mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 05:21:57 +00:00
ThearchyHelios
9727c21d8d
一、结论句裁判
- 报告节点新增 claims: judge：写完、核对完之后，把结论句和它引用的证据摘录交给裁判模型，
判为支持 / 部分支持 / 不支持 / 不是结论，附一句理由；判定只标注，不改写正文
- 送裁判的句子按固定规则预筛：带引用的、含方向词或因果词的送去判，短句、过渡句不送
- 证据摘录：指标写名称、值和原式；查询给 SQL 前 300 字和被引用的行，被引用的列一定给到；遮罩列一律不给；
规则里写明「数字已经由系统核对过，不要复算」
- 一份报告合并成一次调用，超出句数就分批；结构化输出失败时退回 JSON 解析，再失败记为未裁判
- 裁判模型的优先级：节点上写的 > 设置里的「证据裁判模型」> Copilot 的模型；思考关闭；
和写作模型相同时给警告（等于自己审自己）
- rewrite_once（默认关）：把不支持的句子和理由交回写作者只改这几句，再裁一次；
改写稿冒出原稿没有的问题就不采用，保留原稿和原来的判定
二、成本上限
- 每份报告的句数、金额、时长，探索运行每次点击的金额，全局每日金额，每一项都可以自定义，也都可以设成不限；
默认 40 句、$0.05、30 秒，每次点击 $0.01，每天 $2
- 调用前先估算，超出上限按优先级截断：先判带引用的，再判含方向词、因果词的，其余记为未裁判并写明「已到上限」；
不报错，也不让报告失败
- 每日花费按本地日期计数，持久化在设置表里；重跑同一次运行不重复计费；价格目录里没有的模型只计次数，并写明估不出金额
三、正式运行与探索运行
- 正式运行在报告节点内裁判，判定写进报告文档、随文档封存；不支持的按 on_unsupported（degrade / withhold）降档或不予出具，
裁判没跑成记缺口，不能判完整出具
- 探索运行按需裁判：POST /api/runs/{id}/evidence/judge，点开哪句判哪句；判定作为 evidence.judged 事件追加在封存之后，
标「封存后追加」，封存核对仍然一致；只读封存范围内的工件；同一句不重复判；运行在裁判期间被接着跑时不记录
- 正式运行调这个接口返回 409，写明「正式运行的裁判在节点内完成」
四、治理与自动修复
- G3 放开 judge：受管级别写 judge 必须写预算（judge.max_cost_usd，或显式写 null 表示不限）
- 发布前自动修复：缺预算时给选项（默认金额 / 不限），不替人选；
从 require_citation 改成 judge 不算降低要求，judge 改回 off、withhold 放宽成 degrade 算降低要求
- Copilot 手册：受管模板可以写 judge 并带预算；问数据按需裁判，不替人打开结论句策略
五、界面
- 结论句句末放徽标：支持 ◆、部分支持 ◇、不支持 ! 整句浅底、未裁判 ?；不用表示「已核对」的绿色；n / N 也跳到不支持的句子
- 证据面板里「模型的解释」：带「模型判断 · 模型名 · 非确定」，封存后追加的另外标出；
「请模型判断这句」只在探索运行、还没判过时出现，触顶时写明已到上限和怎么调
- 出具横幅多一段结论句计数；设置页新增「证据裁判」一组（裁判模型、各项上限、今日花费），设成不限时写明费用受什么约束；
报告节点可以配 judge；运行轨迹里裁判相关的事件都有中文标题
六、门禁修正
- 引用模式下，图里没有口径卡（只引查询单元格）的，出具契约不再要求 metrics_from，受管门禁也不再要求 required：
strict 加 cells 已经保证每个数都点得开、对不上就不予出具；有口径卡的照旧要求。
以前问数据的图怎么都发不成受管，Copilot 补上契约也会被判「冒出新的问题」作废
778 lines
40 KiB
Python
778 lines
40 KiB
Python
from __future__ import annotations
|
||
|
||
import json
|
||
import re
|
||
from typing import Any
|
||
|
||
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)
|
||
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,
|
||
) -> dict[str, Any]:
|
||
from datetime import datetime, timezone
|
||
|
||
from app.engine.issuance import decide_tier, 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"指标源未解析：{', '.join(unresolved[:5])}")
|
||
|
||
# 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] = {}
|
||
if citations:
|
||
cited = _check_citations(str(report_from), contract, state, ctx, reports or [], cells=cells,
|
||
governed=governed)
|
||
gaps.extend(cited["gaps"])
|
||
unmatched = cited["unmatched"]
|
||
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（路径可能写错了）")
|
||
elif not declared_narrative.strip():
|
||
gaps.append("契约没有声明叙述，数字回指未执行")
|
||
|
||
trace = (
|
||
trace_numbers(narrative, metrics, allow=contract.get("allow_numbers"))
|
||
if narrative.strip()
|
||
else None
|
||
)
|
||
unmatched = trace.unmatched if trace else []
|
||
|
||
# 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，叙述里的数字无从回指")
|
||
|
||
# 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」
|
||
titles = {n.id: n.title for n in ctx.run.spec.nodes}
|
||
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，交来的是成员最后的原话")
|
||
|
||
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)
|
||
|
||
# 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,
|
||
"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(),
|
||
}
|
||
ctx.emit(
|
||
EventType.ISSUANCE,
|
||
tier=tier,
|
||
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,
|
||
)
|
||
return issuance
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 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.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}
|
||
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"契约的 report_from 指向的「{report_from}」不是报告撰写节点，引用核对未执行")
|
||
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（没跑到、被跳过或者失败了），"
|
||
"引用核对未执行")
|
||
return out
|
||
out["doc_artifact"] = doc_artifact
|
||
try:
|
||
doc = load(doc_artifact)
|
||
except ValueError:
|
||
gaps.append(f"报告「{node.title}」的文档和它的哈希对不上，疑似被改过，引用核对未执行")
|
||
return out
|
||
if not isinstance(doc, dict):
|
||
gaps.append(f"报告「{node.title}」的文档在工件库里取不回来，引用核对未执行")
|
||
return out
|
||
if doc.get("node_id") != 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}」的文档和这次运行里它的产出对不上，引用核对未执行")
|
||
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": []}
|
||
for v in checked["violations"]:
|
||
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}」的文档没通过复核：{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}」里有 {len(unknown)} 处可疑实体（{names}{more}）：本次运行的表结构、查询、"
|
||
"结果列里都没有这些名字，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)}
|
||
|
||
# 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）
|
||
hit["cell"] = f"{cite['alias']}.r{cite['locator']['row']}.{cite['locator']['column']}"
|
||
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)
|
||
|
||
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"成果里没有哪个字段是报告「{node.title}」的原文：读者看到的内容没有经过引用核对")
|
||
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")
|
||
#: 证据不支持的结论句怎么处置（claims: judge），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")
|
||
|
||
|
||
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）。正式运行在节点里判过：证据不支持的按 on_unsupported 判档（报告节点和契约取更严的），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「不支持」。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，契约写了也照旧记「这一版还不支持」。
|
||
"""
|
||
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("契约声明了结论句检查（claims），这一版还不支持，结论句没有核对")
|
||
judged = upgraded and written == "judge"
|
||
if upgraded and name == "judge" and not judged:
|
||
out["gaps"].append(f"契约要求结论句由模型裁判（claims: judge），报告「{title}」没有开 claims: judge，"
|
||
"结论句没有裁判")
|
||
if not asks:
|
||
return
|
||
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
|
||
|
||
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": ["开了 claims: judge，文档里却没有裁判结果，结论句没有裁判"]}
|
||
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]]] = {"unsupported": [], "partial": []}
|
||
if inline:
|
||
# 探索运行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），这里只数正式运行的
|
||
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 ""})
|
||
out["gaps"].extend(f"报告「{title}」{gap}" for gap in summary.get("gaps") or [])
|
||
counts = {k: int((summary.get("counts") or {}).get(k) or 0)
|
||
for k in ("supported", "partial", "unsupported", "not_a_claim", "unjudged")}
|
||
out["claims_policy"] = {"policy": "judge", "on_unsupported": on_unsupported, "on_uncited": on_uncited}
|
||
out["uncited"], out["unsupported"] = uncited, flagged["unsupported"]
|
||
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,
|
||
"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"),
|
||
}
|
||
|
||
|
||
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,
|
||
) -> 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。"""
|
||
report = {"node_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 {}),
|
||
}
|
||
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,
|
||
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 {}),
|
||
)
|
||
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, "整形节点没有填表达式")
|
||
try:
|
||
output = eval_expression(expr, tctx)
|
||
except ExpressionError as e:
|
||
raise NodeError(ctx.node.id, f"表达式错误：{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"未知的整形模式：{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 = "检查模板里的引号、逗号，字符串值要用 | 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 = ("让模型交结构化数据请给它配 output_schema；数要进口径卡的，改用 agent 的 output_schema + cite_fields"
|
||
if node.type == NodeType.LLM else
|
||
"让模型交结构化数据请用 output_schema + cite_fields，别用整形节点解析它写的文字")
|
||
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} }}}} 取出来是空的，这里缺一个值：检查路径写对没有、前面的节点有没有产出"
|
||
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} }}}} 出来的已经是合法的 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}，放进 JSON 要写成 {{{{ {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 "它把 JSON 包在了 ``` 代码块里"
|
||
if at_end and _unclosed(text):
|
||
return "它写的 JSON 没有收尾，可能写到一半被截断了"
|
||
if not body.startswith(("{", "[")):
|
||
return "它写的是一段文字，不是 JSON"
|
||
return "常见原因是字符串里有没转义的英文引号"
|
||
|
||
|
||
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
|