mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 02:11:58 +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:
ThearchyHelios
+249
-27
@@ -1811,8 +1811,43 @@ _UNIT_KINDS = frozenset({"m", "i", "v"})
|
||||
#: 渲染结果末尾的单位：最后一个数字后面的字（12,345人次 → 人次，4.57万元 → 万元，12.5% → %）
|
||||
_UNIT_TAIL = re.compile(r"\d([^\d\s][^\d]*)$")
|
||||
_BLANK = " \t\u3000"
|
||||
_SEE_RUN = re.compile(r"\[\[" + _MARKER_GUARD + r"[ \t]*see[ \t]*:[^\[\]\n]*?\]\]"
|
||||
r"(?:[" + _BLANK + r"]*\[\[" + _MARKER_GUARD + r"[ \t]*see[ \t]*:[^\[\]\n]*?\]\])*")
|
||||
#: 依据标记 [[see:…]]（同 MARKER_RE，只认 MARKER_MAX 以内的）
|
||||
_SEE_MARK = r"\[\[" + _MARKER_GUARD + r"[ \t]*see[ \t]*:[^\[\]\n]*?\]\]"
|
||||
#: 括号，每类一对：圆括号、方括号（全角、半角，模型偶尔一对里混用，也算一对）、方头括号、六角括号。核查 agent 拿到的
|
||||
#: 查询结果都以「【证据 Qn】」开头，模型照着写成「【[[see:Q1]]】」并不意外
|
||||
_PAIRS = (("（(", "）)"), ("［[", "］]"), ("【", "】"), ("〔", "〕"))
|
||||
_OPEN, _CLOSE = frozenset("".join(o for o, _ in _PAIRS)), frozenset("".join(c for _, c in _PAIRS))
|
||||
#: 依据和括号、依据和依据之间的分隔标点（「（[[see:Q3]]，原报告…）」「[[see:Q1]]、[[see:Q2]]」）
|
||||
_SEPS = "，、；,;：:"
|
||||
_SEP = frozenset(_SEPS)
|
||||
#: 其中的停顿标点（不含冒号）：依据挂在它后面、紧接着就到句末时，它悬在句末（「华南居首，。」）。冒号后面另起一行
|
||||
#: 常常是它引出的表格、列表，不动
|
||||
_PAUSE = _SEP - frozenset("：:")
|
||||
#: 句末标点：停顿标点悬在它前面时去掉（还有全文结尾）。半角句点不算：后面跟着数字时它是小数点。换行不算：列表项末尾
|
||||
#: 的「；」、软换行处连着两个分句的「，」是写作者自己的，依据挂在它后面正合「挂在句末」的规矩
|
||||
_ENDING = frozenset("。！？!?")
|
||||
#: 依据周围、渲染后会挤在一起的字：空白和分隔标点
|
||||
_GAP = frozenset(_BLANK) | _SEP
|
||||
#: 依据渲染成空、周围的字去掉之后，挨上的两个字会读成一串（_glue）：左边是数字、字母，或者这几个——标记的结尾（渲染
|
||||
#: 出一个数）、百分号、小数点、半角标点（「212,5」读成一个数；半角标点按排版习惯后面本来就隔一个空格）；右边是数字、
|
||||
#: 字母、一个渲染出字的标记，或者紧贴着这几样的半角逗号。中文和数字之间不加空格也读得对，不算
|
||||
_STICKY = frozenset("]%‰.,;:")
|
||||
#: 渲染出来什么都没有的东西（依据标记、只装着依据的括号）在遮盖文本里换成这个字，位置和原文一一对应（_voids）
|
||||
_VOID = "\x00"
|
||||
#: 只装着依据的括号（_PAIRS 里同一类的一对）：里面除了依据（和已经认出来的、套在里面的这种括号）只有空白、分隔标点。
|
||||
#: 头一个依据前面只认空白、标点，免得一长串依据后面没有右括号时来回试
|
||||
_VOID_BRACKET = re.compile("|".join(
|
||||
"[" + re.escape(o) + "][" + _BLANK + _SEPS + "]*" + _VOID + "[" + _BLANK + _SEPS + _VOID + "]*"
|
||||
"[" + re.escape(c) + "]" for o, c in _PAIRS))
|
||||
#: 连着写的几个：中间只隔着空白、分隔标点
|
||||
_VOID_GROUP = re.compile(_VOID + "+(?:[" + _BLANK + _SEPS + "]*" + _VOID + "+)*")
|
||||
#: 其中只隔着空白的一段（两边都是字时，中间的分隔标点可能就是分句的标点，按它断开各算各的）
|
||||
_VOID_RUN = re.compile(_VOID + "+(?:[" + _BLANK + "]*" + _VOID + "+)*")
|
||||
#: 遮盖文本里看得见的一段：一组里夹着的空白、标点，括号连同里面的空白、标点
|
||||
_SOLID = re.compile("[^" + _VOID + "]+")
|
||||
#: 粗体、斜体的星号，粗体的两个下划线，删除线的两个波浪线：前端按 Markdown 渲染，不显示（组装文档时包着标记的粗体
|
||||
#: 也去掉了星号）。删除线画出来的字照样挨着后面的字读
|
||||
_EMPHASIS = re.compile(r"\*+|__+|~~+")
|
||||
|
||||
|
||||
def _rendered_unit(marker: dict[str, Any], catalog: dict[str, Any], snaps: _Snapshots,
|
||||
@@ -1830,8 +1865,9 @@ def _tidy(text: str, catalog: dict[str, Any], snaps: _Snapshots, cells_allowed:
|
||||
- 数字标记后面重复写的单位去掉。[[m:visits]] 渲染出来已经带单位（12,345人次），写作者常常
|
||||
还跟着写一遍（[[m:visits]] 人次），读者看到的是「12,345人次 人次」。单字的单位后面紧跟着字时
|
||||
不动：写作者写的可能是另一个单位（口径卡是「人」，正文写「人次」），硬删会把不一致藏起来；
|
||||
- [[see:…]] 前面的空白去掉。see 渲染成空，「最高月份 [[see:Q2]]。」会剩下「最高月份 。」。
|
||||
see 后面紧跟着字、要靠这个空白隔开的（「A [[see:Q1]]B」）留着；行首的缩进不动。
|
||||
- [[see:…]] 渲染成空，它周围留下的空白、括号、标点去掉（_voids、_see_cuts）：「最高月份 [[see:Q2]]。」会剩下
|
||||
「最高月份 。」，「结果一致（[[see:Q3]]，原报告…）」会剩下「结果一致（，原报告…）」。依据（连同去掉的字）两边
|
||||
挨上的字会读成一串的，补一个空格（_glue）：「4,919.5[[see:Q1]]212 笔」会读成「4,919.5212 笔」。
|
||||
|
||||
prev 是这段文字前面那个字（流式渲染分块时用），判断空白是不是在行首。
|
||||
"""
|
||||
@@ -1847,29 +1883,210 @@ def _tidy(text: str, catalog: dict[str, Any], snaps: _Snapshots, cells_allowed:
|
||||
after = text[hit.end():hit.end() + 1]
|
||||
if len(unit) > 1 or not after.isalnum():
|
||||
cuts.append((marker["end"], hit.end()))
|
||||
for run in _SEE_RUN.finditer(text):
|
||||
start = run.start()
|
||||
while start > 0 and text[start - 1] in _BLANK:
|
||||
start -= 1
|
||||
before = text[start - 1] if start > 0 else prev
|
||||
after = text[run.end():run.end() + 1]
|
||||
# 几个 see 之间的空白也一样渲染成多余的空格
|
||||
parts = list(MARKER_RE.finditer(text, run.start(), run.end()))
|
||||
cuts.extend((m.end(), n.start()) for m, n in zip(parts, parts[1:]) if m.end() < n.start())
|
||||
if start < run.start() and before not in ("", "\n") and not (after.isalnum() or after == "["):
|
||||
cuts.append((start, run.start()))
|
||||
if not cuts:
|
||||
see: list[tuple[int, int]] = []
|
||||
spaces: set[int] = set()
|
||||
if "see" in text: # 没有依据标记就没有要收拾的（流式渲染的每一块都走这里，先挡掉）
|
||||
masked, see = _voids(text)
|
||||
for group in _VOID_GROUP.finditer(masked):
|
||||
see += _see_cuts(text, masked, group.start(), group.end(), prev)
|
||||
spaces = _glue(text, masked, see, prev)
|
||||
if not cuts and not see and not spaces:
|
||||
return text
|
||||
# 几处要去掉的字可能重叠、相接（依据前面的空白、它和括号之间的标点）：按并集去掉，要补空格的那一片换成一个空格；
|
||||
# 补空格的地方没有要去掉的字（依据两边本来就什么都没有、紧贴着下文的半角逗号后面），在那里插一个
|
||||
merged: list[list[int]] = []
|
||||
for a, b in sorted(cuts + see):
|
||||
if merged and a <= merged[-1][1]:
|
||||
merged[-1][1] = max(merged[-1][1], b)
|
||||
else:
|
||||
merged.append([a, b])
|
||||
edits = [(a, b, " " if any(a <= x < b for x in spaces) else "") for a, b in merged]
|
||||
edits += [(x, x, " ") for x in spaces if not any(a <= x < b for a, b in merged)]
|
||||
out, last = [], 0
|
||||
for a, b in sorted(cuts):
|
||||
if a < last:
|
||||
continue
|
||||
out.append(text[last:a])
|
||||
for a, b, put in sorted(edits):
|
||||
out += [text[last:a], put]
|
||||
last = b
|
||||
out.append(text[last:])
|
||||
return "".join(out)
|
||||
|
||||
|
||||
def _voids(text: str) -> tuple[str, list[tuple[int, int]]]:
|
||||
"""渲染出来什么都没有的东西：依据标记，和只装着依据的括号（里面可以隔着空白、分隔标点，套着的也算）。返回遮盖文本
|
||||
（这些地方换成 _VOID，位置和原文一一对应）和这种括号连同里面的空白、标点：要一起去掉的字（「合计 6,635.68
|
||||
（[[see:Q1]]）。」不剩「（）」）。依据标记本身留着：组装文档时句子靠它挂依据（unit.see），渲染时它本来就不显示。
|
||||
|
||||
认出来的括号和依据一样当一个不显示的东西：连着写的几个（「（[[see:Q1]]）（[[see:m:gmv]]）」「[[see:Q1]]
|
||||
（[[see:Q2]]）」）由 _VOID_GROUP 并成一组，外面照 _see_cuts 收拾一遍（「华南居首，（[[see:Q1]]）。」的逗号）。"""
|
||||
chars = list(text.replace(_VOID, "\x01"))
|
||||
for m in MARKER_RE.finditer(text):
|
||||
if m.group("kind") == "see":
|
||||
chars[m.start():m.end()] = _VOID * (m.end() - m.start())
|
||||
masked, cuts = "".join(chars), []
|
||||
while hits := list(_VOID_BRACKET.finditer(masked)):
|
||||
# 一轮认出最里面的一层：套在外面的括号下一轮才只剩依据
|
||||
for hit in hits:
|
||||
cuts += [(s.start(), s.end()) for s in _SOLID.finditer(masked, hit.start(), hit.end())]
|
||||
chars[hit.start():hit.end()] = _VOID * (hit.end() - hit.start())
|
||||
masked = "".join(chars)
|
||||
return masked, cuts
|
||||
|
||||
|
||||
def _see_cuts(text: str, masked: str, start: int, end: int, prev: str) -> list[tuple[int, int]]:
|
||||
"""一组不显示的东西（_VOID_GROUP：依据、只装着依据的括号，连着写的几个，masked[start:end]）渲染成空之后，周围要
|
||||
一起去掉的字：
|
||||
|
||||
- 在括号开头（「（[[see:Q3]]，原报告…）」）：它和括号之间、它后面的标点和空白；在括号末尾（「（同 SQL，
|
||||
[[see:Q2]]）」）同理
|
||||
- 后面就到句末（句末标点、全文结尾），前面、中间有停顿标点的（「华南居首，[[see:Q1]]。」「华南居首，
|
||||
（[[see:Q1]]）。」）：这些停顿标点（前面是冒号的留着：它引出的可能是下一行的表格、列表）；没有停顿标点、到句末
|
||||
只隔着空白的，空白去掉（「华南居首 [[see:Q1]] 。」）
|
||||
- 以上几种里，中间的空白、分隔标点一并去掉（「（同 SQL；[[see:Q1]]、[[see:Q2]]）」）
|
||||
- 两边都是字的，中间的分隔标点可能就是分句的标点：按它断开，每段照 _run_cuts
|
||||
|
||||
不含依据的括号、标点一个字不动。要看的只有这组前后连着的空白、分隔标点，外加再往外的一个字：流式渲染把这些字攒到
|
||||
后面的字来了才放（_TRAILING_SEE），分块渲染和整篇渲染判得一样。prev 是这段文字前面那个字。
|
||||
"""
|
||||
gaps = [(s.start(), s.end()) for s in _SOLID.finditer(masked, start, end)]
|
||||
lo, hi = start, end
|
||||
while lo > 0 and text[lo - 1] in _GAP:
|
||||
lo -= 1
|
||||
while hi < len(text) and text[hi] in _GAP:
|
||||
hi += 1
|
||||
opener = text[lo - 1] if lo > 0 else prev
|
||||
closer = text[hi] if hi < len(text) else ""
|
||||
left, right = text[lo:start], text[end:hi]
|
||||
if opener in _OPEN:
|
||||
return [(lo, start), *gaps, (end, hi)]
|
||||
if closer in _CLOSE:
|
||||
# 和括号之间只有空白的，照依据前的空白处理（行首的缩进不动）
|
||||
return [*gaps, (end, hi), *([(lo, start)] if _SEP & set(left) else _blank_cut(text, start, end, prev))]
|
||||
tail: list[tuple[int, int]] = []
|
||||
if closer in _ENDING or closer == "":
|
||||
if _PAUSE & set(left + "".join(text[a:b] for a, b in gaps)):
|
||||
# 依据后面、句末前面的停顿标点是写作者自己的（「增长 [[see:Q1]]，」），不是依据留下的，不动；只剩空白的去掉
|
||||
return [*gaps, *([] if _SEP & set(right) else [(end, hi)]),
|
||||
*([(lo, start)] if _PAUSE & set(left) else _blank_cut(text, start, end, prev))]
|
||||
if not _SEP & set(right):
|
||||
tail = [(end, hi)]
|
||||
return [*tail, *(cut for run in _VOID_RUN.finditer(masked, start, end)
|
||||
for cut in _run_cuts(text, masked, run.start(), run.end(), prev))]
|
||||
|
||||
|
||||
def _run_cuts(text: str, masked: str, start: int, end: int, prev: str) -> list[tuple[int, int]]:
|
||||
"""两边都是字的一段（_VOID_RUN，中间只隔着空白）：
|
||||
|
||||
- 中间的空白，和前面的空白（_blank_cut）
|
||||
- 夹在两个分隔标点中间（「结果一致，[[see:Q1]]，原报告…」）：去掉后面那个（连同它前面的空白，后面的空白留着
|
||||
隔开下文：「a, [[see:Q1]], b」→「a, b」）。去掉之后前一个贴着下文、读成一串的（「212,[[see:Q1]],5」），_glue
|
||||
补回一个空格
|
||||
"""
|
||||
cuts = [(s.start(), s.end()) for s in _SOLID.finditer(masked, start, end)]
|
||||
cuts += _blank_cut(text, start, end, prev)
|
||||
lo, hi = start, end
|
||||
while lo > 0 and text[lo - 1] in _GAP:
|
||||
lo -= 1
|
||||
while hi < len(text) and text[hi] in _GAP:
|
||||
hi += 1
|
||||
left, right = text[lo:start], text[end:hi]
|
||||
if _SEP & set(left) and _SEP & set(right):
|
||||
cuts.append((end, end + max(i for i, ch in enumerate(right) if ch in _SEP) + 1))
|
||||
return cuts
|
||||
|
||||
|
||||
def _blank_cut(text: str, start: int, end: int, prev: str) -> list[tuple[int, int]]:
|
||||
"""text[start:end]（依据，或者只装着依据的一对括号）渲染成空，它前面的空白去掉：「最高月份 [[see:Q2]]。」会剩下
|
||||
「最高月份 。」。后面紧跟着字、要靠这个空白隔开的（「A [[see:Q1]]B」）留着；行首的缩进不动。"""
|
||||
blank = start
|
||||
while blank > 0 and text[blank - 1] in _BLANK:
|
||||
blank -= 1
|
||||
before = text[blank - 1] if blank > 0 else prev
|
||||
after = text[end:end + 1]
|
||||
if blank < start and before not in ("", "\n") and not (after.isalnum() or after == "["):
|
||||
return [(blank, start)]
|
||||
return []
|
||||
|
||||
|
||||
def _glue(text: str, masked: str, see: list[tuple[int, int]], prev: str) -> set[int]:
|
||||
"""依据渲染成空、它周围的字（see）去掉之后，两边真正挨上的两个看得见的字会读成一串的（_STICKY）：返回要补空格的
|
||||
位置，_tidy 在那里补一个（位置上的字是要去掉的就换成空格，不是就插在它前面）。
|
||||
「[[v:Q1.r0.销售额]]（[[see:Q1]]）2 月」去掉括号读成「4,919.52 月」，「[[v:Q1.r0.销售额]][[see:Q1]][[v:Q1.r0.订单数]]
|
||||
笔」读成「4,919.5212 笔」，是另一个数，而且哪项核对都查不出来。
|
||||
|
||||
每个依据（连着写的一串、只装着依据的括号都算）两边都判，不只看去掉了字的地方：依据前后什么都没有时，两边一样
|
||||
直接挨上。往两边找挨上的字时越过一切不显示的：去掉的字、依据标记、粗体斜体的星号和下划线、删除线的波浪线。
|
||||
两个括号连着写（「（[[see:Q1]]）（[[see:m:gmv]]）」）、数字包着粗体（「**[[v:…]]**[[see:Q1]]2 月」）都照挨上
|
||||
的字判；依据标记的「]」越过去，不当数字（「华南居首。[[see:Q1]]（[[see:Q2]]）2 月」不补）。一片里只补一个
|
||||
（「212,[[see:Q1]],[[see:m:gmv]],5」→「212, 5」）；挨上的是写作者留着的空白的，不补。右边挨上的是紧贴着下文的
|
||||
半角逗号的（「1,200（[[see:Q1]]）,900」读成「1,200,900」），空格补在逗号后面，和上面那种读起来一样。
|
||||
|
||||
空格补在这一片头一个去掉的字那里；依据两边本来什么都没有、没有去掉的字的，补在头一个依据前面。粗体收尾的星号
|
||||
在它左边，粗体照样成对。粗体把依据、括号包在里面时（「**[[v:…]]（[[see:Q1]]）**2 月」）它落在收尾的星号前面：
|
||||
前端的行内 Markdown 照样认这对星号，组装文档时包着标记的星号也去掉了，读者看到的一样。要挪到星号外面得先知道
|
||||
这对星号从哪开始，流式渲染的一块里不一定有。
|
||||
|
||||
越过的这些字（空白、括号、分隔标点、依据、星号、下划线、波浪线）流式渲染都攒到后面的字来了才放（_TRAILING_SEE）：
|
||||
一片总落在同一块里，往左越出这一块就是 prev（它不会是这些字），分块渲染和整篇渲染判得一样。"""
|
||||
cut = bytearray(len(text))
|
||||
for a, b in see:
|
||||
cut[a:b] = b"\x01" * (b - a)
|
||||
gone = bytearray(cut)
|
||||
for void in re.finditer(_VOID + "+", masked):
|
||||
gone[void.start():void.end()] = b"\x01" * (void.end() - void.start())
|
||||
|
||||
spaces: set[int] = set()
|
||||
seen: set[tuple[int, int]] = set()
|
||||
for run in re.finditer(rb"\x01+", gone):
|
||||
p, q = run.start(), run.end()
|
||||
while p > 0:
|
||||
if gone[p - 1]:
|
||||
p -= 1
|
||||
elif text[p - 1] == "*" or (p > 1 and text[p - 2:p] in ("__", "~~")):
|
||||
mark = text[p - 1]
|
||||
while p > 0 and text[p - 1] == mark:
|
||||
p -= 1
|
||||
else:
|
||||
break
|
||||
while q < len(text):
|
||||
if gone[q]:
|
||||
q += 1
|
||||
elif emphasis := _EMPHASIS.match(text, q):
|
||||
q = emphasis.end()
|
||||
else:
|
||||
break
|
||||
if (p, q) in seen:
|
||||
continue
|
||||
seen.add((p, q))
|
||||
left = text[p - 1] if p > 0 else prev
|
||||
if not (_wordish(left) or left in _STICKY):
|
||||
continue
|
||||
if _shows_word(text, q):
|
||||
first = cut.find(b"\x01", run.start(), q)
|
||||
spaces.add(first if first != -1 else run.start())
|
||||
elif text[q:q + 1] == ",":
|
||||
# 右边是写作者紧贴着下文写的半角逗号（「1,200（[[see:Q1]]）,900」）：去掉之后读成「1,200,900」，空格照半角
|
||||
# 逗号的排版习惯补在它后面（同「212,[[see:Q1]],5」→「212, 5」）。下文包着粗体的越过星号看。逗号、星号都是
|
||||
# _TRAILING_SEE 攒着的字，再往后那个字和它们落在同一块里
|
||||
after = q + 1
|
||||
while emphasis := _EMPHASIS.match(text, after):
|
||||
after = emphasis.end()
|
||||
if _shows_word(text, after):
|
||||
spaces.add(q + 1)
|
||||
return spaces
|
||||
|
||||
|
||||
def _shows_word(text: str, i: int) -> bool:
|
||||
"""text[i] 起头的是数字、字母，或者一个渲染出字的标记（依据不算）：挨在一个数后面会读成一串。"""
|
||||
if text[i:i + 1] == "[":
|
||||
marker = MARKER_RE.match(text, i)
|
||||
return marker is not None and marker.group("kind") != "see"
|
||||
return _wordish(text[i:i + 1])
|
||||
|
||||
|
||||
def _wordish(ch: str) -> bool:
|
||||
"""数字（全角的也算）或者半角字母：两个挨在一起读成一个数、一个词。"""
|
||||
return ch.isdecimal() or (ch.isascii() and ch.isalnum())
|
||||
|
||||
|
||||
def _render(text: str, catalog: dict[str, Any], snaps: _Snapshots, cells_allowed: bool, prev: str = "") -> str:
|
||||
text = _tidy(_expand_tables(text, catalog, snaps, cells_allowed, prev), catalog, snaps, cells_allowed, prev)
|
||||
out: list[str] = []
|
||||
@@ -1893,8 +2110,11 @@ def render_markers(text: str, catalog: dict[str, Any], *, cells_allowed: bool =
|
||||
# 流式渲染：让 llm.token 里出现的就是最终的数字
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
#: 文字末尾的空白和 [[see:…]]：流式渲染先攒着，等后面的字来了再一起交给 _tidy
|
||||
_TRAILING_SEE = re.compile(r"(?:[" + _BLANK + r"]|\[\[" + _MARKER_GUARD + r"[ \t]*see[ \t]*:[^\[\]\n]*?\]\])+$")
|
||||
#: 文字末尾的空白、括号、分隔标点、星号、下划线、波浪线和 [[see:…]]：流式渲染先攒着，等后面的字来了再一起交给 _tidy
|
||||
#: （依据周围的这些字去不去掉、补不补空格，要看再往后的那个字，见 _see_cuts、_glue）。半角方括号也在里面：标记收尾的
|
||||
#: 「]]」不算，_safe_cut 另外挑出来
|
||||
_TRAILING_SEE = re.compile(r"(?:[" + _BLANK + re.escape("".join(sorted(_OPEN | _CLOSE))) + _SEPS + "*_~" + r"]|"
|
||||
+ _SEE_MARK + r")+$")
|
||||
|
||||
|
||||
def _safe_cut(text: str, has_unit: Callable[[dict[str, Any]], bool] | None = None) -> int:
|
||||
@@ -1905,9 +2125,9 @@ def _safe_cut(text: str, has_unit: Callable[[dict[str, Any]], bool] | None = Non
|
||||
开头可能是另一个 [。
|
||||
|
||||
整表标记闭合了也先不放：它展开时要看紧跟在后面的那个字（是不是换行），那个字还没到。
|
||||
_tidy 要往后看的也先不放：末尾的空白和 [[see:…]]（后面来的字决定空白去不去掉），
|
||||
以及后面还没攒够 UNIT_LOOKAHEAD 个字的数字标记（后面可能是重复的单位）。has_unit 判断一个
|
||||
标记渲染出来带不带单位，不带的（运行输入里的周次、没有单位的格）不用攒；不给就都攒。
|
||||
_tidy 要往后看的也先不放：末尾的空白、括号、分隔标点、星号、下划线、波浪线和 [[see:…]]（后面来的字决定它们
|
||||
去不去掉、要不要补空格），以及后面还没攒够 UNIT_LOOKAHEAD 个字的数字标记（后面可能是重复的单位）。has_unit 判断
|
||||
一个标记渲染出来带不带单位，不带的（运行输入里的周次、没有单位的格）不用攒；不给就都攒。
|
||||
"""
|
||||
cut = len(text)
|
||||
start = text.rfind("[[")
|
||||
@@ -1924,7 +2144,10 @@ def _safe_cut(text: str, has_unit: Callable[[dict[str, Any]], bool] | None = Non
|
||||
if m and m.end() == held and m.group("kind") == "table":
|
||||
held = start
|
||||
if (trail := _TRAILING_SEE.search(text, 0, held)) and trail.start() < held:
|
||||
held = trail.start()
|
||||
# 攒的一段从一个标记中间开始（标记收尾的「]]」、标记里末尾的空白也是那几种字）：标记整个放出去，从它后面攒起
|
||||
s = text.rfind("[[", 0, trail.start())
|
||||
inside = MARKER_RE.match(text, s, held) if s != -1 else None
|
||||
held = inside.end() if inside and inside.end() > trail.start() else trail.start()
|
||||
numeric = [m for m in MARKER_RE.finditer(text, 0, held) if m.group("kind") in _UNIT_KINDS
|
||||
and (has_unit is None or has_unit(parse_markers(m.group(0))[0]))]
|
||||
if numeric and len(text[numeric[-1].end():held]) < UNIT_LOOKAHEAD \
|
||||
@@ -3055,7 +3278,6 @@ SEE_AND_CODE_RULES = (
|
||||
#: 句中的 [[see:]]：前面是介词、后面接着句子的写法（「以 [[see:Q1]] 为依据」「根据 [[see:Q1]]，」「详见 [[see:Q1]]」）。
|
||||
#: 只认这几种拿得准的：「以」前面是「所、可、加……」的是另一个词（所以、可以），不算；「依据 [[see:…]]。」可能是
|
||||
#: 「这是结论的依据」，不算。认出来就交回写作者重写（report.REWRITE_ONLY：重写不好也不判失败）
|
||||
_SEE_MARK = r"\[\[" + _MARKER_GUARD + r"[ \t]*see[ \t]*:[^\[\]\n]*?\]\]"
|
||||
_INLINE_SEE = re.compile(
|
||||
r"(?<![所可加予得难足用是何之])以[ \t\u3000]*" + _SEE_MARK + r"[ \t\u3000]*为"
|
||||
r"|(?:根据|按照|基于|参见|详见)[ \t\u3000]*" + _SEE_MARK
|
||||
|
||||
@@ -125,7 +125,7 @@ CHALLENGE_SYSTEM = (
|
||||
"原因说过了头；refuted 查到的数据和这句结论相矛盾；inconclusive 查不到判断所需的数据\n"
|
||||
"- 没有找到反例不等于结论被推翻：只有数据和结论冲突时才判 refuted\n"
|
||||
"- 结论要有出处：一条查询都没执行成功时，判什么都不会被采信；判 refuted 时，理由里至少要引用一处你自己查询的"
|
||||
"结果（[[v:Qn.rX.列名]] 或 [[see:Qn]]），否则同样不会被采信"
|
||||
"结果：写 [[v:Qn.rX.列名]] 引用一格，或者在那句的句末写 [[see:Qn]]，否则同样不会被采信"
|
||||
)
|
||||
CHALLENGE_TASK_PROMPT = (
|
||||
"要审的结论（报告「{report}」中的一句{section}）：\n「{claim}」\n\n"
|
||||
@@ -771,9 +771,17 @@ class _Review:
|
||||
message=f"反方模型与报告的写作模型相同（「{outcome.model}」），难以发现写作模型自身的错误。"
|
||||
f"请在设置的「核查」一节选择另一个模型，或在节点的「{CHALLENGE_LABEL['model']}」中指定")
|
||||
claim, why = _claim_of(cand, outcome)
|
||||
# 触顶的计在哪一组：时长按剩余时间收紧过的另算一组（_SQUEEZED），说运行时长上限将至
|
||||
group = _SQUEEZED if outcome.limit == "timeout_s" and allot.tightened else outcome.limit
|
||||
if outcome.limit == "daily_max_usd" and outcome.result is None and not outcome.usage.get("calls"):
|
||||
# 派出去就撞上每日金额上限、一次模型都没调用：和后面不再派出的句子是同一回事，同一种说法
|
||||
claim = _claim(cand, "inconclusive", self._capped(), queries=outcome.queries, usage=outcome.usage)
|
||||
elif outcome.result is None and outcome.error == verify_agent.SETTLE_TIMEOUT:
|
||||
# 查完了、交结论那次超过时长上限：照上限说，和时间线上的 challenge_limit 一致。核查 agent 的原话（「收尾时
|
||||
# 超过时长上限，没能交结果」）是内部的说法，不摆上审批卡
|
||||
words = self._limit_words(_SQUEEZED if allot.tightened else "timeout_s")
|
||||
claim = _claim(cand, "inconclusive", f"反查未完成：已达上限（{words}）", queries=outcome.queries,
|
||||
usage=outcome.usage)
|
||||
elif lost and why is None:
|
||||
# 反查了，可该查的源有停用了的：判断照常采信，理由末尾写明哪几个没查（只写在任务里的话，审稿记录和
|
||||
# 审批卡上的「经受住」看着就像每个源都查过了）
|
||||
@@ -782,8 +790,7 @@ class _Review:
|
||||
claims.append(claim)
|
||||
usage = _add_usage(usage, outcome.usage)
|
||||
if outcome.limit in _CLAIM_LIMITS:
|
||||
squeezed = outcome.limit == "timeout_s" and allot.tightened
|
||||
counted = hit.setdefault(_SQUEEZED if squeezed else outcome.limit, [0, 0])
|
||||
counted = hit.setdefault(group, [0, 0])
|
||||
counted[0] += 1
|
||||
counted[1] += why is not None
|
||||
broke = outcome.limit == "daily_max_usd"
|
||||
|
||||
@@ -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"]
|
||||
|
||||
@@ -0,0 +1,115 @@
|
||||
"""反方审稿的理由：依据写进括号不再断句，占位词不当理由（期 5 真实模型验收 M1）。
|
||||
|
||||
真实模型验收里，反方的理由「先按订单状态做全量概览核对口径（[[see:Q1]]）：」「复算写作者的原查询（同 SQL 同过滤条件，
|
||||
[[see:Q2]]）」渲染成「（）：」「（同 SQL 同过滤条件，）」：审稿记录（rationale 和 rationale_doc）随正式运行封存，审批卡上
|
||||
照片段原文显示。
|
||||
|
||||
模型一律是 MockChatModel 剧本，区域、金额都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from app.core import artifact_store
|
||||
from tests.test_challenge_node import ( # noqa: F401 - 夹具按名字引入
|
||||
NET,
|
||||
REFUNDS,
|
||||
Models,
|
||||
Play,
|
||||
engine_up,
|
||||
finish,
|
||||
graph,
|
||||
output_of,
|
||||
shop,
|
||||
)
|
||||
|
||||
BROKEN = ("（）", "（，", "，）", "()", "(,", ",)")
|
||||
|
||||
|
||||
async def _record(monkeypatch, shop, plays) -> tuple[dict, dict]:
|
||||
Models(monkeypatch, shop, plays=plays)
|
||||
row = await finish(graph(shop))
|
||||
assert row.status == "succeeded", row.error
|
||||
out = await output_of(row.id, "chal")
|
||||
return artifact_store.load(out["challenge_artifact"]), row.output["_issuance"]
|
||||
|
||||
|
||||
async def test_see_in_brackets_does_not_break_the_rationale(monkeypatch, shop):
|
||||
plays = {
|
||||
"华南销售额超过华东": Play([NET], "refuted", "扣除退款后华南为 [[v:Q1.r1.净销售额]]（[[see:Q1]]），"
|
||||
"仍低于华东的 [[v:Q1.r0.净销售额]]（同口径，[[see:Q1]]）。"),
|
||||
"居各区域之首": Play([REFUNDS], "weakened", "先核对退款口径（[[see:Q1]]）：华东含 [[v:Q1.r0.退款笔数]] 笔退款"
|
||||
"（[[see:Q1]]，扣除后仍居首）。"),
|
||||
}
|
||||
record, issuance = await _record(monkeypatch, shop, plays)
|
||||
first, second = record["claims"]
|
||||
assert first["rationale"] == "先核对退款口径：华东含 1 笔退款（扣除后仍居首）。"
|
||||
assert second["rationale"] == "扣除退款后华南为 900，仍低于华东的 1,200（同口径）。"
|
||||
for claim in (first, second):
|
||||
doc = claim["rationale_doc"]
|
||||
assert doc["markdown"] == claim["rationale"] and doc["violations"] == []
|
||||
assert not any(mark in claim["rationale"] for mark in BROKEN)
|
||||
# 括号里的依据仍挂在句子上：判「被推翻」的引用到了自己的查询，照常采信、出具降档
|
||||
assert second["verdict"] == "refuted" and issuance["tier"] == "degraded"
|
||||
|
||||
|
||||
async def test_see_brackets_written_together_do_not_glue_two_numbers(monkeypatch, shop):
|
||||
"""两个只装着依据的括号连着写、数字包着粗体（复核第 2 轮）：去掉之后两个数不能挨成一个（「9001,200」），句末前悬着的
|
||||
逗号也去掉。"""
|
||||
plays = {
|
||||
"华南销售额超过华东": Play([NET], "refuted", "扣除退款后华南、华东分别为 **[[v:Q1.r1.净销售额]]**（[[see:Q1]]）"
|
||||
"（[[see:Q1]]）[[v:Q1.r0.净销售额]]，华南仍低于华东，（[[see:Q1]]）（[[see:Q1]]）。"),
|
||||
"居各区域之首": Play([REFUNDS], "stands", "华东退款 [[v:Q1.r0.退款笔数]] 笔，扣除后仍居首。[[see:Q1]]"),
|
||||
}
|
||||
record, _ = await _record(monkeypatch, shop, plays)
|
||||
refuted = next(claim for claim in record["claims"] if claim["verdict"] == "refuted")
|
||||
assert refuted["rationale"] == "扣除退款后华南、华东分别为 900 1,200，华南仍低于华东。"
|
||||
assert refuted["rationale_doc"]["markdown"] == refuted["rationale"]
|
||||
assert refuted["rationale_doc"]["violations"] == []
|
||||
|
||||
|
||||
async def test_a_bare_see_between_two_numbers_and_other_brackets_do_not_break_the_rationale(monkeypatch, shop):
|
||||
"""依据直接夹在两个数之间（两边什么都没有）、写进方头括号（复核第 3 轮）：两个数不能挨成一个（「9001,200」），括号
|
||||
不能剩下「【】」。"""
|
||||
plays = {
|
||||
"华南销售额超过华东": Play([NET], "refuted", "扣除退款后华南、华东分别为 [[v:Q1.r1.净销售额]][[see:Q1]]"
|
||||
"[[v:Q1.r0.净销售额]]，华南仍低于华东【[[see:Q1]]】。"),
|
||||
"居各区域之首": Play([REFUNDS], "stands", "华东退款 [[v:Q1.r0.退款笔数]] 笔，扣除后仍居首。[[see:Q1]]"),
|
||||
}
|
||||
record, _ = await _record(monkeypatch, shop, plays)
|
||||
refuted = next(claim for claim in record["claims"] if claim["verdict"] == "refuted")
|
||||
assert refuted["rationale"] == "扣除退款后华南、华东分别为 900 1,200，华南仍低于华东。"
|
||||
assert refuted["rationale_doc"]["markdown"] == refuted["rationale"]
|
||||
assert refuted["rationale_doc"]["violations"] == []
|
||||
|
||||
|
||||
async def test_a_placeholder_rationale_is_recorded_as_no_rationale(monkeypatch, shop):
|
||||
"""反方交回的理由只是占位词：记录里理由为空（审批卡据此写「反方没有给出理由」），判「被推翻」的照常不予采信。"""
|
||||
plays = {"华南销售额超过华东": Play([NET], "refuted", "placeholder"),
|
||||
"居各区域之首": Play([REFUNDS], "stands", "placeholder")}
|
||||
record, issuance = await _record(monkeypatch, shop, plays)
|
||||
first, second = record["claims"]
|
||||
assert (first["verdict"], first["rationale"], first["rationale_doc"]) == ("stands", "", None)
|
||||
assert second["verdict"] == "inconclusive"
|
||||
assert second["rationale"] == "理由没有引用到任何查询结果，反方给出的「被推翻」不予采信，记为无法判断。反方没有给出理由"
|
||||
assert "placeholder" not in str(record)
|
||||
assert issuance["tier"] == "formal"
|
||||
|
||||
|
||||
async def test_a_rationale_that_shows_nothing_is_recorded_as_no_rationale(monkeypatch, shop):
|
||||
"""反方的理由只写了依据：显示出来是空的。以前按原文判，理由文档照样记下——审批卡按文档画出一片空白，不说「反方没有
|
||||
给出理由」；判「被推翻」的因为引用到了 Q1 被当成有出处，照常采信、出具降档。"""
|
||||
plays = {"华南销售额超过华东": Play([NET], "refuted", "[[see:Q1]]"),
|
||||
"居各区域之首": Play([REFUNDS], "stands", "（[[see:Q1]]）")}
|
||||
record, issuance = await _record(monkeypatch, shop, plays)
|
||||
first, second = record["claims"]
|
||||
assert (first["verdict"], first["rationale"], first["rationale_doc"]) == ("stands", "", None)
|
||||
assert second["verdict"] == "inconclusive" and second["rationale_doc"] is None
|
||||
assert second["rationale"].endswith("反方给出的「被推翻」不予采信，记为无法判断。反方没有给出理由")
|
||||
assert issuance["tier"] == "formal", "没有理由的「被推翻」不能让出具降档"
|
||||
|
||||
|
||||
async def test_a_placeholder_with_a_see_is_still_a_placeholder(monkeypatch, shop):
|
||||
plays = {"华南销售额超过华东": Play([NET], "stands", "placeholder。[[see:Q1]]"),
|
||||
"居各区域之首": Play([REFUNDS], "stands", "无。")}
|
||||
record, _ = await _record(monkeypatch, shop, plays)
|
||||
assert [(c["rationale"], c["rationale_doc"]) for c in record["claims"]] == [("", None), ("", None)]
|
||||
assert "placeholder" not in str(record)
|
||||
@@ -0,0 +1,70 @@
|
||||
"""双盲复算的说明：依据写进括号不再断句，占位词不当说明（期 5 真实模型验收 M1）。
|
||||
|
||||
盲算 agent 和反方审稿、审批质询共用核查 agent 交结果的提示，同样会把 [[see:Q1]] 写进括号；真实模型验收里还见过盲算交回
|
||||
字面的「placeholder」当说明，审批卡上照着显示。
|
||||
|
||||
模型一律是 MockChatModel 剧本，订单库是虚构的（tests/fixtures/dual_check_shop.py）。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
from app.core import artifact_store
|
||||
from tests.fixtures.dual_check_shop import BY_ORDER_DATE, RIGHT, Blind
|
||||
from tests.test_dual_check_node import ( # noqa: F401 - 夹具按名字引入
|
||||
clean_settings,
|
||||
engine_up,
|
||||
graph,
|
||||
pending,
|
||||
shop,
|
||||
start,
|
||||
)
|
||||
|
||||
|
||||
async def _blind_input(monkeypatch, shop, rationale: str) -> dict:
|
||||
name, _ = shop
|
||||
Blind(monkeypatch, name, {"销售额": (RIGHT, "gmv")}, rationale=rationale)
|
||||
run = await start(graph(name, BY_ORDER_DATE))
|
||||
assert run.status == "interrupted", run.error
|
||||
card = artifact_store.load((await pending(run.id)).payload["card_artifact"])
|
||||
[mine] = card["metrics"][0]["blind"]["inputs"]
|
||||
return mine
|
||||
|
||||
|
||||
async def test_see_in_brackets_does_not_break_the_blind_rationale(monkeypatch, shop):
|
||||
mine = await _blind_input(monkeypatch, shop, "按口径说明统计（[[see:Q1]]，已排除退款），结果是 {ref}（[[see:Q1]]）。")
|
||||
assert mine["rationale"] == "按口径说明统计（已排除退款），结果是 330。"
|
||||
doc = mine["rationale_doc"]
|
||||
assert doc["markdown"] == mine["rationale"] and doc["violations"] == []
|
||||
|
||||
|
||||
async def test_see_brackets_written_together_do_not_glue_the_number(monkeypatch, shop):
|
||||
"""两个只装着依据的括号连着写、数字包着粗体（复核第 2 轮）：去掉之后数字不能和后面的字挨成一个数（「3302 月」），
|
||||
句末前悬着的逗号也去掉。"""
|
||||
mine = await _blind_input(monkeypatch, shop, "结果是 **{ref}**（[[see:Q1]]）（[[see:Q1]]）2 月起按支付日统计，"
|
||||
"已排除退款，（[[see:Q1]]）（[[see:Q1]]）。")
|
||||
assert mine["rationale"] == "结果是 330 2 月起按支付日统计，已排除退款。"
|
||||
doc = mine["rationale_doc"]
|
||||
assert doc["markdown"] == mine["rationale"] and doc["violations"] == []
|
||||
|
||||
|
||||
async def test_a_bare_see_before_a_number_and_other_brackets_do_not_break_the_rationale(monkeypatch, shop):
|
||||
"""依据直接夹在数字和后面的数之间（两边什么都没有）、写进六角括号（复核第 3 轮）：数字不能和后面的字挨成一个数
|
||||
（「3302 月」），括号不能剩下「〔〕」。"""
|
||||
mine = await _blind_input(monkeypatch, shop, "结果是 {ref}[[see:Q1]]2 月起按支付日统计，已排除退款〔[[see:Q1]]〕。")
|
||||
assert mine["rationale"] == "结果是 330 2 月起按支付日统计，已排除退款。"
|
||||
doc = mine["rationale_doc"]
|
||||
assert doc["markdown"] == mine["rationale"] and doc["violations"] == []
|
||||
|
||||
|
||||
async def test_a_placeholder_rationale_is_not_shown_as_the_explanation(monkeypatch, shop):
|
||||
mine = await _blind_input(monkeypatch, shop, "placeholder")
|
||||
assert mine["value"] == 330.0, "数照样采用"
|
||||
assert (mine["rationale"], mine["rationale_doc"]) == ("", None)
|
||||
|
||||
|
||||
async def test_a_rationale_that_shows_nothing_is_not_kept(monkeypatch, shop):
|
||||
"""只写了依据（显示出来是空的），或者占位词后面挂着依据：说明为空、不留文档。以前文档照样记下，审批卡按文档画出
|
||||
「理由」两个字，后面什么都没有。"""
|
||||
for said in ("（[[see:Q1]]）", "placeholder。[[see:Q1]]"):
|
||||
mine = await _blind_input(monkeypatch, shop, said)
|
||||
assert mine["value"] == 330.0, "数照样采用"
|
||||
assert (mine["rationale"], mine["rationale_doc"]) == ("", None), said
|
||||
@@ -0,0 +1,600 @@
|
||||
"""依据标记 [[see:…]] 写进括号、夹在标点中间时的渲染（期 5 真实模型验收 M1）。
|
||||
|
||||
[[see:]] 渲染成空，按设计挂在句末。核查 agent（双盲复算、反方审稿、审批质询共用）以前被提示「提到某次查询时写
|
||||
[[see:Q1]]」，模型就把它写进括号、句中：「结果一致（[[see:Q3]]，原报告 SQL 未按 status 过滤）」「合计
|
||||
[[v:Q1.r1.金额合计]]（[[see:Q1]]）」，渲染后剩下「（，原报告…）」「（）」，正式运行里随封存永久留着。
|
||||
|
||||
- 括号（圆括号、方括号的全角半角，一对里混用也算；方头括号【】、六角括号〔〕）里只有依据的，连括号一起去掉；去掉之后
|
||||
整对括号当一个依据，外面照下面几条再收拾一遍（括号前悬着的停顿标点、括号前后的空白）。连着写的几个
|
||||
（「（[[see:Q1]]）（[[see:m:gmv]]）」）并成一组收拾
|
||||
- 依据渲染成空、周围的字去掉之后，两边真正挨上的两个字会粘成一串的（「4,919.5（[[see:Q1]]）2 月」，依据两边什么都
|
||||
没有的「4,919.5[[see:Q1]]212 笔」也是），补一个空格：粘起来读成另一个数，哪项核对都查不出来。每个依据两边都判，
|
||||
找挨上的字时越过一切不显示的：依据、只装着依据的括号、粗体斜体的星号和下划线、删除线的波浪线；依据标记的「]」
|
||||
不当数字。一片里最多补一个，挨上的是写作者留着的空白的不补；半角逗号夹着依据、去掉后一个之后同理
|
||||
（「212,[[see:Q1]],5」→「212, 5」），右边是紧贴着下文的半角逗号的（「1,200（[[see:Q1]]）,900」）补在逗号后面
|
||||
- 依据在括号开头、后面跟着「，」「、」「；」「：」的，标点一起去掉；在括号末尾、前面紧挨着标点的同理
|
||||
- 依据夹在两个分隔标点中间的，去掉后一个；后面就到句末（句末标点、全文结尾）、前面或几个依据之间有「，」「、」
|
||||
「；」的，去掉这些标点，到句末只隔着空白的去掉空白；依据后面、句末前面的标点是写作者自己的，不动
|
||||
- 换行不算句末：列表项末尾的「；」、软换行处的「，」是写作者自己的，依据挂在它后面正合「挂在句末」的规矩，照旧留着
|
||||
- 括号里、句末前几个依据之间的分隔标点一并去掉；两边都是字时，依据之间的标点可能是分句的标点，照旧留着
|
||||
- 依据标记本身留在原文里：句子照旧挂着这条依据（see、cites），只是不显示
|
||||
- 不含依据的括号、没有依据的文档一个字不动：内容哈希和改动之前一样
|
||||
- 流式渲染分块输出拼起来仍然等于整篇渲染
|
||||
- 已知限制：依据写在「依据」「见」「取自」「来源」这类引出语后面时，引出语悬空（「依据。」「（来源）」）。渲染时去不掉，
|
||||
靠核查 agent 的提示词避免，模型不照做时核查 agent 记一条日志（test_verify_rationale_rules）
|
||||
|
||||
数字、表名、地名都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import random
|
||||
import re
|
||||
|
||||
import pytest
|
||||
|
||||
from app.core.artifact_store import canonical_json, content_hash
|
||||
from app.engine.evidence import (
|
||||
StreamRenderer,
|
||||
build_catalog,
|
||||
compose_doc,
|
||||
iter_units,
|
||||
render_markers,
|
||||
verify_doc,
|
||||
)
|
||||
|
||||
ART = "a" * 64
|
||||
SNAP_ID = "b" * 64
|
||||
#: 一次查询的快照：华南 4,919.5 元 212 笔、华东 1,891.8 元 113 笔
|
||||
SNAP = {"columns": ["区域", "销售额", "订单数"], "rows": [["华南", 4919.5, 212], ["华东", 1891.8, 113]],
|
||||
"row_count": 2, "truncated": False, "elapsed_ms": 1, "sql": "SELECT 1", "source": "shop"}
|
||||
CARD = {"kind": "metric_set", "caliber": "周报口径", "caliber_version": "v1", "artifact": ART,
|
||||
"metrics": [{"id": "gmv", "name": "销售额", "unit": "元", "value": 45678.5, "decimals": 1,
|
||||
"format": "thousands", "status": "ok", "expression": "…"},
|
||||
{"id": "yoy", "name": "同比增幅", "unit": "%", "value": 12.5, "decimals": 1,
|
||||
"format": "plain", "status": "ok", "expression": "…"}]}
|
||||
LEDGER = [{"kind": "metric_set", "node_id": "cal", "exec": 1, "artifact": ART, "caliber": "周报口径",
|
||||
"version": "v1", "metrics": ["gmv", "yoy"]},
|
||||
{"kind": "query", "node_id": "fetch", "exec": 1, "artifact": SNAP_ID, "call_id": "c1",
|
||||
"tool": "db_query__shop", "source": "shop", "columns": SNAP["columns"], "rows": 2, "truncated": False}]
|
||||
LOADER = {SNAP_ID: SNAP}.__getitem__
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def catalog():
|
||||
return build_catalog(nodes={"cal": CARD}, ledger=LEDGER, inputs={"week": "2026-W37"})
|
||||
|
||||
|
||||
def render(raw: str, catalog) -> str:
|
||||
return render_markers(raw, catalog, loader=LOADER)
|
||||
|
||||
|
||||
def compose(raw: str, catalog) -> dict:
|
||||
return compose_doc(raw, catalog, node_id="write", run_id="r1", loader=LOADER)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 括号里只有依据：连括号一起去掉
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
("先按状态核对口径（[[see:Q1]]）：已付 [[v:Q1.r0.订单数]] 笔。", "先按状态核对口径：已付 212 笔。"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）。", "合计 4,919.5。"),
|
||||
("换口径不改变排序([[see:Q1]])，结论成立。", "换口径不改变排序，结论成立。"),
|
||||
("结果一致（[[see:Q1]])。", "结果一致。"), # 一对括号全角半角混用
|
||||
("结果一致（[[see:Q1]] [[see:m:gmv]]）。", "结果一致。"),
|
||||
("结果一致（[[see:Q1]]、[[see:m:gmv]]）。", "结果一致。"),
|
||||
("结果一致（ [[see:Q1]]， ）。", "结果一致。"),
|
||||
("结果一致 （[[see:Q1]]）。", "结果一致。"), # 括号前的空白一起去掉
|
||||
("前一句 （[[see:Q1]]）后一句", "前一句 后一句"), # 后面紧跟着字：留一个空格隔开
|
||||
("末尾（[[see:Q1]]）", "末尾"),
|
||||
])
|
||||
def test_a_bracket_holding_only_see_is_dropped(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
def test_indentation_before_a_see_only_bracket_is_kept(catalog):
|
||||
assert render("- 第一条\n （[[see:Q1]]）续行", catalog) == "- 第一条\n 续行"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 依据在括号开头 / 末尾，紧挨着分隔标点
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 在括号开头：后面的标点（和空白）一起去掉
|
||||
("结果一致（[[see:Q1]]，原报告未按状态过滤）。", "结果一致（原报告未按状态过滤）。"),
|
||||
("结果一致（[[see:Q1]]、原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
("结果一致（[[see:Q1]]；原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
("结果一致（[[see:Q1]]：原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
("结果一致([[see:Q1]], 原报告未过滤)。", "结果一致(原报告未过滤)。"),
|
||||
("结果一致（[[see:Q1]] 原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
("结果一致（ [[see:Q1]]，[[see:m:gmv]]，原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
# 在括号末尾：前面的标点（和空白）一起去掉
|
||||
("复算原查询（同 SQL 同过滤条件，[[see:Q1]]）：华南居首。", "复算原查询（同 SQL 同过滤条件）：华南居首。"),
|
||||
("复算原查询(同 SQL, [[see:Q1]])。", "复算原查询(同 SQL)。"),
|
||||
("复算原查询（同 SQL 同过滤条件 [[see:Q1]]）。", "复算原查询（同 SQL 同过滤条件）。"),
|
||||
("订单（[[v:Q1.r0.订单数]] 笔，[[see:Q1]]）", "订单（212 笔）"),
|
||||
("口径（按支付日；[[see:Q1]]、[[see:m:gmv]]）相同。", "口径（按支付日）相同。"),
|
||||
])
|
||||
def test_punctuation_next_to_see_inside_a_bracket_goes_with_it(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 夹在两个分隔标点中间：去掉后一个（括号里外都一样）
|
||||
("口径相同（按支付日，[[see:Q1]]，不含退款）。", "口径相同（按支付日，不含退款）。"),
|
||||
("结果一致，[[see:Q1]]，原报告未过滤。", "结果一致，原报告未过滤。"),
|
||||
("a, [[see:Q1]], b", "a, b"),
|
||||
("a,[[see:Q1]],b", "a, b"), # 半角逗号后面照排版习惯隔一个空格
|
||||
# 前面是分隔标点、后面就到句末：去掉前面那个
|
||||
("华南居首，[[see:Q1]]。", "华南居首。"),
|
||||
("华南居首，[[see:Q1]]", "华南居首"),
|
||||
("华南居首，[[see:Q1]] 。", "华南居首。"),
|
||||
# 几个依据之间的分隔标点
|
||||
("华南居首 [[see:Q1]]，[[see:m:gmv]]。", "华南居首。"),
|
||||
# 后面就到句末、中间只隔着空白：空白一起去掉，不剩「华南居首 。」
|
||||
("华南居首 [[see:Q1]] 。", "华南居首。"),
|
||||
("华南居首 [[see:Q1]] [[see:m:gmv]] 。", "华南居首。"),
|
||||
])
|
||||
def test_punctuation_left_dangling_by_see_is_dropped(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 括号前悬着的停顿标点：去掉括号之后和没写括号的「华南居首，[[see:Q1]]。」一样收拾
|
||||
("华南居首，（[[see:Q1]]）。", "华南居首。"),
|
||||
("华南居首、([[see:Q1]])", "华南居首"),
|
||||
("华南居首 （[[see:Q1]]） 。", "华南居首。"),
|
||||
("华南居首，（[[see:Q1]]），华东第二。", "华南居首，华东第二。"), # 夹在两个逗号中间：去掉后一个
|
||||
("华南居首（（[[see:Q1]]））。", "华南居首。"), # 套着的两层只装着依据，一起去掉
|
||||
# 换行不算句末：括号里外一样留着写作者的标点
|
||||
("华南居首、（[[see:Q1]]）\n下一行", "华南居首、\n下一行"),
|
||||
])
|
||||
def test_a_dropped_bracket_is_tidied_like_a_bare_see(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 括号是两边唯一的分隔：去掉之后数字会粘成另一个数，换成一个空格
|
||||
("华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）2 月起持续下滑。", "华东 4,919.5 2 月起持续下滑。"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）2026 年同比持平。", "合计 4,919.5 2026 年同比持平。"),
|
||||
("订单 [[v:Q1.r0.订单数]]([[see:Q1]])2 个区域。", "订单 212 2 个区域。"),
|
||||
("增幅 [[m:yoy]]（[[see:m:yoy]]）3 个区域。", "增幅 12.5% 3 个区域。"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）[[v:Q1.r0.订单数]] 笔。", "合计 4,919.5 212 笔。"), # 后面的标记渲染出数字
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）[[see:m:gmv]]2 月。", "合计 4,919.5 2 月。"), # 越过不显示的依据看
|
||||
("编号 Q1（[[see:Q1]]）x", "编号 Q1 x"),
|
||||
("212,（[[see:Q1]]）5 个区域", "212, 5 个区域"),
|
||||
# 半角逗号夹着依据：去掉后一个之后，前一个不能直接贴着数字
|
||||
("212,[[see:Q1]],5 个区域", "212, 5 个区域"),
|
||||
# 不会粘的照旧去掉：中文和数字之间、括号前已有空白、后面是中文、隔着全角逗号
|
||||
("居首（[[see:Q1]]）2 月起", "居首2 月起"),
|
||||
("合计 [[v:Q1.r0.销售额]] （[[see:Q1]]）2 月起", "合计 4,919.5 2 月起"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）元", "合计 4,919.5元"),
|
||||
("华南（[[see:Q1]]）居首", "华南居首"),
|
||||
("212，[[see:Q1]]，5 个区域", "212，5 个区域"),
|
||||
])
|
||||
def test_dropping_a_bracket_does_not_glue_numbers_together(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
def test_a_glued_number_would_have_slipped_through_the_checks(catalog):
|
||||
"""粘起来的数没有哪项核对查得出：片段本身对（4,919.5 后面是另一段文字），月份不算裸数字。所以只能在渲染时不粘。"""
|
||||
doc = compose("华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）2 月起持续下滑。", catalog)
|
||||
assert doc["markdown"] == "华东 4,919.5 2 月起持续下滑。" and doc["violations"] == []
|
||||
assert verify_doc(doc, catalog, loader=LOADER)["violations"] == []
|
||||
[unit] = [unit for _, unit in iter_units(doc)]
|
||||
assert [s["alias"] for s in unit["see"]] == ["Q1"]
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 粘不粘看去掉之后真正挨上的两个字：越过依据、只装着依据的括号、粗体斜体的星号和下划线（复核第 2 轮）
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
#: 读者看到的字：前端按 Markdown 渲染，星号、两个下划线不显示
|
||||
_EMPHASIS = re.compile(r"\*+|__+")
|
||||
|
||||
|
||||
def _visible(text: str) -> str:
|
||||
return _EMPHASIS.sub("", text)
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 两个只装着依据的括号连着写（全角半角混用、套着的也一样）：中间什么都不剩，两边会粘的照样隔开
|
||||
("华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）（[[see:m:gmv]]）2 月起持续下滑。", "华东 4,919.5 2 月起持续下滑。"),
|
||||
("订单 [[v:Q1.r0.订单数]]（[[see:Q1]]）（[[see:m:gmv]]）[[v:Q1.r0.订单数]] 笔。", "订单 212 212 笔。"),
|
||||
("增幅 [[m:yoy]]（[[see:m:yoy]]）（[[see:Q1]]）3 个区域。", "增幅 12.5% 3 个区域。"),
|
||||
("增幅 [[m:yoy]]（[[see:m:yoy]]）([[see:Q1]])3 个区域。", "增幅 12.5% 3 个区域。"),
|
||||
("增幅 [[m:yoy]]（[[see:m:yoy]]）（（[[see:Q1]]））3 个区域。", "增幅 12.5% 3 个区域。"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]） （[[see:m:gmv]]）2 月", "合计 4,919.5 2 月"),
|
||||
# 依据和括号连着写、几个依据之间隔着空白：去掉之后两边会粘的，同样隔开
|
||||
("合计 [[v:Q1.r0.销售额]][[see:Q1]]（[[see:m:gmv]]）2 月。", "合计 4,919.5 2 月。"),
|
||||
("合计 [[v:Q1.r0.销售额]][[see:Q1]] [[see:m:gmv]]2 月。", "合计 4,919.5 2 月。"),
|
||||
# 不会粘的照旧：中文和数字之间
|
||||
("华南居首（[[see:Q1]]）（[[see:m:gmv]]）2 月起回升。", "华南居首2 月起回升。"),
|
||||
])
|
||||
def test_see_only_brackets_written_together_do_not_glue_numbers(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 连着写的几个只装着依据的括号当一组收拾：括号前悬着的停顿标点、中间的标点一并去掉
|
||||
("华南居首，（[[see:Q1]]）（[[see:m:gmv]]）。", "华南居首。"),
|
||||
("华南居首，[[see:Q1]]（[[see:m:gmv]]）。", "华南居首。"),
|
||||
("华南居首，（[[see:Q1]]）、（[[see:m:gmv]]）。", "华南居首。"),
|
||||
("结果一致（[[see:Q1]]）（[[see:m:gmv]]），原报告未过滤。", "结果一致，原报告未过滤。"),
|
||||
("结果一致（[[see:Q1]]）（[[see:m:gmv]]，原报告未过滤）。", "结果一致（原报告未过滤）。"),
|
||||
])
|
||||
def test_see_only_brackets_written_together_are_tidied_as_one(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 粗体在括号前、后、把括号包在里面：星号不显示，挨上的是两边的数
|
||||
("销售额 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）2 月起持续下滑。", "销售额 4,919.5 2 月起持续下滑。"),
|
||||
("华东 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）[[v:Q1.r0.订单数]] 笔。", "华东 4,919.5 212 笔。"),
|
||||
("华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）**[[v:Q1.r0.订单数]] 笔**", "华东 4,919.5 212 笔"),
|
||||
("**[[v:Q1.r0.销售额]]（[[see:Q1]]）**2 月", "4,919.5 2 月"),
|
||||
("华东 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）（[[see:m:gmv]]）**[[v:Q1.r0.订单数]]** 笔", "华东 4,919.5 212 笔"),
|
||||
("__[[v:Q1.r0.销售额]]__（[[see:Q1]]）2 月", "4,919.5 2 月"),
|
||||
("*[[v:Q1.r0.销售额]]*（[[see:Q1]]）2 月", "4,919.5 2 月"),
|
||||
# 依据前的空白：后面紧跟着的是星号，越过它看
|
||||
("合计 [[v:Q1.r0.销售额]] [[see:Q1]]**2 月**起", "合计 4,919.5 2 月起"),
|
||||
# 不会粘的照旧
|
||||
("合计 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）元", "合计 4,919.5元"),
|
||||
])
|
||||
def test_bold_around_a_number_does_not_hide_the_glue(catalog, raw, shown):
|
||||
assert _visible(render(raw, catalog)) == shown
|
||||
assert _visible(compose(raw, catalog)["markdown"]) == shown
|
||||
|
||||
|
||||
def test_bold_still_wraps_the_number_after_the_bracket_is_dropped(catalog):
|
||||
doc = compose("销售额 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）2 月起持续下滑。", catalog)
|
||||
[number] = [seg for _, unit in iter_units(doc) for seg in unit["segments"] if seg.get("ref")]
|
||||
assert number["text"] == "4,919.5" and number.get("strong") is True
|
||||
assert render("销售额 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）2 月", catalog) == "销售额 **4,919.5** 2 月"
|
||||
|
||||
|
||||
@pytest.mark.xfail(strict=True, reason="已知限制：粗体把括号包在里面时，补的空格落在收尾的星号前面（「**4,919.5 **2 月」）。"
|
||||
"前端的行内 Markdown 照样认这对星号，组装出的文档去掉了包着标记的星号，读者看到的一样；"
|
||||
"挪到星号外面得先知道这对星号从哪开始，流式渲染的一块里不一定有")
|
||||
def test_the_space_lands_outside_a_bold_that_wraps_the_bracket(catalog):
|
||||
assert render("销售额 **[[v:Q1.r0.销售额]]（[[see:Q1]]）**2 月", catalog) == "销售额 **4,919.5** 2 月"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 依据直接夹在两个数之间：两边没有括号、空白、标点，也不能让两个数并成一个（复核第 3 轮）
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 依据两边什么都没有：渲染成空之后两边直接挨上
|
||||
("华东 [[v:Q1.r0.销售额]][[see:Q1]][[v:Q1.r0.订单数]] 笔。", "华东 4,919.5 212 笔。"),
|
||||
("华东 [[v:Q1.r0.销售额]][[see:Q1]]2 月起持续下滑。", "华东 4,919.5 2 月起持续下滑。"),
|
||||
("增幅 [[m:yoy]][[see:m:yoy]]3 个区域。", "增幅 12.5% 3 个区域。"),
|
||||
("编号 Q1[[see:Q1]]x 区", "编号 Q1 x 区"),
|
||||
("212,[[see:Q1]]5 个区域", "212, 5 个区域"),
|
||||
# 连着写的一串当一个看
|
||||
("华东 [[v:Q1.r0.销售额]][[see:Q1]][[see:m:gmv]]2 月", "华东 4,919.5 2 月"),
|
||||
# 不会粘的照旧：中文和数字之间、后面是中文或单位、写作者自己留了空白、句号后面
|
||||
("华南居首[[see:Q1]]2 月起回升。", "华南居首2 月起回升。"),
|
||||
("合计 [[v:Q1.r0.销售额]][[see:Q1]]元", "合计 4,919.5元"),
|
||||
("合计 [[v:Q1.r0.销售额]][[see:Q1]]%", "合计 4,919.5%"),
|
||||
("合计 [[v:Q1.r0.销售额]][[see:Q1]] 2 月", "合计 4,919.5 2 月"),
|
||||
("合计 [[v:Q1.r0.销售额]] [[see:Q1]]2 月", "合计 4,919.5 2 月"),
|
||||
("华南居首。[[see:Q1]]2 月起回升。", "华南居首。2 月起回升。"),
|
||||
])
|
||||
def test_a_bare_see_between_two_numbers_does_not_glue_them(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 粗体收尾后紧跟依据再跟数字（组装文档时包着标记的星号去掉了，挨上的是两个数）
|
||||
("**华东 [[v:Q1.r0.销售额]]**[[see:Q1]]2 月", "华东 4,919.5 2 月"),
|
||||
("华东 **[[v:Q1.r0.销售额]]**[[see:Q1]][[v:Q1.r0.订单数]] 笔", "华东 4,919.5 212 笔"),
|
||||
("华东 [[v:Q1.r0.销售额]][[see:Q1]]**2 月**起", "华东 4,919.5 2 月起"),
|
||||
("*[[v:Q1.r0.销售额]]*[[see:Q1]]2 月", "4,919.5 2 月"),
|
||||
])
|
||||
def test_a_bare_see_after_bold_does_not_glue_the_numbers(catalog, raw, shown):
|
||||
assert _visible(render(raw, catalog)) == shown
|
||||
assert _visible(compose(raw, catalog)["markdown"]) == shown
|
||||
|
||||
|
||||
def test_the_space_goes_before_the_see_and_outside_the_bold(catalog):
|
||||
"""补的空格落在依据前面、收尾的星号后面：粗体照样成对，依据照样挂在句子上。一片里有去掉的字时落在头一个去掉的字
|
||||
那里（同复核第 2 轮）：粗体在两段依据中间收尾的，空格照样在星号外面。"""
|
||||
assert render("**华东 [[v:Q1.r0.销售额]]**[[see:Q1]]2 月", catalog) == "**华东 4,919.5** 2 月"
|
||||
assert render("**华东 [[v:Q1.r0.销售额]][[see:Q1]]**（[[see:m:gmv]]）2 月", catalog) == "**华东 4,919.5** 2 月"
|
||||
doc = compose("华东 [[v:Q1.r0.销售额]][[see:Q1]][[v:Q1.r0.订单数]] 笔。", catalog)
|
||||
[unit] = [unit for _, unit in iter_units(doc)]
|
||||
assert [s["alias"] for s in unit["see"]] == ["Q1"]
|
||||
assert [seg["text"] for seg in unit["segments"] if seg.get("ref")] == ["4,919.5", "212"]
|
||||
assert verify_doc(doc, catalog, loader=LOADER)["violations"] == []
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 右边是紧贴着下文的半角逗号：去掉依据（和只装着依据的括号）之后读成一个数（「1,200,900」），空格补在逗号后面
|
||||
("分别为 [[v:Q1.r0.订单数]]（[[see:Q1]]）,5 个区域", "分别为 212, 5 个区域"),
|
||||
("分别为 [[v:Q1.r0.销售额]][[see:Q1]],[[v:Q1.r1.销售额]]。", "分别为 4,919.5, 1,891.8。"),
|
||||
("分别为 [[v:Q1.r0.销售额]]【[[see:Q1]]】,[[v:Q1.r1.销售额]]（[[see:Q1]]）。", "分别为 4,919.5, 1,891.8。"),
|
||||
("212[[see:Q1]],5 个区域", "212, 5 个区域"),
|
||||
("分别为 [[v:Q1.r0.销售额]]（[[see:Q1]]）,**[[v:Q1.r1.销售额]]**。", "分别为 4,919.5, **1,891.8**。"), # 下文包着粗体
|
||||
# 不会读成一串的照旧：逗号后面已有空白、是中文，逗号前面是中文
|
||||
("分别为 [[v:Q1.r0.订单数]]（[[see:Q1]]）, 5 个区域", "分别为 212, 5 个区域"),
|
||||
("分别为 [[v:Q1.r0.订单数]]（[[see:Q1]]）,华东第二", "分别为 212,华东第二"),
|
||||
("分别为 [[v:Q1.r0.订单数]][[see:Q1]],[[see:m:gmv]]华东第二", "分别为 212,华东第二"), # 逗号后面的依据不显示
|
||||
("华南居首（[[see:Q1]]）,5 个区域", "华南居首,5 个区域"),
|
||||
])
|
||||
def test_a_half_width_comma_left_against_the_next_number_gets_a_space(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert _visible(compose(raw, catalog)["markdown"]) == _visible(shown)
|
||||
|
||||
|
||||
def test_a_struck_number_does_not_glue_either(catalog):
|
||||
"""删除线前端照样画成字（淡色），后面紧跟的数挨着它读成一串：星号、下划线之外，两个波浪线也越过去看。"""
|
||||
assert render("~~[[v:Q1.r0.销售额]]~~[[see:Q1]]2 月", catalog) == "~~4,919.5~~ 2 月"
|
||||
assert render("~~[[v:Q1.r0.销售额]]~~（[[see:Q1]]）2 月", catalog) == "~~4,919.5~~ 2 月"
|
||||
assert render("[[v:Q1.r0.销售额]][[see:Q1]]~~2 月~~", catalog) == "4,919.5 ~~2 月~~"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 别的括号：【】［］〔〕[] 只装着依据的，和（）一样去掉（复核第 3 轮）
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
("合计 [[v:Q1.r0.销售额]]【[[see:Q1]]】。", "合计 4,919.5。"),
|
||||
("合计 [[v:Q1.r0.销售额]]［[[see:Q1]]］。", "合计 4,919.5。"),
|
||||
("合计 [[v:Q1.r0.销售额]]〔[[see:Q1]]〕。", "合计 4,919.5。"),
|
||||
("合计 [[v:Q1.r0.销售额]][[[see:Q1]]]。", "合计 4,919.5。"),
|
||||
("结果一致［[[see:Q1]]]。", "结果一致。"), # 方括号一对里全角半角混用
|
||||
("结果一致【[[see:Q1]]，[[see:m:gmv]]】。", "结果一致。"),
|
||||
("结果一致 【 [[see:Q1]] 】 。", "结果一致。"),
|
||||
("前一句 【[[see:Q1]]】后一句", "前一句 后一句"),
|
||||
("末尾[[[see:Q1]]]", "末尾"),
|
||||
# 去掉之后整对括号当一个依据，外面悬着的停顿标点照原规则收拾
|
||||
("华南居首，【[[see:Q1]]】。", "华南居首。"),
|
||||
("华南居首，[[[see:Q1]]]，华东第二。", "华南居首，华东第二。"),
|
||||
# 套着的、和别的括号连着写的
|
||||
("华南居首（【[[see:Q1]]】）。", "华南居首。"),
|
||||
("华南居首【[[see:Q1]]】（[[see:m:gmv]]）〔[[see:Q1]]〕。", "华南居首。"),
|
||||
# 依据在括号开头、末尾，紧挨着分隔标点
|
||||
("结果一致【[[see:Q1]]，原报告未过滤】。", "结果一致【原报告未过滤】。"),
|
||||
("复算原查询〔同 SQL，[[see:Q1]]〕。", "复算原查询〔同 SQL〕。"),
|
||||
("结果一致[[[see:Q1]]; 原报告未过滤]。", "结果一致[原报告未过滤]。"),
|
||||
# 去掉之后两边的数不能并成一个
|
||||
("华东 [[v:Q1.r0.销售额]]【[[see:Q1]]】2 月", "华东 4,919.5 2 月"),
|
||||
("华东 [[v:Q1.r0.销售额]][[[see:Q1]]]2 月", "华东 4,919.5 2 月"),
|
||||
("华东 [[v:Q1.r0.销售额]]〔[[see:Q1]]〕[[v:Q1.r0.订单数]] 笔", "华东 4,919.5 212 笔"),
|
||||
])
|
||||
def test_other_brackets_holding_only_see_are_dropped(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize("raw", [
|
||||
"【注】原样，［1］、〔2026〕、[a] 也是。",
|
||||
"空的【】、［］、〔〕、[] 原样，【，】也是。",
|
||||
"【证据 Q1】华南居首。",
|
||||
])
|
||||
def test_other_brackets_without_see_are_left_alone(catalog, raw):
|
||||
assert render(raw, catalog) == raw
|
||||
assert compose(raw, catalog)["markdown"] == raw
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
# 依据标记的「]」不显示，不当数字：左边真正挨上的是句号、文首
|
||||
("华南居首。[[see:Q1]]（[[see:m:gmv]]）2 月起回升。", "华南居首。2 月起回升。"),
|
||||
("[[see:Q1]]（[[see:Q1]]）2 月", "2 月"),
|
||||
("华南居首。[[see:Q1]] （[[see:m:gmv]]）2 月起回升。", "华南居首。2 月起回升。"), # 中间的空白同几个依据之间的
|
||||
# 一片里最多补一个空格；挨上的本来就隔着空白的不补
|
||||
("212,[[see:Q1]],[[see:m:gmv]],5 个区域", "212, 5 个区域"),
|
||||
("212,（[[see:Q1]]）,[[see:m:gmv]],5 个区域", "212, 5 个区域"),
|
||||
("合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）（[[see:m:gmv]]） 2 月", "合计 4,919.5 2 月"),
|
||||
])
|
||||
def test_a_space_is_put_back_once_and_only_where_it_sticks(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
#: 结构化组合：左边一个渲染出数的标记（可能包着粗体），中间 1–3 段依据、只装着依据的括号（几类括号）、空白、星号，
|
||||
#: 右边以数字、字母或者渲染出数的标记开头
|
||||
_LEFTS = ["合计 [[v:Q1.r0.销售额]]", "合计 **[[v:Q1.r0.销售额]]**", "合计 __[[v:Q1.r0.销售额]]__", "增幅 [[m:yoy]]",
|
||||
"编号 Q1"]
|
||||
_MIDDLES = ["（[[see:Q1]]）", "([[see:m:gmv]])", "（（[[see:Q1]]））", "【[[see:Q1]]】", "[[[see:m:gmv]]]", "[[see:Q1]]",
|
||||
"**", " ", "，"]
|
||||
_RIGHTS = ["2 月", "[[v:Q1.r0.订单数]] 笔", "**[[v:Q1.r0.订单数]] 笔**", "x 区"]
|
||||
#: 写作者原本就隔开了两边的中间段：括号（去掉依据之后还看得见）、空白、标点
|
||||
_SEPARATING = {"（[[see:Q1]]）", "([[see:m:gmv]])", "（（[[see:Q1]]））", "【[[see:Q1]]】", "[[[see:m:gmv]]]", " ", "，"}
|
||||
|
||||
|
||||
def _combos() -> list[tuple[str, str, str]]:
|
||||
"""写作者隔开了两边、或者中间写了依据的组合（中间只有星号的，是写作者自己把两个数写在了一起，不归这里管）。"""
|
||||
middles = [[m] for m in _MIDDLES]
|
||||
middles += [[a, b] for a in _MIDDLES for b in _MIDDLES]
|
||||
middles += [[a, b, c] for a in _MIDDLES for b in _MIDDLES for c in _MIDDLES]
|
||||
return [(left, "".join(middle), right) for left in _LEFTS for right in _RIGHTS for middle in middles
|
||||
if any(m in _SEPARATING or "see" in m for m in middle)]
|
||||
|
||||
|
||||
#: 左边那个数渲染出来的字，后面紧跟着数字、字母，或者半角的逗号、句点再跟数字：读成了另一个数、另一个词
|
||||
_GLUED = re.compile(r"(?:4,919\.5|12\.5%|Q1)(?=[\dA-Za-z]|[,.]\d)")
|
||||
|
||||
|
||||
def test_two_numbers_never_run_together_once_see_is_dropped(catalog):
|
||||
"""写作者隔开了的两边（括号、空白、标点），或者中间只隔着依据的，收拾依据之后不能挨成一串；也不会多出两个连着的
|
||||
空格。"""
|
||||
bad = []
|
||||
for left, middle, right in _combos():
|
||||
raw = f"{left}{middle}{right}。"
|
||||
shown = _visible(render(raw, catalog))
|
||||
writer = _visible(re.sub(r"\[\[see:[^\]]*\]\]", "", raw))
|
||||
if _GLUED.search(shown) or (" " in shown and " " not in writer):
|
||||
bad.append((raw, shown))
|
||||
assert not bad, bad[:10]
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("raw", "shown"), [
|
||||
("华南居首。[[see:Q1]]", "华南居首。"),
|
||||
("华南居首 [[see:Q1]]，华东第二。", "华南居首，华东第二。"),
|
||||
# 两个依据中间的逗号后面接着字：它是分句的逗号，不是依据之间的分隔，照旧留着
|
||||
("华南居首 [[see:Q1]]，[[see:m:gmv]]华东第二。", "华南居首，华东第二。"),
|
||||
("华南居首 [[see:Q1]]，", "华南居首，"), # 依据后面的逗号是写作者自己的，不动
|
||||
("数据如下：[[see:Q1]]\n| 区域 | 销售额 |", "数据如下：\n| 区域 | 销售额 |"), # 冒号引出的是下一行
|
||||
# 换行不算句末：列表项末尾的「；」、软换行处连着两个分句的「，」是写作者自己的
|
||||
("1. 华东居首；[[see:Q1]]\n2. 华南第二；[[see:Q1]]\n3. 华北第三。[[see:Q1]]",
|
||||
"1. 华东居首；\n2. 华南第二；\n3. 华北第三。"),
|
||||
("华东居首，[[see:Q1]]\n华南第二。", "华东居首，\n华南第二。"),
|
||||
("华南居首、华东第二；[[see:Q1]]\n下一行", "华南居首、华东第二；\n下一行"),
|
||||
("- 华东居首；（[[see:Q1]]）\n- 华南第二。", "- 华东居首；\n- 华南第二。"),
|
||||
])
|
||||
def test_punctuation_the_writer_put_there_is_kept(catalog, raw, shown):
|
||||
assert render(raw, catalog) == shown
|
||||
assert compose(raw, catalog)["markdown"] == shown
|
||||
|
||||
|
||||
#: 引出语悬空：「依据」「见」「取自」「来源」「出处」后面（可能隔着冒号、空白）紧跟着句末、停顿标点或括号
|
||||
_DANGLING = re.compile(r"(?:依据|见|取自|来源|出处)[：:]?[ \t]*(?:[。，、；）)！？,;]|$)")
|
||||
|
||||
|
||||
@pytest.mark.xfail(strict=True, reason="已知限制：依据写在引出语后面时，渲染后引出语悬空。引出语在依据前面，渲染时要去掉"
|
||||
"它，流式渲染得先攒住每个可能是引出语的字；目前靠核查 agent 的提示词避免这种写法，模型"
|
||||
"不照做时记一条日志（verify_agent._LEAD_IN_SEE）")
|
||||
@pytest.mark.parametrize("raw", [
|
||||
"结果见：[[see:Q1]]。",
|
||||
"依据 [[see:Q1]]。",
|
||||
"数据取自 [[see:Q1]]。",
|
||||
"结果取自 [[see:Q1]]，完整未截断。",
|
||||
"按支付日统计（见 [[see:Q1]]）。",
|
||||
"按支付日统计（来源：[[see:Q1]]）。",
|
||||
"按状态拆开（status 取值见 [[see:Q1]]）。",
|
||||
])
|
||||
def test_see_after_a_lead_in_word_leaves_it_dangling(catalog, raw):
|
||||
assert not _DANGLING.search(render(raw, catalog)), render(raw, catalog)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("raw", [
|
||||
"空括号（）原样保留，半角 () 也是。",
|
||||
"说明（含退款，不含测试账号）；另见 (a, b)。",
|
||||
"（，）原样，(;) 也是。",
|
||||
"甲，乙、丙；丁：戊。",
|
||||
])
|
||||
def test_brackets_and_punctuation_without_see_are_left_alone(catalog, raw):
|
||||
assert render(raw, catalog) == raw
|
||||
assert compose(raw, catalog)["markdown"] == raw
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 依据仍挂在句子上
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
def test_the_sentence_still_carries_the_see_that_was_in_brackets(catalog):
|
||||
doc = compose("先按状态核对口径（[[see:Q1]]）：已付 [[v:Q1.r0.订单数]] 笔。"
|
||||
"结果一致（[[see:m:gmv]]，原报告未过滤）。", catalog)
|
||||
assert doc["markdown"] == "先按状态核对口径：已付 212 笔。结果一致（原报告未过滤）。"
|
||||
first, second = [unit for _, unit in iter_units(doc)]
|
||||
assert [s["alias"] for s in first["see"]] == ["Q1"] and first["cites"] == ["Q1"]
|
||||
assert [s["alias"] for s in second["see"]] == ["m:gmv"] and second["cites"] == ["m:gmv"]
|
||||
assert doc["stats"]["see"] == 2 and doc["stats"]["uncited_claims"] == 0
|
||||
assert verify_doc(doc, catalog, loader=LOADER)["violations"] == []
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 没有依据写进括号的文档：渲染和内容哈希都和改动之前一样
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
#: 改动之前（基线 58a6759）组装出来的文档的内容哈希
|
||||
BEFORE = {
|
||||
"plain": ("## 本周概览\n\n本周（[[i:week]]）销售额 [[m:gmv]]（含退款，不含测试账号），同比 [[m:yoy]]；"
|
||||
"华南 [[v:Q1.r0.销售额]]（[[v:Q1.r0.订单数]] 笔）、华东 [[v:Q1.r1.销售额]]。\n\n"
|
||||
"- 空括号（）原样保留，半角 (a, b) 也是\n- 分隔：甲，乙、丙；丁\n\n"
|
||||
"| 区域 | 销售额（元） |\n| --- | --- |\n| 华南（含线上） | [[v:Q1.r0.销售额]] |",
|
||||
"49e7a6e4e430f0cbc495b5ab2b5e319f146bbee2c7147d663494922aaf6773ba"),
|
||||
"see_at_end": ("本周销售额 [[m:gmv]]（含退款）。[[see:m:gmv]]同比 [[m:yoy]]，增幅放缓 [[see:m:yoy]]。\n\n"
|
||||
"华南居首（[[v:Q1.r0.订单数]] 笔），华东第二。[[see:Q1]]\n- 列表项 [[see:Q1]]\n",
|
||||
"1bb0bc31fffa6de8e906d24022dae7fa3cf9b43aeda47d319498f1593d5a3684"),
|
||||
}
|
||||
|
||||
|
||||
@pytest.mark.parametrize("name", list(BEFORE))
|
||||
def test_documents_without_see_in_brackets_hash_as_before(catalog, name):
|
||||
raw, digest = BEFORE[name]
|
||||
assert content_hash(canonical_json(compose(raw, catalog))) == digest
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 流式渲染：分块输出拼起来仍然等于整篇渲染
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
def _stream(catalog, raw: str, sizes: list[int]) -> list[str]:
|
||||
renderer, out, i = StreamRenderer(catalog, loader=LOADER), [], 0
|
||||
for size in sizes:
|
||||
if i >= len(raw):
|
||||
break
|
||||
out.append(renderer.feed(raw[i:i + size]))
|
||||
i += size
|
||||
if i < len(raw):
|
||||
out.append(renderer.feed(raw[i:]))
|
||||
out.append(renderer.flush())
|
||||
return out
|
||||
|
||||
|
||||
@pytest.mark.parametrize("raw", [
|
||||
"结果一致（[[see:Q1]]，原报告未按状态过滤），合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）。下一句",
|
||||
"前一句 （[[see:Q1]]）后一句，复算原查询（同 SQL，[[see:Q1]]）：华南居首，[[see:Q1]]。",
|
||||
"口径相同（按支付日，[[see:Q1]]，不含退款）。a, [[see:Q1]], b\n华南、[[see:Q1]]\n末尾（[[see:Q1]]）",
|
||||
"华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）2 月，合计 [[v:Q1.r0.销售额]]（[[see:Q1]]）[[see:m:gmv]][[v:Q1.r0.订单数]] 笔"
|
||||
"，212,[[see:Q1]],5 个",
|
||||
"华南居首，（[[see:Q1]]）。华南居首 [[see:Q1]] 。1. 华东居首；[[see:Q1]]\n2. 华南第二，（[[see:Q1]]），华东第二",
|
||||
# 粗体、连着写的括号：往两边找挨上的字要越过星号、下划线，流式渲染得把它们和依据一起攒着
|
||||
"销售额 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）2 月，华东 [[v:Q1.r0.销售额]]（[[see:Q1]]）（[[see:m:gmv]]）"
|
||||
"**[[v:Q1.r0.订单数]] 笔**",
|
||||
"华南居首。[[see:Q1]]（[[see:m:gmv]]）2 月，212,[[see:Q1]],[[see:m:gmv]],5 个，合计 [[v:Q1.r0.销售额]] [[see:Q1]]"
|
||||
"__2 月__",
|
||||
"**[[v:Q1.r0.销售额]]（[[see:Q1]]）**2 月；华南居首，（[[see:Q1]]）（[[see:m:gmv]]）。order_（[[see:Q1]]）_x",
|
||||
# 依据直接夹在两个数之间、别的括号（复核第 3 轮）：方括号、波浪线也得和依据一起攒着
|
||||
"华东 [[v:Q1.r0.销售额]][[see:Q1]][[v:Q1.r0.订单数]] 笔，[[v:Q1.r0.销售额]][[see:Q1]][[see:m:gmv]]2 月，"
|
||||
"**华东 [[v:Q1.r0.销售额]]**[[see:Q1]]2 月",
|
||||
"合计 [[v:Q1.r0.销售额]]【[[see:Q1]]】2 月，[[v:Q1.r0.销售额]][[[see:Q1]]]2 月，华南居首，［[[see:Q1]]]。"
|
||||
"结果一致【[[see:Q1]]，原报告未过滤】〔[[see:m:gmv]]〕[[v:Q1.r0.订单数]] 笔",
|
||||
"~~[[v:Q1.r0.销售额]]~~[[see:Q1]]2 月，[[v:Q1.r0.销售额 ]][[see:Q1]][[v:Q1.r0.订单数]]，a[0] 与 [b] 原样]",
|
||||
"分别为 [[v:Q1.r0.销售额]]（[[see:Q1]]）,[[v:Q1.r1.销售额]]，212[[see:Q1]],5 个，[[m:gmv]][[see:Q1]],[[m:gmv]] 元",
|
||||
])
|
||||
@pytest.mark.parametrize("size", range(1, 9))
|
||||
def test_stream_tidies_see_in_brackets_like_the_full_render(catalog, raw, size):
|
||||
out = _stream(catalog, raw, [size] * len(raw))
|
||||
assert "".join(out) == render(raw, catalog), out
|
||||
assert all("[[see:" not in piece for piece in out)
|
||||
|
||||
|
||||
def _corpus(n: int) -> list[str]:
|
||||
rng = random.Random(20261006)
|
||||
pieces = ["[[see:Q1]]", "[[see:m:gmv]]", "[[v:Q1.r0.销售额]]", "[[m:gmv]]", "（", "）", "(", ")", "，", "、",
|
||||
"；", "：", ",", ";", " ", " ", "。", "\n", "文字", "元", "[[", "]]", "[", "2", "x", "%", "！",
|
||||
"**", "*", "__", "_", "（[[see:Q1]]）", "【", "】", "［", "］", "〔", "〕", "]", "~~", "【[[see:Q1]]】",
|
||||
"[[[see:m:gmv]]]"]
|
||||
return ["".join(rng.choice(pieces) for _ in range(rng.randint(1, 18))) for _ in range(n)]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("raw", _corpus(200))
|
||||
def test_stream_matches_the_full_render_on_random_brackets(catalog, raw):
|
||||
rng = random.Random(raw)
|
||||
full = render(raw, catalog)
|
||||
for _ in range(3):
|
||||
sizes = [rng.randint(1, 12) for _ in range(len(raw))]
|
||||
assert "".join(_stream(catalog, raw, sizes)) == full
|
||||
|
||||
|
||||
def test_stream_matches_the_full_render_on_the_structured_combos(catalog):
|
||||
"""结构化组合（数、依据、只装着依据的括号、空白、星号、下划线）逐字、每 4 个字一块地送进去，拼起来和整篇渲染一样。"""
|
||||
bad = []
|
||||
for left, middle, right in _combos():
|
||||
raw = f"{left}{middle}{right}。"
|
||||
full = render(raw, catalog)
|
||||
bad += [(raw, out) for size in (1, 4) if "".join(out := _stream(catalog, raw, [size] * len(raw))) != full]
|
||||
assert not bad, bad[:5]
|
||||
@@ -0,0 +1,117 @@
|
||||
"""审批质询的回答：依据写进括号不再断句，收尾超时记时长上限，占位词不当回答（期 5 真实模型验收 M1）。
|
||||
|
||||
真实模型验收里，质询 agent 写的「结果一致（[[see:Q3]]，原报告 SQL 未按 status 过滤）」「合计 [[v:Q1.r1.金额合计]]
|
||||
（[[see:Q1]]）」渲染成「（，原报告…）」「（）」，随正式运行封存；另一次回答查完了，收尾那次调用超过时长上限，报错写着
|
||||
「回答超过时长上限（180 秒）」，answer.limit 却是 null，界面上缺了「已达时长上限」。
|
||||
|
||||
回答只写了依据（「（[[see:Q1]]）」）时显示出来是空的：以前照样交回，封存之后的追问末尾补上「注：以上回答依据的是数据源
|
||||
当前的数据…」，回答正文只剩这句系统说明、却记成回答成功。现在按没给回答记。回答就是一个「无。」的，照常是回答。
|
||||
|
||||
模型一律是 MockChatModel 剧本，区域、金额都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from app.core import artifact_store
|
||||
from app.engine import verify_agent
|
||||
from app.engine.inquiry import NO_ANSWER
|
||||
from tests.test_inquiry_api import ( # noqa: F401 - 夹具按名字引入
|
||||
EAST,
|
||||
NO_TEST,
|
||||
QUESTION,
|
||||
Models,
|
||||
Play,
|
||||
ask,
|
||||
client,
|
||||
engine_up,
|
||||
events,
|
||||
finish,
|
||||
pause,
|
||||
pending,
|
||||
report_doc,
|
||||
set_verify,
|
||||
shop,
|
||||
unit_of,
|
||||
)
|
||||
from tests.test_verify_rationale_rules import slow_final
|
||||
|
||||
#: 依据写进了括号：开头跟着逗号的、只有依据的
|
||||
BRACKETED = ("仍居各区域之首（[[see:Q1]]，已排除测试账号）：华东为 [[v:Q1.r0.销售额]]（[[see:Q1]]），"
|
||||
"华南为 [[v:Q1.r1.销售额]]（同口径，[[see:Q1]]）。")
|
||||
BRACKETED_TEXT = "仍居各区域之首（已排除测试账号）：华东为 1,200，华南为 900（同口径）。"
|
||||
#: 两个只装着依据的括号连着写、数字包着粗体（复核第 2 轮）：去掉括号之后两个数不能挨成一个（「1,200900」），句末前悬着的
|
||||
#: 逗号也去掉
|
||||
TOGETHER = ("华东、华南分别为 **[[v:Q1.r0.销售额]]**（[[see:Q1]]）（[[see:Q1]]）[[v:Q1.r1.销售额]]，"
|
||||
"华东仍居各区域之首，（[[see:Q1]]）（[[see:Q1]]）。")
|
||||
TOGETHER_TEXT = "华东、华南分别为 1,200 900，华东仍居各区域之首。"
|
||||
#: 依据直接夹在两个数之间（两边什么都没有）、写进方头括号（复核第 3 轮）：两个数不能挨成一个，括号不能剩下「【】」
|
||||
BARE = "华东、华南分别为 [[v:Q1.r0.销售额]][[see:Q1]][[v:Q1.r1.销售额]]，华东仍居各区域之首【[[see:Q1]]】。"
|
||||
BARE_TEXT = "华东、华南分别为 1,200 900，华东仍居各区域之首。"
|
||||
|
||||
|
||||
async def _asked(client, shop) -> tuple[dict, list]:
|
||||
paused = await pause(shop)
|
||||
[approval] = await pending(paused.id)
|
||||
r = await ask(client, paused.id, report="write", unit=None, question=QUESTION, approval_id=approval.id)
|
||||
assert r.status_code == 200, r.text
|
||||
return r.json()["inquiry"], await events(paused.id, "approval.inquiry")
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("said", "shown"), [(BRACKETED, BRACKETED_TEXT), (TOGETHER, TOGETHER_TEXT), (BARE, BARE_TEXT)],
|
||||
ids=["bracketed", "together", "bare"])
|
||||
async def test_see_in_brackets_does_not_break_the_answer(monkeypatch, shop, client, said, shown):
|
||||
Models(monkeypatch, shop, plays={"排除测试账号": Play([NO_TEST], said)})
|
||||
item, [event] = await _asked(client, shop)
|
||||
answer = item["answer"]
|
||||
assert answer["ok"] is True and answer["text"] == shown
|
||||
doc = artifact_store.load(answer["doc_artifact"])
|
||||
assert doc["markdown"] == shown and doc["source"] == said
|
||||
assert doc["violations"] == []
|
||||
assert event.data["answer_text"] == shown
|
||||
|
||||
|
||||
async def test_a_final_answer_that_runs_out_of_time_reports_the_time_limit(monkeypatch, shop, client):
|
||||
"""质询 agent 查完了、自己停下，交回答那次调用超过时长上限：报错说时长上限，limit 也是 timeout_s（事件里同样）。"""
|
||||
await set_verify({"inquiry": {"timeout_s": 0.5}})
|
||||
monkeypatch.setattr(verify_agent, "SETTLE_GRACE_S", 0.3)
|
||||
models = Models(monkeypatch, shop)
|
||||
slow_final(monkeypatch)
|
||||
item, [event] = await _asked(client, shop)
|
||||
answer = item["answer"]
|
||||
assert answer["ok"] is False and answer["text"] == "" and answer["doc_artifact"] is None
|
||||
assert answer["error"] == "回答超过时长上限（0.5 秒），未能给出回答，可缩小问题范围后重新提问"
|
||||
assert answer["limit"] == "timeout_s"
|
||||
assert [q["sql"] for q in item["queries"]] == [NO_TEST], "查到的照记"
|
||||
assert (event.data["limit"], event.data["error"]) == ("timeout_s", answer["error"])
|
||||
assert models.finals == [], "交回答那次没等到回复"
|
||||
|
||||
|
||||
async def test_a_placeholder_answer_is_no_answer(monkeypatch, shop, client):
|
||||
Models(monkeypatch, shop, plays={"排除测试账号": Play([NO_TEST], "placeholder")})
|
||||
item, _ = await _asked(client, shop)
|
||||
answer = item["answer"]
|
||||
assert (answer["ok"], answer["text"], answer["limit"]) == (False, "", None)
|
||||
assert answer["error"] == NO_ANSWER
|
||||
|
||||
|
||||
async def test_an_answer_that_shows_nothing_is_no_answer_even_after_the_seal(monkeypatch, shop, client):
|
||||
"""封存之后的追问：回答只写了依据，显示出来是空的。以前回答文档照样组装，末尾补上「查的是现在的数据」那句说明，
|
||||
回答正文只剩系统说明、却记成回答成功。"""
|
||||
Models(monkeypatch, shop, plays={"排除测试账号": Play([NO_TEST], "（[[see:Q1]]）")})
|
||||
done = await finish(shop)
|
||||
assert done.status == "succeeded", done.error
|
||||
r = await ask(client, done.id, report="write", unit=unit_of(await report_doc(done.id), EAST), question=QUESTION)
|
||||
assert r.status_code == 200, r.text
|
||||
answer = r.json()["inquiry"]["answer"]
|
||||
assert (answer["ok"], answer["text"], answer["doc_artifact"]) == (False, "", None)
|
||||
assert answer["error"] == NO_ANSWER
|
||||
|
||||
|
||||
async def test_a_one_word_answer_is_an_answer(monkeypatch, shop, client):
|
||||
"""问的是「有没有」，答「无。」：这是回答，不是占位词，不能报「模型没有给出回答内容」让提问人重问、再花一次钱。"""
|
||||
Models(monkeypatch, shop, plays={"排除测试账号": Play([NO_TEST], "无。[[see:Q1]]")})
|
||||
item, [event] = await _asked(client, shop)
|
||||
answer = item["answer"]
|
||||
assert (answer["ok"], answer["text"], answer["error"]) == (True, "无。", None)
|
||||
assert event.data["answer_text"] == "无。"
|
||||
@@ -0,0 +1,243 @@
|
||||
"""核查 agent 交理由的写法和收尾超时记的上限（期 5 真实模型验收 M1）。
|
||||
|
||||
- 交结果的提示（VERIFY_FINAL_PROMPT，双盲复算、反方审稿、审批质询共用）和写作者一致：依据写在句末，不放进括号或句子
|
||||
中间。以前写的是「提到某次查询时写 [[see:Q1]]」，模型就写成「结果一致（[[see:Q3]]，原报告…）」，反方审稿的系统提示
|
||||
自己也把它写在括号里（「（[[v:Qn.rX.列名]] 或 [[see:Qn]]）」）
|
||||
- 写进了括号的依据，理由文档里照样渲染干净（evidence._tidy），这句仍挂着那条依据
|
||||
- 模型自己停了工具调用、收尾那次调用超过了时长上限：limit 照实记 timeout_s。以前是 None——质询的报错写着「回答超过
|
||||
时长上限（180 秒）」，limit 却是空的，界面上缺了「已达时长上限」。不走结构化输出的模型（普通调用收尾）也一样
|
||||
- 交回的理由是空的、只有标点，或者只是一个占位词（字面的 placeholder、N/A）：当作没给理由，不把占位词当理由展示。
|
||||
按渲染出来的正文判：只写了依据的（[[see:Q1]]、「（[[see:Q1]]）」）显示出来是空的，占位词后面挂着依据的显示出来只剩
|
||||
占位词，都算没给。理由就是结果本身时（审批质询，结构里只有理由一项），「无」「none」是回答，不算占位
|
||||
- 依据写在「依据」「见」「取自」这类引出语后面（渲染时去不掉，引出语悬空）：记一条日志，验收时查得到
|
||||
|
||||
模型一律是 MockChatModel 剧本，数据源是现造的 SQLite 文件，数字都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import logging
|
||||
import re
|
||||
import time
|
||||
|
||||
import pytest
|
||||
|
||||
from app.engine import verify_agent
|
||||
from app.engine.evidence import iter_units
|
||||
from app.engine.inquiry import INQUIRY_SCHEMA, INQUIRY_SYSTEM, INQUIRY_TASK_PROMPT
|
||||
from app.engine.nodes.challenge import CHALLENGE_SYSTEM, CHALLENGE_TASK_PROMPT
|
||||
from app.engine.nodes.dual_check import BLIND_SYSTEM, BLIND_TASK_PROMPT
|
||||
from app.engine.verify_agent import VERIFY_FINAL_PROMPT, VERIFY_SYSTEM
|
||||
from app.providers.mock_model import MockChatModel
|
||||
from tests.test_verify_agent import VALID_SUM, Script, clean_spend, q, run, shop # noqa: F401 - 夹具按名字引入
|
||||
|
||||
#: 括号里写着依据标记：「（[[v:Qn.rX.列名]] 或 [[see:Qn]]）」「（[[see:Q1]]）」这种示范
|
||||
_BRACKETED_SEE = re.compile(r"[（(][^（）()]*\[\[see:[^\]]*\]\][^（）()]*[）)]")
|
||||
FOUND = {"value": {"alias": "Q1", "row": 0, "column": "total"}, "verdict": "一致"}
|
||||
|
||||
|
||||
def slow_final(monkeypatch, seconds: float = 5.0) -> None:
|
||||
"""交结果那次调用（消息里带着 VERIFY_FINAL_PROMPT：结构化输出、退回普通调用都是）迟迟不返回，查库的几轮照常。"""
|
||||
original = MockChatModel._agenerate
|
||||
|
||||
async def held(chat, messages, stop=None, run_manager=None, **kwargs):
|
||||
if any(VERIFY_FINAL_PROMPT in str(m.content) for m in messages):
|
||||
await asyncio.sleep(seconds)
|
||||
return await original(chat, messages, stop, run_manager, **kwargs)
|
||||
|
||||
monkeypatch.setattr(MockChatModel, "_agenerate", held)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 提示词
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
def test_the_final_prompt_puts_see_at_the_end_of_the_sentence():
|
||||
assert "提到某次查询时写 [[see:Q1]]" not in VERIFY_FINAL_PROMPT
|
||||
assert "句末" in VERIFY_FINAL_PROMPT and "括号" in VERIFY_FINAL_PROMPT
|
||||
assert re.search(r"。\[\[see:Q\d\]\]", VERIFY_FINAL_PROMPT), "示范的写法：依据挂在句号后面"
|
||||
assert "渲染后不显示" in VERIFY_FINAL_PROMPT
|
||||
|
||||
|
||||
@pytest.mark.parametrize("prompt", [VERIFY_SYSTEM, VERIFY_FINAL_PROMPT, CHALLENGE_SYSTEM, CHALLENGE_TASK_PROMPT,
|
||||
INQUIRY_SYSTEM, INQUIRY_TASK_PROMPT, BLIND_SYSTEM, BLIND_TASK_PROMPT])
|
||||
def test_no_verify_prompt_shows_see_inside_brackets(prompt):
|
||||
assert not _BRACKETED_SEE.search(prompt), _BRACKETED_SEE.search(prompt).group(0)
|
||||
|
||||
|
||||
async def test_the_final_call_is_told_the_rule(monkeypatch, shop):
|
||||
name, _ = shop
|
||||
script = Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": "合计 [[v:Q1.r0.total]]。[[see:Q1]]"})
|
||||
await run(name)
|
||||
kind, messages = script.calls[-1]
|
||||
assert kind == "final" and VERIFY_FINAL_PROMPT in str(messages[-1].content)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 写进括号的依据
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
async def test_see_written_in_brackets_renders_clean_in_the_rationale(monkeypatch, shop):
|
||||
name, _ = shop
|
||||
said = "未作废订单的金额合计是 [[v:Q1.r0.total]]（[[see:Q1]]），口径一致（[[see:Q1]]，已排除作废单）。"
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
shown = "未作废订单的金额合计是 230，口径一致（已排除作废单）。"
|
||||
assert out.result["rationale"] == shown and out.rationale["markdown"] == shown
|
||||
assert out.rationale["source"] == said, "原文照存，依据标记还在"
|
||||
[unit] = [unit for _, unit in iter_units(out.rationale)]
|
||||
assert [s["alias"] for s in unit["see"]] == ["Q1", "Q1"] and unit["cites"] == ["Q1"]
|
||||
assert out.rationale["violations"] == []
|
||||
|
||||
|
||||
async def test_see_brackets_written_together_do_not_glue_two_numbers(monkeypatch, shop):
|
||||
"""两个只装着依据的括号连着写、数字包着粗体（复核第 2 轮）：去掉括号之后两个数不能挨成一个（「2303 笔」），句末前
|
||||
悬着的逗号也去掉。"""
|
||||
name, _ = shop
|
||||
said = ("未作废订单合计 **[[v:Q1.r0.total]]**（[[see:Q1]]）（[[see:Q2]]）[[v:Q2.r0.n]] 笔都已核对，口径一致，"
|
||||
"（[[see:Q1]]）（[[see:Q2]]）。")
|
||||
count = "SELECT COUNT(*) AS n FROM orders WHERE is_deleted = 0"
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1"), q(name, count, "c2")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
shown = "未作废订单合计 230 3 笔都已核对，口径一致。"
|
||||
assert out.result["rationale"] == shown and out.rationale["markdown"] == shown
|
||||
[unit] = [unit for _, unit in iter_units(out.rationale)]
|
||||
assert [s["alias"] for s in unit["see"]] == ["Q1", "Q2", "Q1", "Q2"]
|
||||
numbers = [seg for seg in unit["segments"] if seg.get("ref")]
|
||||
assert [(seg["text"], bool(seg.get("strong"))) for seg in numbers] == [("230", True), ("3", False)]
|
||||
assert out.rationale["violations"] == []
|
||||
|
||||
|
||||
async def test_a_bare_see_between_two_numbers_and_other_brackets_render_clean(monkeypatch, shop):
|
||||
"""依据直接夹在两个数之间（两边什么都没有）、写进方括号（复核第 3 轮）：两个数不能挨成一个（「2303 笔」），括号不能
|
||||
剩下「［］」；两条依据照样挂在句子上。"""
|
||||
name, _ = shop
|
||||
said = "未作废订单合计 [[v:Q1.r0.total]][[see:Q1]][[v:Q2.r0.n]] 笔都已核对［[[see:Q2]]］。"
|
||||
count = "SELECT COUNT(*) AS n FROM orders WHERE is_deleted = 0"
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1"), q(name, count, "c2")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
shown = "未作废订单合计 230 3 笔都已核对。"
|
||||
assert out.result["rationale"] == shown and out.rationale["markdown"] == shown
|
||||
[unit] = [unit for _, unit in iter_units(out.rationale)]
|
||||
assert [s["alias"] for s in unit["see"]] == ["Q1", "Q2"]
|
||||
assert out.rationale["violations"] == []
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 收尾超时
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize("structured", [True, False], ids=["structured", "plain"])
|
||||
async def test_a_final_call_that_runs_out_of_time_records_the_time_limit(monkeypatch, shop, structured):
|
||||
"""模型查了一次就自己停下（没撞上任何上限），交结果那次调用却一直不返回：到点放弃，limit 是 timeout_s，和原因一致。"""
|
||||
name, _ = shop
|
||||
monkeypatch.setattr(verify_agent, "SETTLE_GRACE_S", 0.3)
|
||||
script = Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": "合计 [[v:Q1.r0.total]]。"})
|
||||
slow_final(monkeypatch)
|
||||
if not structured:
|
||||
def refuse(self, schema, **kwargs):
|
||||
raise NotImplementedError("不支持结构化输出")
|
||||
|
||||
monkeypatch.setattr(MockChatModel, "with_structured_output", refuse)
|
||||
started = time.monotonic()
|
||||
out = await run(name, timeout_s=0.5)
|
||||
assert time.monotonic() - started < 3.0, "收尾那次只等到宽限结束"
|
||||
assert script.kinds == ["step", "step"], "查了一次、自己停下，交结果那次没等到回复"
|
||||
assert out.answered is True and [qd["alias"] for qd in out.queries] == ["Q1"]
|
||||
assert out.result is None and out.error == "收尾时超过时长上限，没能交结果"
|
||||
assert out.limit == "timeout_s"
|
||||
|
||||
|
||||
async def test_a_final_call_that_runs_out_of_time_after_the_query_limit_records_the_time_limit(monkeypatch, shop):
|
||||
"""循环先撞上查询次数上限、收尾又超时：没交成结果是因为超时，limit 记 timeout_s，和原因一致——记成查询次数上限的话，
|
||||
质询会显示「未能回答（已达查询次数上限）：回答超过时长上限…」，双盲复算会说盲算查满了次数（查满了照样能交结果）。
|
||||
金额被拒的收尾照旧报循环里先撞上的那个（见 test_verify_truncation）。"""
|
||||
name, _ = shop
|
||||
monkeypatch.setattr(verify_agent, "SETTLE_GRACE_S", 0.3)
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1"), q(name, "SELECT COUNT(*) AS n FROM orders", "c2")],
|
||||
{**FOUND, "rationale": "x"})
|
||||
slow_final(monkeypatch)
|
||||
out = await run(name, timeout_s=0.5, max_queries=1)
|
||||
assert [qd["alias"] for qd in out.queries] == ["Q1"], "查满了一次"
|
||||
assert out.result is None and out.error == "收尾时超过时长上限，没能交结果"
|
||||
assert out.limit == "timeout_s"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 占位词当理由
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
@pytest.mark.parametrize("said", ["placeholder", "PLACEHOLDER", " ", "N/A", "（略）", "…", "<rationale>", "TODO", "无",
|
||||
"暂无。", "None"])
|
||||
async def test_a_placeholder_rationale_counts_as_none(monkeypatch, shop, said):
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
assert out.result["value"] == 230, "数照样取"
|
||||
assert out.result["rationale"] == "" and out.rationale is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize("said", ["[[see:Q1]]", "（[[see:Q1]]）", " [[see:Q1]]。", "placeholder。[[see:Q1]]",
|
||||
"N/A（[[see:Q1]]）", "TODO [[see:Q1]]"])
|
||||
async def test_a_rationale_that_shows_nothing_counts_as_none(monkeypatch, shop, said):
|
||||
"""只写了依据、或者占位词后面挂着依据：原文里看着有字，显示出来没有理由。以前按原文判，标记里的字（see、Q1）也算，
|
||||
理由文档照样交回——审批卡按文档画出一片空白，判「被推翻」的还因为引用到了 Q1 被当成有出处。"""
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
assert out.result["value"] == 230, "数照样取"
|
||||
assert out.result["rationale"] == "" and out.rationale is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize("said", ["无异常。", "见 [[v:Q1.r0.total]]。", "略高于上周。", "[[v:Q1.r0.total]]",
|
||||
"一致。[[see:Q1]]"])
|
||||
async def test_a_short_real_rationale_is_kept(monkeypatch, shop, said):
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": said})
|
||||
out = await run(name)
|
||||
assert out.result["rationale"] and out.rationale is not None
|
||||
|
||||
|
||||
@pytest.mark.parametrize(("said", "shown"), [("无。", "无。"), ("None.", "None."), ("暂无。[[see:Q1]]", "暂无。")])
|
||||
async def test_nothing_is_an_answer_when_the_rationale_is_the_answer(monkeypatch, shop, said, shown):
|
||||
"""结构里只有理由一项（审批质询）：理由就是回答。问「有没有测试账号」，答「无。」是一个回答，不能当成没给、让提问人
|
||||
重新提问再花一次钱。"""
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {"rationale": said})
|
||||
out = await run(name, output_schema=INQUIRY_SCHEMA)
|
||||
assert out.result["rationale"] == shown and out.rationale["markdown"] == shown
|
||||
|
||||
|
||||
@pytest.mark.parametrize("said", ["placeholder", "null", "[[see:Q1]]", "N/A。[[see:Q1]]"])
|
||||
async def test_a_placeholder_answer_still_counts_as_none(monkeypatch, shop, said):
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {"rationale": said})
|
||||
out = await run(name, output_schema=INQUIRY_SCHEMA)
|
||||
assert out.result["rationale"] == "" and out.rationale is None
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 依据写在引出语后面
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
async def test_see_after_a_lead_in_word_is_logged(monkeypatch, shop, caplog):
|
||||
"""「依据 [[see:Q1]]。」「（来源：[[see:Q1]]）」渲染后引出语悬空，渲染时去不掉（见 test_evidence_see_brackets 的已知
|
||||
限制）：记一条日志，写明是哪几处，验收时查得到。只记标记和引出语，不记数。"""
|
||||
name, _ = shop
|
||||
said = "合计 [[v:Q1.r0.total]]，依据 [[see:Q1]]。口径一致（来源：[[see:Q1]]）。"
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": said})
|
||||
with caplog.at_level(logging.WARNING, logger="app.engine.verify_agent"):
|
||||
out = await run(name, node_id="chal", run_id="run-1")
|
||||
assert out.rationale is not None, "照常交回"
|
||||
[record] = [r for r in caplog.records if "引出语" in r.getMessage()]
|
||||
message = record.getMessage()
|
||||
assert "依据 [[see:Q1]]" in message and "来源：[[see:Q1]]" in message and "run-1" in message
|
||||
assert "230" not in message
|
||||
|
||||
|
||||
async def test_see_at_the_end_of_the_sentence_is_not_logged(monkeypatch, shop, caplog):
|
||||
name, _ = shop
|
||||
Script(monkeypatch, [q(name, VALID_SUM, "c1")], {**FOUND, "rationale": "合计 [[v:Q1.r0.total]]，结果一致。[[see:Q1]]"})
|
||||
with caplog.at_level(logging.WARNING, logger="app.engine.verify_agent"):
|
||||
await run(name)
|
||||
assert not [r for r in caplog.records if "引出语" in r.getMessage()]
|
||||
@@ -0,0 +1,139 @@
|
||||
"""核查 agent 收尾超时之后，反方审稿、双盲复算对外怎么说（期 5 真实模型验收 M1 的复核）。
|
||||
|
||||
核查 agent 查完了、自己停下，交结果那次调用超过了时长上限：verify_agent 照实记 limit=timeout_s（以前模型自己停下时是
|
||||
None，test_verify_rationale_rules 钉住了这一层）。两个节点据此按时长上限说，不再把核查 agent 的原话（「收尾时超过时长
|
||||
上限，没能交结果」）摆上审批卡、时间线：
|
||||
|
||||
- 反方审稿：这句计入 challenge_limit（limit=timeout_s，claims_skipped 算上它），不再记成 challenge_unreviewed；审稿记录
|
||||
里这句的理由是「反查未完成：已达上限（每句时长上限 N 秒）」。时长按剩余时间收紧过的那组带 cause=run_deadline，理由
|
||||
说运行时长上限将至
|
||||
- 双盲复算：无法复算的原因是「盲算已达每项时长上限（N 秒）」，发 dual_check_limit；收紧过的说运行时长上限将至，不发
|
||||
dual_check_limit（它叫人调高每项的上限，调高了只会更快撞线），时间线另起一条带 cause 的 dual_check_unchecked
|
||||
|
||||
交结果那次用 slow_final 拖住，收尾宽限调到 0.3 秒；收紧过的那组把 allot() 看到的钟拨到「给下游留出余量、扣掉收尾宽限
|
||||
之后只剩 0.8 秒」（真的钟照走，核查 agent 按真的 0.8 秒到点）。模型一律是 MockChatModel 剧本，数据源是现造的 SQLite
|
||||
文件，区域、金额都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from app.core import artifact_store
|
||||
from app.engine import deadline, verify_agent
|
||||
from tests.fixtures.dual_check_shop import RIGHT, Blind
|
||||
from tests.test_challenge_node import ( # noqa: F401 - 夹具按名字引入
|
||||
Models,
|
||||
engine_up,
|
||||
finish,
|
||||
graph,
|
||||
logs,
|
||||
output_of,
|
||||
shop,
|
||||
)
|
||||
from tests.test_dual_check_node import clean_settings, dual_output, events, start # noqa: F401
|
||||
from tests.test_dual_check_node import graph as dual_graph
|
||||
from tests.test_dual_check_node import logs as dual_logs
|
||||
from tests.test_dual_check_node import shop as dual_shop # noqa: F401
|
||||
from tests.test_verify_rationale_rules import slow_final
|
||||
|
||||
#: 收紧过的那组：allot() 算出的查库时长
|
||||
ROOM = 0.8
|
||||
#: 收紧过的时长到点时，反方审稿每句的理由、双盲复算的无法复算原因（默认 600 秒的一段）
|
||||
SQUEEZED_CLAIM = "反查未完成：已达上限（本次运行即将达到 600 秒的时长上限，反查时长已按剩余时间收紧）"
|
||||
SQUEEZED_BLIND = "本次运行即将达到 600 秒的时长上限，盲算时长已按剩余时间收紧，到时未能完成"
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
def settles_slowly(monkeypatch):
|
||||
"""交结果那次调用拖 5 秒，收尾宽限只给 0.3 秒：查完了的核查一律在收尾时超时。"""
|
||||
monkeypatch.setattr(verify_agent, "SETTLE_GRACE_S", 0.3)
|
||||
slow_final(monkeypatch)
|
||||
|
||||
|
||||
def squeeze(monkeypatch) -> None:
|
||||
"""allot() 看到的「现在」：这一段的截止时刻扣掉给下游的余量（下游只有成果节点，一份）、收尾宽限，再往前 ROOM 秒。
|
||||
每开始一项都一样，所以每一项的时长都收紧到 ROOM。"""
|
||||
def now() -> float:
|
||||
found = deadline.current()
|
||||
return found.deadline - deadline.reserve(found.seconds) - verify_agent.SETTLE_GRACE_S - ROOM - 0.05
|
||||
|
||||
monkeypatch.setattr(deadline, "_now", now)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 反方审稿
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
async def test_a_claim_whose_verdict_runs_out_of_time_counts_as_hitting_the_limit(monkeypatch, shop, settles_slowly):
|
||||
models = Models(monkeypatch, shop)
|
||||
row = await finish(graph(shop, timeout_s=0.5), run_class="exploratory")
|
||||
assert row.status == "succeeded", row.error
|
||||
assert [kind for kind, _, _ in models.calls].count("final") == 0, "交结论那次都没等到回复"
|
||||
|
||||
[limit] = await logs(row.id, "challenge_limit")
|
||||
assert (limit["limit"], limit["claims_skipped"]) == ("timeout_s", 2)
|
||||
assert "cause" not in limit and "每句时长上限 0.5 秒" in limit["message"] and "2 句未能给出结论" in limit["message"]
|
||||
assert await logs(row.id, "challenge_unreviewed") == [], "触了上限的由 challenge_limit 说，不另记成未能反查"
|
||||
|
||||
out = await output_of(row.id, "chal")
|
||||
assert out["counts"]["inconclusive"] == 2 and out["unreviewed"] == 2
|
||||
record = artifact_store.load(out["challenge_artifact"])
|
||||
for claim in record["claims"]:
|
||||
assert claim["verdict"] == "inconclusive" and claim["rationale_doc"] is None
|
||||
assert claim["rationale"] == "反查未完成：已达上限（每句时长上限 0.5 秒）"
|
||||
assert len(claim["queries"]) == 1, "查到的照记"
|
||||
assert "没能交结果" not in str(record)
|
||||
assert row.output["_issuance"]["tier"] == "formal", "没给出结论的句子不让出具降档"
|
||||
|
||||
|
||||
async def test_a_squeezed_claim_whose_verdict_runs_out_of_time_says_the_run_deadline(monkeypatch, shop,
|
||||
settles_slowly):
|
||||
Models(monkeypatch, shop)
|
||||
squeeze(monkeypatch)
|
||||
row = await finish(graph(shop), run_class="exploratory")
|
||||
assert row.status == "succeeded", row.error
|
||||
|
||||
[limit] = await logs(row.id, "challenge_limit")
|
||||
assert (limit["limit"], limit["claims_skipped"], limit["claims"]) == ("timeout_s", 2, 2)
|
||||
assert limit["cause"] == deadline.RUN_DEADLINE and limit["run_seconds"] == 600
|
||||
assert "每句时长上限" not in limit["message"], "收紧过的不叫人调高每句的上限"
|
||||
assert await logs(row.id, "challenge_unreviewed") == []
|
||||
record = artifact_store.load((await output_of(row.id, "chal"))["challenge_artifact"])
|
||||
assert [c["rationale"] for c in record["claims"]] == [SQUEEZED_CLAIM, SQUEEZED_CLAIM]
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 双盲复算
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
async def test_a_blind_check_whose_result_runs_out_of_time_names_the_time_limit(monkeypatch, dual_shop,
|
||||
settles_slowly):
|
||||
name, _ = dual_shop
|
||||
blind = Blind(monkeypatch, name, {"销售额": (RIGHT, "gmv")})
|
||||
run = await start(dual_graph(name, RIGHT), options={"verify": {"dual_check": {"timeout_s": 0.5}}})
|
||||
assert run.status == "succeeded", run.error
|
||||
assert blind.runs == 0, "交结果那次没等到回复"
|
||||
evs = await events(run.id)
|
||||
assert [(d["limit"], d["metric"]) for d in dual_logs(evs, "dual_check_limit")] == [("timeout_s", "gmv")]
|
||||
out = dual_output(run, evs)
|
||||
assert out["status"] == "unchecked" and out["metrics"][0]["status"] == "unchecked"
|
||||
assert out["metrics"][0]["reason"] == "盲算已达每项时长上限（0.5 秒）"
|
||||
assert out["issuance_gaps"] == ["指标「销售额」无法复算：盲算已达每项时长上限（0.5 秒）"]
|
||||
[said] = dual_logs(evs, "dual_check_unchecked")
|
||||
assert "cause" not in said and said["metrics"][0]["reason"] == "盲算已达每项时长上限（0.5 秒）"
|
||||
assert "没能交结果" not in str(out)
|
||||
|
||||
|
||||
async def test_a_squeezed_blind_check_whose_result_runs_out_of_time_says_the_run_deadline(monkeypatch, dual_shop,
|
||||
settles_slowly):
|
||||
name, _ = dual_shop
|
||||
Blind(monkeypatch, name, {"销售额": (RIGHT, "gmv")})
|
||||
squeeze(monkeypatch)
|
||||
run = await start(dual_graph(name, RIGHT))
|
||||
assert run.status == "succeeded", run.error
|
||||
evs = await events(run.id)
|
||||
assert dual_logs(evs, "dual_check_limit") == [], "收紧过的不叫人调高每项的上限"
|
||||
out = dual_output(run, evs)
|
||||
assert out["metrics"][0]["status"] == "unchecked" and out["metrics"][0]["reason"] == SQUEEZED_BLIND
|
||||
[said] = dual_logs(evs, "dual_check_unchecked")
|
||||
assert said["cause"] == deadline.RUN_DEADLINE and said["run_seconds"] == 600
|
||||
Reference in New Issue
Block a user