
||
if m := _VAR_HEAD.match(path):
|
||
var = m.group(1) or m.group(2)
|
||
writers = [n for n in ctx.run.spec.nodes if str(n.config.get("assign_to") or "").strip() == var]
|
||
# 好几个节点写同一个变量时说不准是谁，宁可给原来的提示也不点错名
|
||
return writers[0] if len(writers) == 1 else None
|
||
return None
|