起草引擎和后台任务的几处收紧:提示词要求请写的每一列都列出来(拿不准的字段留空、列不漏),免得模型留空的列再被补问;纯文本那条路的格式示例改用「表类型」「度量类型」占位,不让模型照抄示例值;从正文取 JSON 时依次试多个括号作开头,说明文字里带括号、外面多包一层都认得;用量记在单次调用的时限之外,落库慢了不会把已经回来的调用算成超时;后台任务收尾写库出错只记日志;补上从正文取 JSON、任务开始时模型已用不了(整体 failed)的测试

This commit is contained in:
ThearchyHelios
2026-10-13 15:36:24 +08:00
parent c1a70ecf38
commit b988a5ae18
5 changed files with 117 additions and 36 deletions
+4 -3
View File
@@ -1449,6 +1449,7 @@ status 状态;attribute 属性)。
规矩:
- 只写数据事实,不写计算公式,不写指标怎么算。
- 拿不准的字段留空,不要猜。列名、表名照原样写,不要翻译或改写。
- 请你写的每一列都在 columns 里写一项,拿不准的字段可以留空,但不要漏掉列:漏掉的列会被再问一次。
- 标着「系统说明」的内容是导入时按核对结果生成的,以它为准:不要重复,也不要与之矛盾。
- 标着「已确认」的目录项由人工确认过,以它为准。
- 表名下一行写着「这次只写……」的表,只写它说的那部分:只写表级的项时不写列,只写列出的几列时不写表级的项、\
@@ -1464,10 +1465,10 @@ CATALOG_DRAFT_BACKGROUND_PROMPT = "使用者补充的背景和要求:\n{lines}
#: 纯文本那条路(结构化输出用不了时):没有 JSON Schema 约束,要把格式说清
CATALOG_DRAFT_TEXT_PROMPT = """只输出一个 JSON 对象,不要输出别的文字,格式如下(拿不准的字段不写):
{"tables": [{"table": "表名", "label": "中文名", "grain": "一行代表什么", "kind": "fact", "keys": ["列名"], \
{"tables": [{"table": "表名", "label": "中文名", "grain": "一行代表什么", "kind": "表类型", "keys": ["列名"], \
"business_date": {"column": "列名", "rule": "一句规则", "timezone": "时区"}, \
"columns": [{"name": "列名", "label": "中文名", "meaning": "含义", "unit": "单位", "measure": "flow"}]}]}
kind 取 fact、dimension、snapshot、log、config 之一;measure 取 flow、stock、ratio、identifier、status、attribute 之一。"""
"columns": [{"name": "列名", "label": "中文名", "meaning": "含义", "unit": "单位", "measure": "度量类型"}]}]}
表类型写 fact、dimension、snapshot、log、config 之一;度量类型写 flow、stock、ratio、identifier、status、attribute 之一。"""
#: 拼表结构描述用的固定片段(写给模型)
CATALOG_DRAFT_PROMPT_LABELS = {
+36 -25
View File
@@ -347,15 +347,18 @@ def _array(text: str, i: int, depth: int) -> tuple[Any, int, bool]:
out.append(value)
def _json_prefix(text: str) -> tuple[Any, bool] | None:
"""正文里的第一个 JSON 值:(值, 写完了没有)。前面可以有一句话、围栏,后面可以有别的字;不是 JSON 返回 None。"""
for start in sorted(k for k in (text.find("{"), text.find("[")) if k >= 0):
try:
value, _, complete = _value(text, start, 0)
except (_Bad, RecursionError):
continue
return value, complete
return None
#: 正文里最多试几个 { 或 [ 作为 JSON 的开头(前面的说明文字里也可能有括号)
_MAX_STARTS = 16
def _starts(text: str) -> list[int]:
out: list[int] = []
for i, ch in enumerate(text):
if ch in "{[":
out.append(i)
if len(out) >= _MAX_STARTS:
break
return out
def _entry(item: Any) -> dict[str, Any] | None:
@@ -398,16 +401,22 @@ def _tables_of(root: Any) -> list[dict[str, Any]] | None:
def _parse_text(text: str, cut: bool) -> _Reply | None:
found = _json_prefix(text)
if found is None:
return None
root, complete = found
entries = _tables_of(root)
if entries is None:
return None
if complete:
return _Reply(entries=entries)
return _Reply(entries=entries, truncated=True, exhausted=cut)
"""正文里的表对象。前面可以有一句话、围栏,后面可以有别的字;JSON 没写完的(截断)只收已经写完的部分。
认不出约定的格式返回 None。"""
empty: _Reply | None = None
for start in _starts(text):
try:
root, _, complete = _value(text, start, 0)
except (_Bad, RecursionError):
continue
entries = _tables_of(root)
if entries is None:
continue
reply = _Reply(entries=entries) if complete else _Reply(entries=entries, truncated=True, exhausted=cut)
if entries:
return reply
empty = empty or reply # {} 或 {"tables": []}:后面还可能有真的那一份
return empty
def _cut_off(raw: Any) -> bool:
@@ -738,7 +747,10 @@ class _Run:
try:
async with model_gate():
await self._hook(self.hooks.call_started, [p.table for p in pieces])
return await asyncio.wait_for(self._call(pieces, mode), timeout=C.MODEL_TIMEOUT_S)
reply, raw = await asyncio.wait_for(self._call(pieces, mode), timeout=C.MODEL_TIMEOUT_S)
# 用量记在时限之外:落库慢了不能把已经回来的这次调用算成超时
await self._usage(raw)
return reply
except asyncio.CancelledError:
raise
except Exception as e: # noqa: BLE001 - 按原因分别处理
@@ -760,16 +772,15 @@ class _Run:
continue
return _Reply(error=self._call_error(e, kind, retried))
async def _call(self, pieces: list[_Piece], mode: str) -> _Reply:
async def _call(self, pieces: list[_Piece], mode: str) -> tuple[_Reply, Any]:
"""一次调用:(读好的回复, 原始回复)。原始回复给用量用(测试的假模型没有,是 None)。"""
messages = self._messages(pieces, mode)
if mode == "structured":
runnable = self.model.with_structured_output(C.CATALOG_DRAFT_PROMPT_SCHEMA, include_raw=True)
raw, parsed = _unpack(await runnable.ainvoke(messages))
await self._usage(raw)
return _structured_reply(raw, parsed)
return _structured_reply(raw, parsed), raw
raw = await self.model.ainvoke(messages)
await self._usage(raw)
return _text_reply(raw)
return _text_reply(raw), raw
async def _length_reply(self, e: BaseException) -> _Reply:
"""LengthFinishReasonError:额度用完了。异常里带着那次的回复,能捡回来的照捡。"""
+16 -8
View File
@@ -407,21 +407,29 @@ class _JobRun:
current = asyncio.current_task()
if current is not None:
current.uncancel()
await self._settle()
if self.stop_reason == "user":
await self._write(_close, self.job_id, "cancelled", None, "cancelled", None)
await self._end("cancelled", None, "cancelled", None)
else:
await self._write(_close, self.job_id, "interrupted", INTERRUPTED_ERROR, "cancelled", INTERRUPTED_ERROR)
await self._end("interrupted", INTERRUPTED_ERROR, "cancelled", INTERRUPTED_ERROR)
except _JobFailed as e:
await self._settle()
await self._write(_close, self.job_id, "failed", str(e), "failed", str(e))
await self._end("failed", str(e), "failed", str(e))
except Exception as e: # noqa: BLE001 - 任务整体出错:照实记下,不留在「起草中」
from app.core.errors import describe_exception, first_line
logger.exception("数据目录起草任务 %s 出错", self.job_id)
await self._settle()
reason = JOB_ERROR.format(line=first_line(e) or describe_exception(e))
await self._write(_close, self.job_id, "failed", reason, "failed", reason)
await self._end("failed", reason, "failed", reason)
async def _end(self, job_status: str, job_error: str | None, task_status: str, task_error: str | None) -> None:
"""停止、出错时收尾:等手上的库写入写完,再把没结束的表和任务记下来。收尾本身写不进去只记日志(下次启动时
startup 会把它标成中断)。"""
try:
await self._settle()
await self._write(_close, self.job_id, job_status, job_error, task_status, task_error)
except asyncio.CancelledError:
raise
except Exception: # noqa: BLE001
logger.exception("数据目录起草任务 %s 收尾时写库出错", self.job_id)
async def _run(self) -> None:
if not await self._write(self._claim):
@@ -617,7 +625,7 @@ class DraftJobs:
if run is not None:
await asyncio.wait([run.task], timeout=timeout) # type: ignore[list-item]
async def shutdown(self, *, grace: float = 5.0) -> None:
async def shutdown(self, *, grace: float = 3.0) -> None:
"""关停:停掉进行中的任务(标中断),最多等 grace 秒收尾。"""
runs = [run for job_id in list(self._runs) if (run := self._alive(job_id)) is not None]
for run in runs:
@@ -553,3 +553,41 @@ def test_budget_words_match_the_other_drafts():
from app.data import recipe_ai
assert D.BUDGET_WORDS == recipe_ai._BUDGET_WORDS
# --------------------------------------------------------------------------
# 从正文里取 JSON
# --------------------------------------------------------------------------
@pytest.mark.parametrize("text, tables, truncated", [
# 前面有一句话、带括号的说明、围栏
('按 {table} 的格式写好了:\n```json\n{"tables": [{"table": "a", "label": "甲"}]}\n```', ["a"], False),
('[说明] 草稿:{"tables": [{"table": "a"}]} 以上', ["a"], False),
# 外面多包了一层
('{"result": {"tables": [{"table": "a"}, {"table": "b"}]}}', ["a", "b"], False),
# 截断在字符串、数字、字面量、键名中间:只收写完了的
('{"tables": [{"table": "a", "label": "甲"}, {"table": "b", "label": "乙', ["a", "b"], True),
('{"tables": [{"table": "a", "columns": [{"name": "x", "label": "某"}, {"name": "y", "len": 12', ["a"], True),
('{"tables": [{"table": "a", "keys": ["x", "y"', ["a"], True),
('{"tables": [{"table": "a"}, {"tab', ["a"], True),
('{"tables": [{"table": "a", "flag": tr', ["a"], True),
])
def test_json_is_taken_from_the_text(text, tables, truncated):
reply = D._parse_text(text, cut=False)
assert reply is not None and [e["table"] for e in reply.entries] == tables
assert reply.truncated is truncated
def test_cut_values_are_dropped_and_complete_ones_kept():
text = '{"tables": [{"table": "a", "label": "甲", "keys": ["x", "y"], "columns": [{"name": "x", "label": "某"}, ' \
'{"name": "y", "label": "半'
entry = D._parse_text(text, cut=True).entries[0]
assert entry == {"table": "a", "label": "甲", "keys": ["x", "y"], "columns": [{"name": "x", "label": "某"}]}
cut_keys = D._parse_text('{"tables": [{"table": "a", "label": "甲", "keys": ["x", "y"', cut=True).entries[0]
assert cut_keys == {"table": "a", "label": "甲"}, "没写完的业务主键整项不要,不能只留前半"
@pytest.mark.parametrize("text", ["好的,我来写", "{坏的 json}", ""])
def test_not_json_is_not_a_reply(text):
assert D._parse_text(text, cut=False) is None
+23
View File
@@ -215,6 +215,29 @@ async def test_a_table_whose_model_part_failed_is_failed_but_structure_is_writte
assert by["channels"]["status"] == "ok"
async def test_a_job_that_cannot_start_is_failed_with_the_reason(client, monkeypatch):
"""建任务时起草模型还能用,开始起草时已经用不了(接入刚被停用):任务整体 failed,error 写原因,没开始的表也记 failed。"""
sid = await _source()
calls: list[int] = []
async def flaky(session):
calls.append(1)
if len(calls) > 1:
raise catalog.CatalogModelUnavailable("起草模型使用的模型接入「虚构接入」已停用")
return FakeModel(REPLIES), "fake-model"
monkeypatch.setattr(catalog, "resolve_draft_model", flaky)
job = (await client.post(f"{_base(sid)}/draft-jobs", json={"tables": ["visits", "gates"],
"use_model": True})).json()["job"]
await finished(job["id"])
got = await get_job(client, sid, job["id"])
assert got["status"] == "failed" and got["error"] == "起草模型使用的模型接入「虚构接入」已停用"
assert {t["status"] for t in got["tasks"]} == {"failed"} and got["finished_at"]
assert all(t["error"] == got["error"] for t in got["tasks"])
retry = await client.post(f"{_base(sid)}/draft-jobs/{job['id']}/retry")
assert retry.status_code == 409 and retry.json()["code"] == "catalog_draft_model_unavailable"
async def test_busy_source_is_409_and_the_active_job_is_reachable(client, monkeypatch):
sid = await _source()
model = Held(REPLIES, hold={"visits"})