mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-18 09:01:59 +00:00
合入期 5 验收修复 M1:核查 agent 正文里的依据标记不再断句——只装着依据的括号(含【】[]〔〕[])连同里面的停顿标点一起去掉,去掉之后按真正挨上的两个字补空格,两个数不再并成一个;核查 agent 的提示词改成依据写在句末;收尾超时照实记 timeout_s;没给理由时照实说没有给出理由
This commit is contained in:
@@ -54,6 +54,7 @@ import asyncio
|
||||
import inspect
|
||||
import json
|
||||
import logging
|
||||
import re
|
||||
import time
|
||||
from collections.abc import Callable, Iterable
|
||||
from dataclasses import asdict, dataclass, field, replace
|
||||
@@ -101,6 +102,9 @@ _REF_DEPTH = 8
|
||||
RATIONALE = "rationale"
|
||||
#: 触顶原因
|
||||
LIMITS = ("max_queries", "timeout_s", "max_cost_usd", "daily_max_usd", "max_steps")
|
||||
#: 收尾那次调用超过时长上限、没交成结果时的原话(limit 记 timeout_s)。是内部的说法,给人看时调用方换成自己的话:质询的
|
||||
#: _error_words、反方审稿的「反查未完成:已达上限(…)」、双盲复算按 limit 写的无法复算原因
|
||||
SETTLE_TIMEOUT = "收尾时超过时长上限,没能交结果"
|
||||
#: 核查模型的输出到了上限被截断、没交成结果时给人看的话:双盲复算的无法复算、反方审稿的「反查未完成」、质询的报错都照
|
||||
#: 它说,也是 verify_truncated 日志的原话。cut 是 budget.cut_words 的「核查模型的输出达到上限(N token)被截断」。开头只说
|
||||
#: 「未能给出结果」:质询里提问人看到的就是这句,它问的是问题,不是在核查
|
||||
@@ -125,7 +129,10 @@ VERIFY_FINAL_PROMPT = (
|
||||
"row 是行号(从 0 数),column 是列名。系统按引用从查询结果里取值,你自己写的数字不会被采用;查不到就写 null,"
|
||||
"不要估算\n"
|
||||
"- rationale 写理由,简短说清依据。理由里出现的每个数字都写成 [[v:Q1.r0.列名]] 这样的引用,系统会换成查询结果"
|
||||
"里的值;不要直接写数字,提到某次查询时写 [[see:Q1]]\n"
|
||||
"里的值;不要直接写数字\n"
|
||||
"- 要指明一句话依据的是哪次查询,在这句的句末写 [[see:Q1]],比如「……结果一致。[[see:Q3]]」。[[see:…]] 渲染后"
|
||||
"不显示,只能挂在句末:不要放进括号,也不要写进句子中间、写在「依据」「见」「取自」这类字后面,否则会留下「()」"
|
||||
"「(,」「依据。」这样的断句\n"
|
||||
"- 其余字段按各自的说明填写"
|
||||
)
|
||||
_CELL_REF: dict[str, Any] = {
|
||||
@@ -687,6 +694,26 @@ def _parsed(value: Any) -> dict[str, Any] | None:
|
||||
return value if isinstance(value, dict) else None
|
||||
|
||||
|
||||
#: 只是占位词的理由(只留字母、数字和汉字,不分大小写):模型照着结构填了个字面的 placeholder、N/A,等于没给理由
|
||||
_PLACEHOLDERS = frozenset({"placeholder", "todo", "tbd", "na", "null", "nil", "rationale", "reason", "略", "理由"})
|
||||
#: 说「没有」的词:理由是另一个结果的说明时(双盲复算的数、反方审稿的结论),它等于没给理由;理由就是结果本身时(审批
|
||||
#: 质询,结构里只有理由一项)它可能就是回答——问「有没有测试账号」,答「无。」
|
||||
_NOTHING = frozenset({"none", "无", "暂无"})
|
||||
#: 依据写在引出语后面(「依据 [[see:Q1]]。」「结果见:[[see:Q1]]」「(来源:[[see:Q1]])」):依据渲染后不显示,引出语悬在
|
||||
#: 那里(「依据。」),渲染时去不掉(evidence._see_cuts 只收拾依据周围的空白、括号和标点)。交结果的提示要求写在句末,
|
||||
#: 模型不照做时记一条日志,验收时查得到
|
||||
_LEAD_IN_SEE = re.compile(r"(?:依据|见|取自|来源|出处)[ \t\u3000]*[::]?[ \t\u3000]*\[\[[ \t]*see[ \t]*:[^\[\]\n]*\]\]")
|
||||
|
||||
|
||||
def _no_rationale(shown: str, *, answer: bool = False) -> bool:
|
||||
"""渲染出来的理由等于没给:空的、只有标点(只写了依据:[[see:Q1]]、「([[see:Q1]])」渲染出来是空的),或者只是一个
|
||||
占位词(「placeholder。[[see:Q1]]」渲染出来只剩占位词)。按渲染后的正文判,不按原文:标记里的字(see、Q1)不显示。
|
||||
answer:理由就是交回的结果本身(审批质询的回答),说「没有」的词不算。当没给理由处理(理由为空、没有理由文档),
|
||||
审批卡照「没有给出理由」说,不把占位词当理由展示,判「被推翻」的也不当有出处。"""
|
||||
word = "".join(ch for ch in shown if ch.isalnum()).casefold()
|
||||
return not word or word in _PLACEHOLDERS or (not answer and word in _NOTHING)
|
||||
|
||||
|
||||
@dataclass
|
||||
class _State:
|
||||
"""一次核查过程中攒下的东西:出意外时也照这些交回,用了多少、查了什么不丢。"""
|
||||
@@ -1032,10 +1059,14 @@ async def _verify(state: _State, *, spec: ModelSpec, system: str, task: str, sou
|
||||
if parsed is None:
|
||||
if settle_cut:
|
||||
return await report_cut("verify_result", settler.max_tokens)
|
||||
why = "收尾时超过时长上限,没能交结果" if isinstance(failure, asyncio.TimeoutError) else \
|
||||
timed_out = isinstance(failure, asyncio.TimeoutError)
|
||||
why = SETTLE_TIMEOUT if timed_out else \
|
||||
f"核查模型没有按结构交结果({explain_model_error(failure, model_id)})" if failure is not None else \
|
||||
"核查模型没有按结构交结果"
|
||||
return outcome(limit=stopped, answered=answered, error=why)
|
||||
# 收尾超时:没交成结果是因为时长上限,limit 照实记 timeout_s,和原因对得上——模型自己停了查询(stopped 为 None)
|
||||
# 时以前记成没触顶,质询的报错说时长上限、界面却没有「已达时长上限」;循环里先查满了次数的,记成查询次数上限
|
||||
# 会让双盲复算、反方审稿说错原因(查满了照样能交结果)。上面金额被拒的收尾仍报循环里先撞上的那个
|
||||
return outcome(limit="timeout_s" if timed_out else stopped, answered=answered, error=why)
|
||||
declared = list((resolved.get("properties") or {}) if _is_object(resolved) else {})
|
||||
if declared and not any(k in parsed for k in declared):
|
||||
# 一个要求的字段都没交:不是「查不到」,是没按结构交。不这么分,字段名对不上的回复会被当成什么都没查到
|
||||
@@ -1050,6 +1081,13 @@ async def _verify(state: _State, *, spec: ModelSpec, system: str, task: str, sou
|
||||
doc = None
|
||||
if isinstance(rationale, str) and rationale.strip():
|
||||
doc = compose_doc(rationale, build_catalog(nodes={}, ledger=ledger), node_id=owner, run_id=run_id or None)
|
||||
if hits := _LEAD_IN_SEE.findall(rationale):
|
||||
logger.warning("核查 agent 的理由里依据写在了引出语后面,渲染后引出语会悬空(运行 %s,节点 %s):%s",
|
||||
run_id or "-", owner, "、".join(hits[:5]))
|
||||
# 按渲染出来的判:只写了依据的、占位词后面挂着依据的,原文里有字,显示出来没有理由。结构里只有理由一项(审批
|
||||
# 质询)时理由就是回答,「无」是一个回答
|
||||
if _no_rationale(doc["markdown"], answer=declared == [RATIONALE]):
|
||||
doc = None
|
||||
result[RATIONALE] = doc["markdown"] if doc else ""
|
||||
return outcome(result=result, cells=cells.cells, rationale=doc, limit=stopped, answered=answered)
|
||||
|
||||
@@ -1130,4 +1168,5 @@ async def _run_tool(tool: Any, name: str, args: dict[str, Any], call_id: str, re
|
||||
return content
|
||||
|
||||
|
||||
__all__ = ["LIMITS", "RATIONALE", "SETTLE_GRACE_S", "Default", "VerifyOutcome", "result_schema", "run_verify_agent"]
|
||||
__all__ = ["LIMITS", "RATIONALE", "SETTLE_GRACE_S", "SETTLE_TIMEOUT", "Default", "VerifyOutcome", "result_schema",
|
||||
"run_verify_agent"]
|
||||
|
||||
Reference in New Issue
Block a user