mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 00:11:59 +00:00
起草引擎按验收修正：列分段均分、超时拆小、段大小随模型减半、每段 8 次封顶、提醒、超时与没写出来的计数、模型留空的列、不插队、原因写清哪几段哪几列
- 均分：超过 40 列的表，列分段数是列数除以 40 向上取整，各段列数平均（相差不超过 1），不再留下 2 到 8 列的尾段（实测 会思考的模型常对这种小段只给空壳）；表级那段、窄表拼段、parts_total 不变。 - 拆不开的段（不到 10 列）只给空壳：再问一次，结构化输出换纯文本，已经走纯文本的原样再问，仍是空壳才记没写出来。 - 超时当作这一段太大：不再原样重试同样大小的段，按额度用完的办法拆小再问；网络错误、429、5xx 仍原样退避重试。 - 段大小随模型调整：一次起草里额度用完或超时累计 2 次，之后新开始的段列数上限减半（40 → 20 → 10，最低 10），新开始 的段超过上限就先按上限均分开（多表的段按表重新拼）；任务记当前的上限 part_columns。 - 封顶：一段（连同它拆出来的、补问的、重试的）最多调用 8 次，到了次数还没写完的记没写出来、原因写清。止损照旧。 - 提醒：最近 6 次调用有 4 次以上没写出来（额度用完、超时、空壳、出错），任务的 notice 写一句提醒（次数、原因，建议在 「设置 → 偏好设置」的「起草模型」换一个模型、减少表或停止）；之后最近 6 次里少于 2 次就清掉。 - 用量：进度回调 usage 换成 call_done（CallReport：用量、没写出来的原因种类、当前的提醒和列数上限），任务的用量加 timeouts、failed_calls；超时的调用 token 照实记 0。 - 留空的列：请模型写了、模型也列出了，中文名和含义却都没给（或者给的都没通过校验）的列记在 ModelOutcome.blank，逐表 记 blank_columns 和最多 5 个列名 blank_sample，任务上是合计，目录清单的 last_draft 也带列数；表的状态不因此变。已确认 的、库注释已经说清的列不算。补问时把漏的列列出来、字段却留空的，算留空，不再记成「漏了这些列」。 - 公平排队：宽表的表级那段写完，它的列分段排到已经排着的段后面（先来先服务），不再插到队首让窄表一直等。 - 原因：几段因为同一个原因没写出来时说「这几段」，同一段拆出来的几块仍是「这一段」；只有真拆小过才补「已拆小重试仍 没写出」（截断后只补剩下的不算）；每段写上列的范围，例如「第 5 段（列名 到 列名，共 33 列）」，窄表写出是哪几列。 - 库：catalog_draft_jobs 加 timeouts、failed_calls、part_columns、notice，catalog_draft_tasks 加 blank_columns、 blank_sample，旧库启动时补列（存量回填 0 和空）。假模型加 hang（头几次调用一直不回，用来模拟超时）。
This commit is contained in:
ThearchyHelios
@@ -144,7 +144,8 @@ def _summary_row(name: str, meta: dict[str, Any] | None, entry: catalog.CatalogE
|
||||
"version": entry.version if entry else 0,
|
||||
"updated_at": _iso(entry.updated_at) if entry else None,
|
||||
"updated_by": entry.updated_by if entry else None,
|
||||
# 这张表最近一次起草任务里的那一项：{status, job_id, finished_at, error, model_error}，没起草过是 null
|
||||
# 这张表最近一次起草任务里的那一项：{status, job_id, finished_at, error, model_error, blank_columns}，没起草过
|
||||
# 是 null
|
||||
"last_draft": last_draft,
|
||||
}
|
||||
|
||||
|
||||
@@ -1463,12 +1463,23 @@ def draft_structure(source: Any, table: str) -> TableDraft:
|
||||
|
||||
#: 每次调用最多几张表、合计多少列（data/catalog_draft.py 按它分段）。实测每列约 50 个输出 token（不算思考），
|
||||
#: 会思考的模型思考也算在输出额度里：一段太大，想完就没额度写正文。列数超过 MODEL_BATCH_COLUMNS 的宽表拆成
|
||||
#: 「表级一段 + 列分段」，每个列分段也不超过这个列数
|
||||
#: 「表级一段 + 列分段」，列分段数是列数除以它向上取整，各段列数平均（相差不超过 1），不留很小的尾段
|
||||
MODEL_BATCH_TABLES = 4
|
||||
MODEL_BATCH_COLUMNS = 40
|
||||
#: 一次起草里额度用完或超时累计这么多次，之后新开始的段列数上限减半（MODEL_BATCH_COLUMNS → 一半 → …），
|
||||
#: 最低 MODEL_MIN_PART_COLUMNS 列
|
||||
MODEL_SHRINK_AFTER = 2
|
||||
MODEL_MIN_PART_COLUMNS = 10
|
||||
#: 一段（连同它拆出来的、补问的、重试的）最多调用几次模型，到了次数还没写完的记没写出来
|
||||
MODEL_PART_MAX_CALLS = 8
|
||||
#: 提醒：最近 MODEL_NOTICE_WINDOW 次调用里有 MODEL_NOTICE_ON 次以上没写出来时提醒换模型；之后少于 MODEL_NOTICE_OFF
|
||||
#: 次时清掉
|
||||
MODEL_NOTICE_WINDOW = 6
|
||||
MODEL_NOTICE_ON = 4
|
||||
MODEL_NOTICE_OFF = 2
|
||||
#: 全进程同时进行的模型调用最多几个（所有起草任务、同步起草共用）
|
||||
MODEL_CONCURRENCY = 3
|
||||
#: 单次调用最多等多久（秒）。会思考的模型想得久，一段 40 列常常要几分钟
|
||||
#: 单次调用最多等多久（秒）。会思考的模型想得久，一段 40 列常常要几分钟；超时了拆小再问（同额度用完）
|
||||
MODEL_TIMEOUT_S = 420
|
||||
#: 输出额度。会思考的模型（经 OpenAI 兼容网关的推理模型等）思考也算在里面：8192 时上百列的表常常想完就没额度
|
||||
#: 写正文。保留思考（不关），加大额度；服务方嫌额度太大时退回 MODEL_FALLBACK_MAX_TOKENS（同 copilot、recipe_ai）
|
||||
@@ -1477,10 +1488,10 @@ MODEL_FALLBACK_MAX_TOKENS = 8192
|
||||
#: 一段没写出来（额度用完、只给空壳）时拆成两半再问：最多拆几层、拆到最少几列为止
|
||||
MODEL_SPLIT_DEPTH = 3
|
||||
MODEL_SPLIT_MIN_COLUMNS = 5
|
||||
#: 止损：连续这么多段拆到底仍一项没写出来，这次起草后面的段不再拆小重试（一段 40 列拆到底要 15 次调用，模型明显
|
||||
#: 用不了时不该照这个倍数烧额度）
|
||||
#: 止损：连续这么多段拆到底仍一项没写出来，这次起草后面的段不再拆小重试（模型明显用不了时不该一段一段拆到底烧额度；
|
||||
#: 每段另有 MODEL_PART_MAX_CALLS 次封顶）
|
||||
MODEL_GIVE_UP_AFTER = 2
|
||||
#: 网络错误、429、5xx、超时最多重试几次，每次重试前等多久（秒，依次取用）
|
||||
#: 网络错误、429、5xx 最多重试几次，每次重试前等多久（秒，依次取用）。超时不原样重试，当作这一段太大、拆小再问
|
||||
MODEL_RETRIES = 2
|
||||
MODEL_RETRY_DELAYS_S = (10.0, 30.0)
|
||||
#: 不指定表时起草使用次数最多的前几张
|
||||
@@ -1794,6 +1805,8 @@ class TableDraftResult:
|
||||
status: str = "ok"
|
||||
#: 模型分几段起草（不用模型时为 None）
|
||||
parts_total: int | None = None
|
||||
#: 模型拿不准、留空了的列（catalog_draft.ModelOutcome.blank，按表结构的顺序）；表的状态不因此变
|
||||
blank: list[str] = field(default_factory=list)
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -1837,6 +1850,7 @@ async def write_draft(session: AsyncSession, source: Any, table: str, modeled: A
|
||||
result.model_error = modeled.error
|
||||
if modeled.items:
|
||||
result.model_items = modeled.items
|
||||
result.blank = list(getattr(modeled, "blank", None) or [])
|
||||
model_draft = TableDraft(notes=modeled.notes, covered=frozenset({"llm"}) if modeled.status == "ok"
|
||||
else frozenset())
|
||||
for _attempt in range(3):
|
||||
@@ -2298,8 +2312,10 @@ __all__ = [
|
||||
"CatalogEntry", "CatalogInvalid", "CatalogModelUnavailable", "CatalogPathError", "DRAFT_DEFAULT_TABLES",
|
||||
"DraftModel", "DraftReport", "INFERRED_MARK", "ITEM_SOURCES", "ITEM_STATUSES", "JoinEdge", "KIND_LABEL",
|
||||
"MEASURE_LABEL", "MODEL_BATCH_COLUMNS", "MODEL_BATCH_TABLES", "MODEL_CONCURRENCY", "MODEL_EMPTY_REASON",
|
||||
"MODEL_FALLBACK_MAX_TOKENS", "MODEL_MAX_TOKENS", "MODEL_RETRIES", "MODEL_RETRY_DELAYS_S", "MODEL_SPLIT_DEPTH",
|
||||
"MODEL_SPLIT_MIN_COLUMNS", "MODEL_TIMEOUT_S", "MergeStats", "NEW_RELATION_PATH", "PATCH_MAX_CHANGES",
|
||||
"MODEL_FALLBACK_MAX_TOKENS", "MODEL_MAX_TOKENS", "MODEL_MIN_PART_COLUMNS", "MODEL_NOTICE_OFF", "MODEL_NOTICE_ON",
|
||||
"MODEL_NOTICE_WINDOW", "MODEL_PART_MAX_CALLS", "MODEL_RETRIES", "MODEL_RETRY_DELAYS_S", "MODEL_SHRINK_AFTER",
|
||||
"MODEL_SPLIT_DEPTH", "MODEL_SPLIT_MIN_COLUMNS", "MODEL_TIMEOUT_S", "MergeStats", "NEW_RELATION_PATH",
|
||||
"PATCH_MAX_CHANGES",
|
||||
"PROFILE_NOTE_PREFIX", "PROFILE_VERIFY_COVERAGE", "WRITE_CONFLICT_ERROR", "default_tables",
|
||||
"missing_table_error", "pick_tables", "write_draft",
|
||||
"PatchChange", "PatchPlan", "QUERY_SNAPSHOT_KIND", "REVIEW_ACTIONS", "TABLE_FIELDS", "TABLE_FIELD_LABEL",
|
||||
|
||||
+358
-110
@@ -9,15 +9,21 @@
|
||||
|
||||
**分段。** 每次调用最多 MODEL_BATCH_TABLES 张表、合计 MODEL_BATCH_COLUMNS 列：列数不超过它的表按顺序拼在一起
|
||||
（整表段）；超过它的宽表拆成「表级一段 + 列分段」——表级那段给全部列名和类型做背景、只要表级的项（中文名、粒度、
|
||||
表类型、业务主键、业务日期）；列分段每段 MODEL_BATCH_COLUMNS 列，带着表名、表注释和表级起草结果做背景，只要这几列。
|
||||
最后合并成这张表的一份，照旧经 catalog._from_model 逐项校验（起草要求不能让模型绕过校验）。
|
||||
表类型、业务主键、业务日期）；列分段数是列数除以 MODEL_BATCH_COLUMNS 向上取整、各段列数平均（相差不超过 1，以前
|
||||
按 40 列定长切，留下几列的尾段，实测会思考的模型常对这种小段只给空壳），带着表名、表注释和表级起草结果做背景，只要
|
||||
这几列。最后合并成这张表的一份，照旧经 catalog._from_model 逐项校验（起草要求不能让模型绕过校验）。
|
||||
|
||||
**段大小随模型调整。** 一次起草里额度用完或超时累计 MODEL_SHRINK_AFTER 次，之后新开始的段列数上限减半（40 → 20
|
||||
→ 10，最低 MODEL_MIN_PART_COLUMNS）：新开始的段（整表段、列分段）超过上限就先按上限均分开再问。已经开始的段照旧
|
||||
按失败原因拆小。
|
||||
|
||||
**按失败原因处理，不原样重发。**
|
||||
|
||||
- 思考加输出用完额度（finish_reason=length、LengthFinishReasonError、正文为空而有思考）、只给空壳：把这一段拆成
|
||||
两半再问（多表的段按表拆，单表的按列拆），最小 MODEL_SPLIT_MIN_COLUMNS 列，最多拆 MODEL_SPLIT_DEPTH 层。
|
||||
拆不开的空壳（结构化输出给的）换纯文本问一次：有的网关的结构化输出只给空壳、纯文本却能写——写出来了，这次起草
|
||||
后面都走纯文本。
|
||||
- 思考加输出用完额度（finish_reason=length、LengthFinishReasonError、正文为空而有思考）、只给空壳、超时（当作这一段
|
||||
太大，不原样重试同样大小的段）：把这一段拆成两半再问（多表的段按表拆，单表的按列拆），最小
|
||||
MODEL_SPLIT_MIN_COLUMNS 列，最多拆 MODEL_SPLIT_DEPTH 层。拆不开的空壳再问一次：结构化输出给的换纯文本（有的网关的
|
||||
结构化输出只给空壳、纯文本却能写——写出来了，这次起草后面都走纯文本），已经走纯文本的原样再问；仍是空壳才记没写
|
||||
出来。
|
||||
- 结构化输出回了一段说明文字（不是 JSON、也没停在额度上）：网关接受了结构化输出的参数却没照做（例如开了丢弃不认识
|
||||
的参数），提示词里又没有格式说明，模型就写成了说明文字。这次起草后面改走纯文本，这一段换纯文本问一次。有思考
|
||||
不等于额度用完：只有停在额度上、或者正文为空而有思考才算。
|
||||
@@ -27,20 +33,30 @@
|
||||
- 结构化输出这个接口不支持（400 / 422 提到 response_format、json_schema、tool……，或者模型类不支持）：这次起草
|
||||
后面改走纯文本、从正文取 JSON（纯文本那条路另外说清 JSON 的格式）。
|
||||
- 服务方嫌额度太大：退回 MODEL_FALLBACK_MAX_TOKENS 重来，这次起草后面都用它（catalog.DraftModel.shrink）。
|
||||
- 网络错误、429、5xx、超时：等一会儿重试（最多 MODEL_RETRIES 次，退避）。单次调用的时限是 MODEL_TIMEOUT_S。SDK
|
||||
- 网络错误、429、5xx：等一会儿重试（最多 MODEL_RETRIES 次，退避）。单次调用的时限是 MODEL_TIMEOUT_S。SDK
|
||||
自己不再重试（catalog.resolve_draft_model 建模型时 max_retries=0），实际重试次数和报错写的一致。
|
||||
- 别的错误：这一段没写出来，原因照实写。
|
||||
- 封顶：一段（连同它拆出来的、补问的、重试的）最多调用 MODEL_PART_MAX_CALLS 次，到了次数还没写完的记没写出来。
|
||||
- 止损：连续 MODEL_GIVE_UP_AFTER 段拆到底仍一项没写出来（模型明显用不了，例如思考总是吃满额度），这次起草后面的
|
||||
段不再拆小重试、没写出来就直接记下，报错建议更换起草模型；有一段写出来了就恢复。不然一段拆到底要几十次调用，
|
||||
起草几十张宽表会烧掉数百万 token、跑几个小时。
|
||||
段不再拆小重试、没写出来就直接记下，报错建议更换起草模型；有一段写出来了就恢复。
|
||||
- 提醒：最近 MODEL_NOTICE_WINDOW 次调用里有 MODEL_NOTICE_ON 次以上没写出来（额度用完、超时、空壳、出错），给一句
|
||||
提醒（次数和原因，建议换起草模型、减少表或停止），经进度回调交给任务；之后少于 MODEL_NOTICE_OFF 次时清掉。以前
|
||||
看不出调用在反复失败：一个任务 55 分钟只结束 1 张表，界面上只有段进度和用量。
|
||||
|
||||
**原因。** 没写出来的原因只写进这张表的 model_error：几段因为同一个原因没写出来时说「这几段」，每段写上列的范围
|
||||
（「第 5 段（列名 到 列名，共 33 列）」，列名照表结构原样），真拆小重试过的才补「已拆小重试仍没写出」。
|
||||
|
||||
**留空的列。** 请模型写了、模型也列出了，中文名和含义却都没给（或者给的都没通过校验）的列是模型拿不准，不补问，
|
||||
记在 ModelOutcome.blank 里（表的状态不因此变）。已确认的、库注释已经说清的列不算（提示词里它们以此为准）。
|
||||
|
||||
**并发。** 全进程最多 MODEL_CONCURRENCY 个调用同时进行（model_gate：按事件循环各一个信号量，所有起草任务和同步
|
||||
起草共用，先来先放）。一次起草里同时只起草 MODEL_CONCURRENCY 段（一段拆小、补缺都做完再起草下一段），同时在等
|
||||
名额的调用也最多这么多个：不把几百段一下子都排上去，几次起草（不同数据源的任务、同步起草）的调用交替排队，后开始
|
||||
的不会等前一个排着的段全部调完。
|
||||
的不会等前一个排着的段全部调完。一次起草里按表的先后排队：宽表的表级那段写完，它的列分段排到已经排着的段后面
|
||||
（先来先服务），不插到窄表前面。
|
||||
|
||||
**进度。** 调用方给 Hooks：开始调用、每次调用的用量、一段结束、一张表的模型部分全部结束。后台起草任务据此落库，
|
||||
每张表一结束就并入目录（data/catalog_jobs.py）。
|
||||
**进度。** 调用方给 Hooks：开始调用、每次调用的结果（用量、没写出来的原因、当前的提醒和列数上限）、一段结束、一张表
|
||||
的模型部分全部结束。后台起草任务据此落库，每张表一结束就并入目录（data/catalog_jobs.py）。
|
||||
|
||||
交给模型的只有表结构（列、类型、主键、外键、唯一约束、注释）和使用者补充的背景，不发任何数据值。写给模型的文字都在
|
||||
catalog.py 名字带 _SYSTEM / _PROMPT 的常量里；这里的 REASON_* 是给人看的（拼进这张表的 model_error）。
|
||||
@@ -59,6 +75,7 @@ from dataclasses import dataclass, field
|
||||
from typing import Any
|
||||
|
||||
from app.data import catalog as C
|
||||
from app.data.draft_model import DRAFT_MODEL_PLACE
|
||||
from app.data.names import name_key
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
@@ -75,21 +92,33 @@ STRUCTURED_WORDS = ("response_format", "json_schema", "json_object", "tool", "fu
|
||||
#: 一张表的表级的项在「没写出来」记录里的键（不会和列名撞上）
|
||||
_TABLE = "\0table"
|
||||
|
||||
# ---- 一段没写出来的原因（给人看，拼进这张表的 model_error）
|
||||
REASON_EXHAUSTED = "模型的思考和输出用完了额度（{budget} 个 token），这一段没写出来"
|
||||
# ---- 一段没写出来的原因（给人看，拼进这张表的 model_error）。{seg} 在拼成一句时换成「这一段」或「这几段」：同一个
|
||||
#: 原因落在几段上（宽表的几个列分段）说「这几段」，同一段拆出来的几块仍是「这一段」
|
||||
SEG = "{seg}"
|
||||
SEG_ONE = "这一段"
|
||||
SEG_MANY = "这几段"
|
||||
REASON_EXHAUSTED = "模型的思考和输出用完了额度（{budget} 个 token），{seg}没写出来"
|
||||
REASON_SPLIT = "；已拆小重试仍没写出"
|
||||
REASON_HOLLOW = "模型只给了空壳，没有写出可用的内容"
|
||||
REASON_TRUNCATED = "模型的输出被截断，这一段没写完"
|
||||
REASON_TRUNCATED = "模型的输出被截断，{seg}没写完"
|
||||
REASON_FORMAT = "模型的回复不是约定的格式"
|
||||
REASON_MISSING = "模型的回复里没有这张表"
|
||||
REASON_UNANSWERED = "模型的回复里漏了这些列，补问一次仍没有给出"
|
||||
REASON_TIMEOUT = "模型 {seconds} 秒内没有写完这一段"
|
||||
REASON_TIMEOUT = "模型 {seconds} 秒内没有写完{seg}"
|
||||
REASON_CALL = "调用模型出错：{line}"
|
||||
REASON_RETRIED = "（已重试 {n} 次）"
|
||||
REASON_CRASH = "起草这张表时出错：{line}"
|
||||
#: 一段到了调用次数的上限（MODEL_PART_MAX_CALLS）还没写完
|
||||
REASON_CAPPED = "每段最多调用 {n} 次模型，{seg}到了次数仍没写完，已停止重试"
|
||||
#: 止损（连续几段拆到底仍一项没写出来，后面的段不再拆小重试）时补在原因后面
|
||||
REASON_HALTED = "；模型连续 {n} 段都没写出来，已停止拆小重试，建议更换起草模型"
|
||||
|
||||
# ---- 提醒（最近几次调用多数没写出来），经进度回调交给任务（CatalogDraftJob.notice）
|
||||
NOTICE = ("最近 {calls} 次调用有 {failed} 次没写出来（{kinds}）。可以在{place}换一个模型、减少这次起草的表，"
|
||||
"或者停止这次起草")
|
||||
#: 一次调用没写出来的原因种类 → 提醒里的说法（按这个顺序列）
|
||||
FAILURE_LABEL = {"exhausted": "额度用完", "timeout": "超时", "hollow": "只给空壳", "error": "出错"}
|
||||
|
||||
|
||||
# ==========================================================================
|
||||
# 对外的形状
|
||||
@@ -111,7 +140,8 @@ class ModelOutcome:
|
||||
|
||||
status：ok 各段都写出来了；partial 有的段写出来、有的没有（error 写清哪部分、为什么）；failed 一项都没写出来
|
||||
（error 写原因）。notes 是写出来的那些（已逐项校验，llm / proposed）；items 是项数。parts_total 分几段起草
|
||||
（整表段是 1，宽表是 1 + 列分段数），parts_done 已经结束了几段。"""
|
||||
（整表段是 1，宽表是 1 + 列分段数），parts_done 已经结束了几段。blank：模型拿不准、留空了的列（按表结构的顺序；
|
||||
表的状态不因此变）。"""
|
||||
|
||||
status: str
|
||||
notes: dict[str, Any]
|
||||
@@ -119,6 +149,26 @@ class ModelOutcome:
|
||||
error: str | None
|
||||
parts_total: int
|
||||
parts_done: int
|
||||
blank: list[str] = field(default_factory=list)
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class CallReport:
|
||||
"""一次调用的结果（进度回调 call_done 收到的）。
|
||||
|
||||
failure：没写出来的原因种类，写出来了是 None——exhausted 额度用完（一项都没捡回来）、timeout 超时、hollow 只给空壳
|
||||
（回复里没有可用的内容）、error 出错（会重试的那次也算一次）。超时的调用拿不到用量，token 记 0（网关那边可能已经
|
||||
生成了）。notice、part_columns 是这次调用之后的：提醒（没有为 None）、新开始的段每段最多几列。"""
|
||||
|
||||
input_tokens: int
|
||||
output_tokens: int
|
||||
failure: str | None
|
||||
notice: str | None
|
||||
part_columns: int
|
||||
|
||||
@property
|
||||
def timeout(self) -> bool:
|
||||
return self.failure == "timeout"
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -126,22 +176,37 @@ class Hooks:
|
||||
"""进度回调，都是 async，出错只记日志、不影响起草。
|
||||
|
||||
- call_started(表名列表)：一次调用拿到并发名额、马上要发出去了；
|
||||
- usage(输入 token, 输出 token)：每次调用一回（拿不到用量的是 0），条数就是调用次数；
|
||||
- call_done(CallReport)：每次调用一回（条数就是调用次数），带着用量、没写出来的原因、当前的提醒和列数上限；
|
||||
- part_done(表名)：这张表的一段结束了（写没写出来都算）；
|
||||
- table_done(表名, ModelOutcome)：这张表的模型部分全部结束。"""
|
||||
|
||||
call_started: Callable[[list[str]], Awaitable[None]] | None = None
|
||||
usage: Callable[[int, int], Awaitable[None]] | None = None
|
||||
call_done: Callable[[CallReport], Awaitable[None]] | None = None
|
||||
part_done: Callable[[str], Awaitable[None]] | None = None
|
||||
table_done: Callable[[str, ModelOutcome], Awaitable[None]] | None = None
|
||||
|
||||
|
||||
def parts_for(columns: int) -> int:
|
||||
"""一张有这么多列的表分几段起草：不超过 MODEL_BATCH_COLUMNS 列是 1 段，否则是表级一段加列分段。"""
|
||||
"""一张有这么多列的表分几段起草：不超过 MODEL_BATCH_COLUMNS 列是 1 段，否则是表级一段加列分段（列数除以
|
||||
MODEL_BATCH_COLUMNS 向上取整）。"""
|
||||
width = max(1, int(C.MODEL_BATCH_COLUMNS))
|
||||
return 1 if columns <= width else 1 + -(-columns // width)
|
||||
|
||||
|
||||
def _even(columns: Sequence[str], parts: int) -> list[tuple[str, ...]]:
|
||||
"""按顺序切成 parts 段，各段列数平均（相差不超过 1，多出来的放在前面几段）。"""
|
||||
columns = tuple(columns)
|
||||
parts = max(1, min(int(parts), len(columns) or 1))
|
||||
size, extra = divmod(len(columns), parts)
|
||||
out: list[tuple[str, ...]] = []
|
||||
start = 0
|
||||
for i in range(parts):
|
||||
n = size + (1 if i < extra else 0)
|
||||
out.append(columns[start:start + n])
|
||||
start += n
|
||||
return out
|
||||
|
||||
|
||||
async def run_model_draft(model: Any, source: Any, tables: Sequence[str], *,
|
||||
existing: Mapping[str, Mapping[str, Any]] | None = None,
|
||||
background: Background | None = None,
|
||||
@@ -237,25 +302,36 @@ class _Reply:
|
||||
error: str | None = None
|
||||
#: 结构化输出回了一段说明文字（error 是「不是约定的格式」）：换纯文本还可能写出来
|
||||
prose: bool = False
|
||||
#: 这次调用超时了（当作这一段太大：拆小再问，不原样重试）
|
||||
timeout: bool = False
|
||||
#: 这一段的调用次数用完了，这次没有调用
|
||||
capped: bool = False
|
||||
#: 这次调用的用量 (输入, 输出)，还没报给进度回调；没有调用（capped）的是 None
|
||||
usage: tuple[int, int] | None = None
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class _Why:
|
||||
"""一部分没写出来的原因。base 给人看的原因；split 拆小重试过；halted 本来还能拆小，因为前面连续几段都没写出来
|
||||
（止损）才没拆。"""
|
||||
"""一部分没写出来的原因。base 给人看的原因（可以带 SEG，拼成一句时换成「这一段」「这几段」）；split 真拆小重试过；
|
||||
halted 本来还能拆小，因为前面连续几段都没写出来（止损）才没拆。"""
|
||||
|
||||
base: str
|
||||
split: bool = False
|
||||
halted: bool = False
|
||||
|
||||
|
||||
def _join(whys: Iterable[_Why]) -> str:
|
||||
"""几处没写出来的原因拼成一句：同样的原因只说一次（最多说两种），拆小重试过、止损的说明各在最后补一次（只看说出来
|
||||
的那几种）。不然宽表的表级那段拆不开、列分段拆过时，同一句原因会说两遍。"""
|
||||
whys = list(whys)
|
||||
shown = list(dict.fromkeys(w.base for w in whys))[:2]
|
||||
told = [w for w in whys if w.base in shown]
|
||||
text = "；".join(shown)
|
||||
def _join(failures: Iterable[tuple[_Why, int]]) -> str:
|
||||
"""几处没写出来的原因拼成一句。failures 是 (原因, 属于这张表的第几段)。
|
||||
|
||||
同样的原因只说一次（最多说两种）；同一个原因落在几段上说「这几段」，同一段拆出来的几块仍是「这一段」；拆小重试过、
|
||||
止损的说明各在最后补一次（只看说出来的那几种）。不然宽表的表级那段拆不开、列分段拆过时，同一句原因会说两遍。"""
|
||||
failures = list(failures)
|
||||
parts: dict[str, set[int]] = {}
|
||||
for why, part in failures:
|
||||
parts.setdefault(why.base, set()).add(part)
|
||||
shown = list(parts)[:2]
|
||||
told = [w for w, _ in failures if w.base in shown]
|
||||
text = "；".join(base.replace(SEG, SEG_MANY if len(parts[base]) > 1 else SEG_ONE) for base in shown)
|
||||
if any(w.split for w in told):
|
||||
text += REASON_SPLIT
|
||||
if any(w.halted for w in told):
|
||||
@@ -263,6 +339,22 @@ def _join(whys: Iterable[_Why]) -> str:
|
||||
return text
|
||||
|
||||
|
||||
def _columns_text(order: Sequence[str], columns: Iterable[str], *, count: bool) -> str:
|
||||
"""说清是哪几列（列名照表结构原样，按表结构的顺序）：一到三列逐个写；更多的连着就写「头 到 尾」，不连着写前三个加
|
||||
「等」。count：后面带上列数（「，共 33 列」）。"""
|
||||
pos = {name: i for i, name in enumerate(order)}
|
||||
cols = sorted(dict.fromkeys(columns), key=lambda c: pos.get(c, len(pos)))
|
||||
n = len(cols)
|
||||
if n <= 3:
|
||||
text = "、".join(cols)
|
||||
elif pos.get(cols[-1], -1) - pos.get(cols[0], 0) + 1 == n:
|
||||
text = f"{cols[0]} 到 {cols[-1]}"
|
||||
else:
|
||||
head = "、".join(cols[:3])
|
||||
return f"{head} 等 {n} 列" if count else f"{head} 等"
|
||||
return f"{text}，共 {n} 列" if count else text
|
||||
|
||||
|
||||
class _Cut:
|
||||
"""被截断的容器（对象或数组）：value 里只有已经完整的成员，最后一个没写完的容器成员也包成 _Cut 留着，好往里捡。"""
|
||||
|
||||
@@ -541,6 +633,12 @@ def _text_reply(raw: Any) -> _Reply:
|
||||
return _Reply(exhausted=cut or _thought(raw))
|
||||
|
||||
|
||||
def _usage_of(raw: Any) -> tuple[int, int]:
|
||||
"""一次调用的用量 (输入, 输出)：拿不到的是 0（测试的假模型没有原始回复）。"""
|
||||
meta = getattr(raw, "usage_metadata", None) or {}
|
||||
return int(meta.get("input_tokens") or 0), int(meta.get("output_tokens") or 0)
|
||||
|
||||
|
||||
def _unpack(out: Any) -> tuple[Any, Any]:
|
||||
"""with_structured_output(include_raw=True) 交回 {raw, parsed, parsing_error}；测试的假模型直接交回解析好的对象。"""
|
||||
if isinstance(out, dict) and "raw" in out and ("parsed" in out or "parsing_error" in out):
|
||||
@@ -602,6 +700,12 @@ class _Table:
|
||||
"""这一段请写的部分拿到了内容。"""
|
||||
return (piece.table_level and self.got_table) or any(c in self.got_columns for c in piece.columns)
|
||||
|
||||
def part_of(self) -> dict[str, int]:
|
||||
"""每一列（表级的项是 _TABLE）属于这张表的第几段。"""
|
||||
out = {col: piece.part for piece in self.parts for col in piece.columns}
|
||||
out[_TABLE] = 1
|
||||
return out
|
||||
|
||||
|
||||
def _restrict(notes: Mapping[str, Any], piece: _Piece) -> dict[str, Any]:
|
||||
"""只留这一段请模型写的部分：没请写表级的项就不要表级的项，列只要请写的那几列。"""
|
||||
@@ -624,6 +728,27 @@ class _Work:
|
||||
then: list[_Work] = field(default_factory=list)
|
||||
|
||||
|
||||
@dataclass
|
||||
class _Budget:
|
||||
"""一段（新开始的一段，连同它拆出来的、补问的、重试的）还能调用几次模型。"""
|
||||
|
||||
left: int
|
||||
|
||||
|
||||
def _pack(pieces: list[_Piece], cap: int) -> list[list[_Piece]]:
|
||||
"""多表的段按顺序重新拼成每段不超过 cap 列（超过 cap 的表自成一段）。"""
|
||||
groups: list[list[_Piece]] = []
|
||||
used = 0
|
||||
for piece in pieces:
|
||||
n = len(piece.columns)
|
||||
if not groups or used + n > cap:
|
||||
groups.append([])
|
||||
used = 0
|
||||
groups[-1].append(piece)
|
||||
used += n
|
||||
return groups
|
||||
|
||||
|
||||
def _halves(pieces: list[_Piece]) -> tuple[list[_Piece], list[_Piece]] | None:
|
||||
"""拆成两半：多表的段按表拆；单表的按列拆，两半都至少 MODEL_SPLIT_MIN_COLUMNS 列（表级的项跟着前一半）。拆不开返回 None。"""
|
||||
if len(pieces) > 1:
|
||||
@@ -659,6 +784,12 @@ class _Run:
|
||||
self.gate = asyncio.Semaphore(max(1, int(C.MODEL_CONCURRENCY)))
|
||||
#: 连续几段拆到底仍一项没写出来（止损用，有一段写出来就清零）
|
||||
self.dead_streak = 0
|
||||
#: 新开始的段每段最多几列：额度用完或超时（strikes）累计 MODEL_SHRINK_AFTER 次减半一次
|
||||
self.part_columns = max(1, int(C.MODEL_BATCH_COLUMNS))
|
||||
self.strikes = 0
|
||||
#: 最近几次调用没写出来的原因种类（写出来了是 None），提醒看它
|
||||
self.recent: collections.deque[str | None] = collections.deque(maxlen=max(1, int(C.MODEL_NOTICE_WINDOW)))
|
||||
self.notice: str | None = None
|
||||
|
||||
@property
|
||||
def halted(self) -> bool:
|
||||
@@ -669,8 +800,9 @@ class _Run:
|
||||
|
||||
async def run(self) -> dict[str, ModelOutcome]:
|
||||
"""按计划的先后起草：同时只起草 MODEL_CONCURRENCY 段，一段（连同拆小、补缺）做完再取下一段；宽表的表级那段
|
||||
写完，它的列分段排到最前面。不一下子把所有段（包括宽表的全部列分段）都排上去：那样后开始的另一次起草要等这次
|
||||
排着的段全部调完；拆小重试也排在所有段的第一次调用后面，一段要等几乎所有调用都做完才结束，止损形同虚设。"""
|
||||
写完，它的列分段排到队尾（先来先服务：排在已经排着的窄表段后面，不插队）。不一下子把所有段（包括宽表的全部列
|
||||
分段）都排上去：那样后开始的另一次起草要等这次排着的段全部调完；拆小重试也排在所有段的第一次调用后面，一段要
|
||||
等几乎所有调用都做完才结束，止损形同虚设。"""
|
||||
queue: collections.deque[_Work] = collections.deque()
|
||||
for unit in self._plan():
|
||||
if isinstance(unit, list):
|
||||
@@ -704,7 +836,8 @@ class _Run:
|
||||
return {name: self._outcome(t) for name, t in self.tables.items()}
|
||||
|
||||
def _plan(self) -> list[list[str] | str]:
|
||||
"""分段：窄表按顺序拼成整表段（列表），宽表一个一项（表名）。各项按第一张表在 tables 里的位置排。"""
|
||||
"""分段：窄表按顺序拼成整表段（列表），宽表一个一项（表名），列分段均分（_even）。各项按第一张表在 tables 里的
|
||||
位置排。"""
|
||||
width = max(1, int(C.MODEL_BATCH_COLUMNS))
|
||||
most = max(1, int(C.MODEL_BATCH_TABLES))
|
||||
units: list[list[str] | str] = []
|
||||
@@ -715,8 +848,7 @@ class _Run:
|
||||
n = len(st.columns)
|
||||
if n > width:
|
||||
st.parts = [_Piece(name, (), True, 1)] + [
|
||||
_Piece(name, tuple(st.columns[i:i + width]), False, k + 2)
|
||||
for k, i in enumerate(range(0, n, width))]
|
||||
_Piece(name, cols, False, k + 2) for k, cols in enumerate(_even(st.columns, -(-n // width)))]
|
||||
units.append(name)
|
||||
continue
|
||||
st.parts = [_Piece(name, tuple(st.columns), True, 1)]
|
||||
@@ -728,20 +860,40 @@ class _Run:
|
||||
return units
|
||||
|
||||
async def _work(self, work: _Work, queue: collections.deque[_Work], left: dict[str, int]) -> None:
|
||||
"""起草一段；它之后的段（宽表的列分段）排到最前面；一张表的段都结束了，这张表的模型部分就结束了。"""
|
||||
await self._guarded(work.pieces)
|
||||
"""起草一段；它之后的段（宽表的列分段）排到队尾；一张表的段都结束了，这张表的模型部分就结束了。"""
|
||||
await self._start(work.pieces)
|
||||
for piece in work.pieces:
|
||||
await self._part_done(piece.table)
|
||||
queue.extendleft(reversed(work.then))
|
||||
queue.extend(work.then)
|
||||
for piece in work.pieces:
|
||||
left[piece.table] -= 1
|
||||
if not left[piece.table]:
|
||||
await self._table_done(piece.table)
|
||||
|
||||
async def _start(self, pieces: list[_Piece]) -> None:
|
||||
"""开始起草一段。超过当前的列数上限（额度用完、超时多了会减半）的先按上限分开：多表的段按表重新拼，单张表的
|
||||
按列均分（表级的项跟着第一块，先写它，后面几块拿它做背景）。分开的每一块各算一段（各有调用次数的上限）。"""
|
||||
cap = max(1, self.part_columns)
|
||||
if sum(len(p.columns) for p in pieces) <= cap:
|
||||
await self._guarded(pieces)
|
||||
return
|
||||
if len(pieces) > 1:
|
||||
await asyncio.gather(*(self._start(group) for group in _pack(pieces, cap)))
|
||||
return
|
||||
piece = pieces[0]
|
||||
chunks = _even(piece.columns, -(-len(piece.columns) // cap))
|
||||
first = [_Piece(piece.table, chunks[0], piece.table_level, piece.part)]
|
||||
rest = [[_Piece(piece.table, cols, False, piece.part)] for cols in chunks[1:]]
|
||||
if piece.table_level:
|
||||
await self._guarded(first)
|
||||
await asyncio.gather(*(self._guarded(group) for group in rest))
|
||||
else:
|
||||
await asyncio.gather(*(self._guarded(group) for group in [first, *rest]))
|
||||
|
||||
async def _guarded(self, pieces: list[_Piece]) -> None:
|
||||
"""起草一段；意料之外的错误只记这一段没写出来，不影响别的段和别的表。"""
|
||||
try:
|
||||
await self._segment(pieces, 0, False)
|
||||
await self._segment(pieces, 0, False, _Budget(max(1, int(C.MODEL_PART_MAX_CALLS))))
|
||||
except asyncio.CancelledError:
|
||||
raise
|
||||
except Exception as e: # noqa: BLE001
|
||||
@@ -762,111 +914,132 @@ class _Run:
|
||||
|
||||
# ---------------------------------------------------------------- 一段
|
||||
|
||||
async def _segment(self, pieces: list[_Piece], depth: int, followup: bool, *, mode: str | None = None,
|
||||
last: bool = False) -> None:
|
||||
async def _segment(self, pieces: list[_Piece], depth: int, followup: bool, budget: _Budget, *,
|
||||
halved: bool = False, mode: str | None = None, last: bool = False) -> None:
|
||||
"""问一段，拿到的并进各表；按失败原因拆开、补缺、记下没写出来的。
|
||||
|
||||
depth：已经拆了几层（截断后只补剩下的也算一层）。followup：这一段是在补缺（补过就不再补，见 _after_followup）。
|
||||
budget：这一段还能调用几次。halved：这几块是真拆小（_split）出来的——原因里「已拆小重试」只在这时写。
|
||||
mode / last：换纯文本问的那一次——拆不开的空壳、结构化输出回了说明文字（再不行就记没写出来）。"""
|
||||
reply = await self._ask(pieces, mode)
|
||||
reply = await self._ask(pieces, budget, mode)
|
||||
if reply.capped:
|
||||
self._fail(pieces, self._capped(halved))
|
||||
return
|
||||
if reply.timeout:
|
||||
await self._report(reply, wrote=False)
|
||||
await self._split(pieces, depth, "timeout", followup, budget, halved)
|
||||
return
|
||||
if reply.error is not None:
|
||||
await self._report(reply, wrote=False)
|
||||
if reply.prose and not last:
|
||||
await self._prose_to_text(pieces, depth, followup)
|
||||
await self._prose_to_text(pieces, depth, followup, budget, halved)
|
||||
return
|
||||
self._fail(pieces, reply.error)
|
||||
self._fail(pieces, _Why(reply.error, split=halved))
|
||||
return
|
||||
gains = self._take(pieces, reply.entries)
|
||||
anything = any(g.items for g in gains)
|
||||
await self._report(reply, wrote=anything)
|
||||
if last and anything and self.mode != "text":
|
||||
# 结构化输出只给空壳、纯文本写出来了：这个模型的结构化输出靠不住，这次起草后面都走纯文本
|
||||
self.mode = "text"
|
||||
logger.info("起草数据目录：结构化输出只给了空壳，纯文本写出来了，后面改走纯文本")
|
||||
if reply.exhausted or reply.truncated:
|
||||
rest = self._remainder(pieces, gains, truncated=True)
|
||||
rest = self._remainder(pieces, gains, truncated=True, followup=followup)
|
||||
if not rest:
|
||||
return
|
||||
kind = "exhausted" if reply.exhausted else "truncated"
|
||||
if not anything:
|
||||
await self._split(rest, depth, kind, followup)
|
||||
await self._split(rest, depth, kind, followup, budget, halved)
|
||||
elif depth < int(C.MODEL_SPLIT_DEPTH):
|
||||
await self._segment(rest, depth + 1, followup) # 捡回了一部分：只补剩下的
|
||||
# 捡回了一部分：只补剩下的（不是拆小）
|
||||
await self._segment(rest, depth + 1, followup, budget, halved=halved)
|
||||
else:
|
||||
self._fail(rest, self._reason(kind, split=True))
|
||||
self._fail(rest, self._reason(kind, split=halved))
|
||||
return
|
||||
if not anything and not followup:
|
||||
if last:
|
||||
halted = self._stopped(pieces, depth)
|
||||
self._fail(pieces, lambda p: self._reason("hollow", split=depth > 0, piece=p, halted=halted))
|
||||
self._fail(pieces, lambda p: self._reason("hollow", split=halved, piece=p, halted=halted))
|
||||
return
|
||||
await self._split(pieces, depth, "hollow", followup)
|
||||
await self._split(pieces, depth, "hollow", followup, budget, halved)
|
||||
return
|
||||
rest = self._remainder(pieces, gains, truncated=False)
|
||||
rest = self._remainder(pieces, gains, truncated=False, followup=followup)
|
||||
if not rest:
|
||||
return
|
||||
if followup:
|
||||
await self._after_followup(rest, depth, last=last)
|
||||
await self._after_followup(rest, depth, budget, halved, last=last)
|
||||
return
|
||||
await self._segment(rest, depth, True)
|
||||
await self._segment(rest, depth, True, budget, halved=halved)
|
||||
|
||||
async def _after_followup(self, rest: list[_Piece], depth: int, *, last: bool) -> None:
|
||||
async def _after_followup(self, rest: list[_Piece], depth: int, budget: _Budget, halved: bool, *,
|
||||
last: bool) -> None:
|
||||
"""补问过仍然没给的部分（补过就不再补）。
|
||||
|
||||
- 这张表已经拿到了内容：没列出来的列记没写出来（这张表是 partial，旧的模型起草项不删）。模型列出了、字段留空
|
||||
的列是拿不准，不在这里（_remainder 只把没列出来的列交过来）；
|
||||
- 这张表到现在还一项都没有（回复里没有这张表、只给了空壳）：照第一次只给空壳的规矩，拆小再问，拆不开换纯文本
|
||||
问一次。拆出来的半段是新的一段（还能补缺一次），层数照算。"""
|
||||
- 这张表已经拿到了内容：补问时仍没列出来的列记没写出来（这张表是 partial）。补问时列出了、字段留空的列是拿不准
|
||||
（记留空），不在这里（_remainder 只把没列出来的列交过来）；
|
||||
- 这张表到现在还一项都没有（回复里没有这张表、只给了空壳）：照第一次只给空壳的规矩，拆小再问，拆不开再问一次。
|
||||
拆出来的半段是新的一段（还能补缺一次），层数照算。"""
|
||||
given = [p for p in rest if self.tables[p.table].has_any()]
|
||||
empty = [p for p in rest if not self.tables[p.table].has_any()]
|
||||
if given:
|
||||
self._fail(given, _Why(REASON_UNANSWERED))
|
||||
self._fail(given, _Why(REASON_UNANSWERED, split=halved))
|
||||
if not empty:
|
||||
return
|
||||
if last:
|
||||
halted = self._stopped(empty, depth)
|
||||
self._fail(empty, lambda p: self._reason("hollow", split=depth > 0, piece=p, halted=halted))
|
||||
self._fail(empty, lambda p: self._reason("hollow", split=halved, piece=p, halted=halted))
|
||||
return
|
||||
await self._split(empty, depth, "hollow", False)
|
||||
await self._split(empty, depth, "hollow", False, budget, halved)
|
||||
|
||||
async def _prose_to_text(self, pieces: list[_Piece], depth: int, followup: bool) -> None:
|
||||
async def _prose_to_text(self, pieces: list[_Piece], depth: int, followup: bool, budget: _Budget,
|
||||
halved: bool) -> None:
|
||||
"""结构化输出回了一段说明文字：网关接受了结构化输出的参数却没照做。这次起草后面都走纯文本（另带格式说明），
|
||||
这一段换纯文本问一次。纯文本额度用完了还能照常拆小（拆出来的半段也走纯文本，不再先问一遍结构化输出）。"""
|
||||
if self.mode != "text":
|
||||
self.mode = "text"
|
||||
logger.info("起草数据目录：结构化输出回了说明文字，后面改走纯文本")
|
||||
await self._segment(pieces, depth, followup, mode="text", last=True)
|
||||
await self._segment(pieces, depth, followup, budget, halved=halved, mode="text", last=True)
|
||||
|
||||
def _stopped(self, pieces: list[_Piece], depth: int) -> bool:
|
||||
"""这几段本来还能拆小，因为前面连续几段都没写出来（止损）才没拆。"""
|
||||
return self.halted and depth < int(C.MODEL_SPLIT_DEPTH) and _halves(pieces) is not None
|
||||
|
||||
async def _split(self, pieces: list[_Piece], depth: int, kind: str, followup: bool) -> None:
|
||||
async def _split(self, pieces: list[_Piece], depth: int, kind: str, followup: bool, budget: _Budget,
|
||||
halved: bool) -> None:
|
||||
"""拆成两半再问（额度用完、超时、只给空壳、截断后一项都没捡回来）。拆不开的空壳再问一次：结构化输出给的换纯文本，
|
||||
已经走纯文本的原样再问；别的原因拆不开就记没写出来。"""
|
||||
halves = _halves(pieces) if depth < int(C.MODEL_SPLIT_DEPTH) else None
|
||||
halted = halves is not None and self.halted
|
||||
if halves is not None and not halted:
|
||||
first, second = halves
|
||||
if (len(first) == len(second) == 1 and first[0].table == second[0].table and first[0].table_level):
|
||||
# 同一张表按列拆、前一半带着表级的项：先写前一半，后一半拿它做背景
|
||||
await self._segment(first, depth + 1, followup)
|
||||
await self._segment(second, depth + 1, followup)
|
||||
await self._segment(first, depth + 1, followup, budget, halved=True)
|
||||
await self._segment(second, depth + 1, followup, budget, halved=True)
|
||||
return
|
||||
results = await asyncio.gather(self._segment(first, depth + 1, followup),
|
||||
self._segment(second, depth + 1, followup), return_exceptions=True)
|
||||
results = await asyncio.gather(self._segment(first, depth + 1, followup, budget, halved=True),
|
||||
self._segment(second, depth + 1, followup, budget, halved=True),
|
||||
return_exceptions=True)
|
||||
for result in results:
|
||||
if isinstance(result, BaseException):
|
||||
raise result
|
||||
return
|
||||
if kind == "hollow" and self.mode == "structured":
|
||||
await self._segment(pieces, depth, followup, mode="text", last=True)
|
||||
if kind == "hollow":
|
||||
await self._segment(pieces, depth, followup, budget, halved=halved, mode="text", last=True)
|
||||
return
|
||||
self._fail(pieces, lambda p: self._reason(kind, split=depth > 0, piece=p, halted=halted))
|
||||
self._fail(pieces, lambda p: self._reason(kind, split=halved, piece=p, halted=halted))
|
||||
|
||||
def _remainder(self, pieces: list[_Piece], gains: list[_Gain], *, truncated: bool) -> list[_Piece]:
|
||||
def _remainder(self, pieces: list[_Piece], gains: list[_Gain], *, truncated: bool,
|
||||
followup: bool) -> list[_Piece]:
|
||||
"""还没写出来的部分：回复里没有这张表、或者这张表一项可用内容都没有的整段重问；有内容的只问没列出来的列
|
||||
（列出了、字段留空的算模型拿不准，不再问）。截断的回复里表级的项也没拿到的，表级的项一起重问（没截断时表级
|
||||
空着算模型拿不准：这张表在回复里，表级的字段留空和列的字段留空一样）。"""
|
||||
空着算模型拿不准：这张表在回复里，表级的字段留空和列的字段留空一样）。
|
||||
|
||||
补问（followup）时这张表已经有内容了：补问的列列出来了就算答过（字段留空是拿不准），只交回没列出来的列。"""
|
||||
out: list[_Piece] = []
|
||||
for piece, gain in zip(pieces, gains):
|
||||
if not gain.named or not gain.items:
|
||||
answered_blank = followup and gain.named and self.tables[piece.table].has_any()
|
||||
if not gain.named or (not gain.items and not answered_blank):
|
||||
out.append(piece)
|
||||
continue
|
||||
cols = tuple(c for c in piece.columns if c not in gain.answered)
|
||||
@@ -877,40 +1050,49 @@ class _Run:
|
||||
|
||||
# ---------------------------------------------------------------- 调用
|
||||
|
||||
async def _ask(self, pieces: list[_Piece], mode: str | None = None) -> _Reply:
|
||||
"""调一次模型（按原因重来：额度被拒、结构化输出不支持、网络错误……），交回读好的回复。"""
|
||||
async def _ask(self, pieces: list[_Piece], budget: _Budget, mode: str | None = None) -> _Reply:
|
||||
"""调一次模型（按原因重来：额度被拒、结构化输出不支持、网络错误……），交回读好的回复。
|
||||
|
||||
这一段的调用次数用完了交回 capped（不调用）。超时不原样重试：交回 timeout，由调用方当作这一段太大、拆小再问。
|
||||
重来之前的那次调用在这里就报给进度回调；交回的那次由调用方读完内容再报（写没写出来要看校验之后的内容）。"""
|
||||
mode = mode or self.mode
|
||||
retried = 0
|
||||
while True:
|
||||
if budget.left <= 0:
|
||||
return _Reply(capped=True)
|
||||
budget.left -= 1
|
||||
used = getattr(self.model, "budget", None)
|
||||
try:
|
||||
# 先占这次起草自己的名额，再去排全进程的：一次起草排在全进程名额上的调用最多 MODEL_CONCURRENCY 个
|
||||
async with self.gate, model_gate():
|
||||
await self._hook(self.hooks.call_started, [p.table for p in pieces])
|
||||
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)
|
||||
reply.usage = _usage_of(raw)
|
||||
return reply
|
||||
except asyncio.CancelledError:
|
||||
raise
|
||||
except Exception as e: # noqa: BLE001 - 按原因分别处理
|
||||
kind = classify_failure(e)
|
||||
if kind == "length":
|
||||
return await self._length_reply(e)
|
||||
await self._hook(self.hooks.usage, 0, 0)
|
||||
return self._length_reply(e)
|
||||
if kind == "timeout":
|
||||
return _Reply(timeout=True, usage=(0, 0))
|
||||
if kind == "budget" and (getattr(self.model, "budget", None) != used or self._shrink()):
|
||||
await self._called((0, 0), "error")
|
||||
continue
|
||||
if kind == "unsupported" and mode == "structured":
|
||||
if self.mode == "structured":
|
||||
logger.info("起草数据目录：结构化输出用不了（%s），后面改走纯文本", str(e)[:200])
|
||||
self.mode = mode = "text"
|
||||
await self._called((0, 0), "error")
|
||||
continue
|
||||
if kind in ("transient", "timeout") and retried < int(C.MODEL_RETRIES):
|
||||
if kind == "transient" and retried < int(C.MODEL_RETRIES):
|
||||
await self._called((0, 0), "error")
|
||||
delays = tuple(C.MODEL_RETRY_DELAYS_S) or (0.0,)
|
||||
await _sleep(delays[min(retried, len(delays) - 1)])
|
||||
retried += 1
|
||||
continue
|
||||
return _Reply(error=self._call_error(e, kind, retried))
|
||||
return _Reply(error=self._call_error(e, retried), usage=(0, 0))
|
||||
|
||||
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）。"""
|
||||
@@ -922,12 +1104,12 @@ class _Run:
|
||||
raw = await self.model.ainvoke(messages)
|
||||
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。"""
|
||||
@staticmethod
|
||||
def _length_reply(e: BaseException) -> _Reply:
|
||||
"""LengthFinishReasonError：额度用完了。异常里带着那次的回复（和用量），能捡回来的照捡。"""
|
||||
completion = getattr(e, "completion", None)
|
||||
usage = getattr(completion, "usage", None)
|
||||
await self._hook(self.hooks.usage, int(getattr(usage, "prompt_tokens", 0) or 0),
|
||||
int(getattr(usage, "completion_tokens", 0) or 0))
|
||||
tokens = (int(getattr(usage, "prompt_tokens", 0) or 0), int(getattr(usage, "completion_tokens", 0) or 0))
|
||||
try:
|
||||
text = completion.choices[0].message.content or "" # type: ignore[union-attr]
|
||||
except (AttributeError, IndexError, TypeError):
|
||||
@@ -935,8 +1117,8 @@ class _Run:
|
||||
if text.strip():
|
||||
reply = _parse_text(text, True)
|
||||
if reply is not None and reply.entries:
|
||||
return _Reply(entries=reply.entries, truncated=True, exhausted=True)
|
||||
return _Reply(exhausted=True)
|
||||
return _Reply(entries=reply.entries, truncated=True, exhausted=True, usage=tokens)
|
||||
return _Reply(exhausted=True, usage=tokens)
|
||||
|
||||
def _shrink(self) -> bool:
|
||||
shrink = getattr(self.model, "shrink", None)
|
||||
@@ -948,17 +1130,50 @@ class _Run:
|
||||
def _budget(self) -> int:
|
||||
return int(getattr(self.model, "budget", None) or C.MODEL_MAX_TOKENS)
|
||||
|
||||
def _call_error(self, e: BaseException, kind: str, retried: int) -> str:
|
||||
@staticmethod
|
||||
def _call_error(e: BaseException, retried: int) -> str:
|
||||
from app.core.errors import describe_exception, first_line
|
||||
|
||||
tail = REASON_RETRIED.format(n=retried) if retried else ""
|
||||
if kind == "timeout":
|
||||
return REASON_TIMEOUT.format(seconds=_seconds(C.MODEL_TIMEOUT_S)) + tail
|
||||
return REASON_CALL.format(line=first_line(e) or describe_exception(e)) + tail
|
||||
|
||||
async def _usage(self, raw: Any) -> None:
|
||||
meta = getattr(raw, "usage_metadata", None) or {}
|
||||
await self._hook(self.hooks.usage, int(meta.get("input_tokens") or 0), int(meta.get("output_tokens") or 0))
|
||||
# ---------------------------------------------------------------- 每次调用之后：用量、段大小、提醒
|
||||
|
||||
async def _report(self, reply: _Reply, *, wrote: bool) -> None:
|
||||
"""交回的那次调用读完内容之后报给进度回调。"""
|
||||
if reply.usage is None:
|
||||
return
|
||||
if reply.timeout:
|
||||
failure: str | None = "timeout"
|
||||
elif reply.error is not None:
|
||||
failure = "error"
|
||||
elif wrote:
|
||||
failure = None
|
||||
else:
|
||||
failure = "exhausted" if reply.exhausted else "hollow"
|
||||
await self._called(reply.usage, failure, strike=reply.timeout or reply.exhausted)
|
||||
|
||||
async def _called(self, usage: tuple[int, int], failure: str | None, *, strike: bool = False) -> None:
|
||||
"""记下一次调用：额度用完或超时（strike）累计 MODEL_SHRINK_AFTER 次，新开始的段列数上限减半；最近几次调用多数
|
||||
没写出来时提醒，少了清掉。报给进度回调。"""
|
||||
self.recent.append(failure)
|
||||
if strike:
|
||||
self.strikes += 1
|
||||
width = max(1, int(C.MODEL_BATCH_COLUMNS))
|
||||
floor = min(width, max(1, int(C.MODEL_MIN_PART_COLUMNS)))
|
||||
cap = max(floor, width >> (self.strikes // max(1, int(C.MODEL_SHRINK_AFTER))))
|
||||
if cap < self.part_columns:
|
||||
self.part_columns = cap
|
||||
logger.info("起草数据目录：额度用完或超时已有 %d 次，之后新开始的段每段最多 %d 列", self.strikes, cap)
|
||||
failed = [f for f in self.recent if f is not None]
|
||||
if len(failed) >= int(C.MODEL_NOTICE_ON) or (self.notice is not None and len(failed) >= int(C.MODEL_NOTICE_OFF)):
|
||||
counts = collections.Counter(failed)
|
||||
kinds = "、".join(f"{label} {counts[kind]} 次" for kind, label in FAILURE_LABEL.items() if counts[kind])
|
||||
self.notice = NOTICE.format(calls=len(self.recent), failed=len(failed), kinds=kinds, place=DRAFT_MODEL_PLACE)
|
||||
else:
|
||||
self.notice = None
|
||||
await self._hook(self.hooks.call_done, CallReport(int(usage[0]), int(usage[1]), failure, self.notice,
|
||||
self.part_columns))
|
||||
|
||||
# ---------------------------------------------------------------- 拿到的内容
|
||||
|
||||
@@ -1028,9 +1243,11 @@ class _Run:
|
||||
|
||||
def _reason(self, kind: str, *, split: bool, piece: _Piece | None = None, halted: bool = False) -> _Why:
|
||||
if kind == "exhausted":
|
||||
base = REASON_EXHAUSTED.format(budget=self._budget())
|
||||
base = REASON_EXHAUSTED.format(budget=self._budget(), seg=SEG)
|
||||
elif kind == "truncated":
|
||||
base = REASON_TRUNCATED
|
||||
base = REASON_TRUNCATED.format(seg=SEG)
|
||||
elif kind == "timeout":
|
||||
base = REASON_TIMEOUT.format(seconds=_seconds(C.MODEL_TIMEOUT_S), seg=SEG)
|
||||
else:
|
||||
st = self.tables[piece.table] if piece is not None else None
|
||||
if st is not None and st.whole(piece):
|
||||
@@ -1039,6 +1256,10 @@ class _Run:
|
||||
base = REASON_HOLLOW
|
||||
return _Why(base, split=split, halted=halted)
|
||||
|
||||
@staticmethod
|
||||
def _capped(halved: bool) -> _Why:
|
||||
return _Why(REASON_CAPPED.format(n=max(1, int(C.MODEL_PART_MAX_CALLS)), seg=SEG), split=halved)
|
||||
|
||||
# ---------------------------------------------------------------- 收尾
|
||||
|
||||
def _outcome(self, st: _Table) -> ModelOutcome:
|
||||
@@ -1046,31 +1267,57 @@ class _Run:
|
||||
failed_cols = [c for c in st.columns if c in st.failed and c not in st.got_columns]
|
||||
failed_table = _TABLE in st.failed and not st.got_table
|
||||
total = len(st.parts) or 1
|
||||
part_of = st.part_of()
|
||||
if not items:
|
||||
whys = list(st.failed.values())
|
||||
reason = _join(whys) if whys else (C.MODEL_EMPTY_REASON if st.named else REASON_MISSING)
|
||||
failures = [(why, part_of.get(key, 1)) for key, why in st.failed.items()]
|
||||
reason = _join(failures) if failures else (C.MODEL_EMPTY_REASON if st.named else REASON_MISSING)
|
||||
return ModelOutcome("failed", {}, 0, reason, total, st.parts_done)
|
||||
notes = copy.deepcopy(st.notes)
|
||||
blank = self._blank(st, set(failed_cols))
|
||||
if failed_cols or failed_table:
|
||||
return ModelOutcome("partial", notes, items, self._partial_text(st, failed_cols, failed_table), total,
|
||||
st.parts_done)
|
||||
return ModelOutcome("ok", notes, items, None, total, st.parts_done)
|
||||
st.parts_done, blank)
|
||||
return ModelOutcome("ok", notes, items, None, total, st.parts_done, blank)
|
||||
|
||||
def _blank(self, st: _Table, failed: set[str]) -> list[str]:
|
||||
"""模型拿不准、留空了的列：请它写了（不在没写出来的里面），中文名和含义却都没给（或者给的都没通过校验）。已确认
|
||||
的、库注释已经说清的不算——提示词里它们以此为准，模型不写也不缺：已确认（有确证）的中文名或含义、库注释（按
|
||||
catalog.comment_role 记进目录的那种；导入表格的源是系统说明）。"""
|
||||
drafted = st.notes.get("columns") or {}
|
||||
confirmed = C._confirmed_only(st.existing).get("columns") or {}
|
||||
comments = {str(c.get("name")): (c.get("comment") or "").strip()
|
||||
for c in st.meta.get("columns") or [] if isinstance(c, dict)}
|
||||
out: list[str] = []
|
||||
for col in st.columns:
|
||||
if col in failed:
|
||||
continue
|
||||
if any(f in (drafted.get(col) or {}) or f in (confirmed.get(col) or {}) for f in ("label", "meaning")):
|
||||
continue
|
||||
text = comments.get(col)
|
||||
if text and (self.system or C.comment_role(text, col) is not None):
|
||||
continue
|
||||
out.append(col)
|
||||
return out
|
||||
|
||||
@staticmethod
|
||||
def _partial_text(st: _Table, failed_cols: list[str], failed_table: bool) -> str:
|
||||
"""只写出一部分时说清哪部分没写出来、为什么、怎么办。"""
|
||||
"""只写出一部分时说清哪部分没写出来（第几段、哪几列）、为什么、怎么办。"""
|
||||
part_of = st.part_of()
|
||||
keys = ([_TABLE] if failed_table else []) + failed_cols
|
||||
reason = _join(st.failed[k] for k in keys)
|
||||
reason = _join((st.failed[k], part_of.get(k, 1)) for k in keys)
|
||||
n = len(st.columns)
|
||||
if len(st.parts) > 1:
|
||||
part_of = {c: p.part for p in st.parts for c in p.columns}
|
||||
bad = sorted({part_of[c] for c in failed_cols if c in part_of} | ({1} if failed_table else set()))
|
||||
head = f"这张表 {n} 列，分 {len(st.parts)} 段起草，第 {'、'.join(str(b) for b in bad)} 段没写出来"
|
||||
head += f"（共 {len(failed_cols)} 列）" if failed_cols else "（表级的项）"
|
||||
bad: dict[int, list[str]] = {}
|
||||
for col in failed_cols:
|
||||
bad.setdefault(part_of.get(col, 1), []).append(col)
|
||||
said = ["第 1 段（表级的项）"] if failed_table else []
|
||||
said += [f"第 {k} 段（{_columns_text(st.columns, cols, count=True)}）" for k, cols in sorted(bad.items())]
|
||||
head = f"这张表 {n} 列，分 {len(st.parts)} 段起草，{'、'.join(said)}没写出来"
|
||||
elif failed_cols and failed_table:
|
||||
head = f"这张表 {n} 列，有表级的项和 {len(failed_cols)} 列没写出来"
|
||||
head = (f"这张表 {n} 列，有表级的项和 {len(failed_cols)} 列（{_columns_text(st.columns, failed_cols, count=False)}）"
|
||||
"没写出来")
|
||||
elif failed_cols:
|
||||
head = f"这张表 {n} 列，有 {len(failed_cols)} 列没写出来"
|
||||
head = f"这张表 {n} 列，有 {len(failed_cols)} 列（{_columns_text(st.columns, failed_cols, count=False)}）没写出来"
|
||||
else:
|
||||
head = f"这张表 {n} 列，表级的项没写出来"
|
||||
return f"{head}：{reason}。可以只重试这张表"
|
||||
@@ -1196,6 +1443,7 @@ class _Run:
|
||||
return C._notes_lines(confirmed, meta=st.meta, system_notes=False)
|
||||
|
||||
|
||||
__all__ = ["BUDGET_WORDS", "Background", "Hooks", "ModelOutcome", "REASON_CALL", "REASON_EXHAUSTED", "REASON_FORMAT",
|
||||
"REASON_HALTED", "REASON_HOLLOW", "REASON_MISSING", "REASON_RETRIED", "REASON_SPLIT", "REASON_TIMEOUT",
|
||||
"REASON_TRUNCATED", "REASON_UNANSWERED", "classify_failure", "model_gate", "parts_for", "run_model_draft"]
|
||||
__all__ = ["BUDGET_WORDS", "Background", "CallReport", "FAILURE_LABEL", "Hooks", "ModelOutcome", "NOTICE", "REASON_CALL",
|
||||
"REASON_CAPPED", "REASON_EXHAUSTED", "REASON_FORMAT", "REASON_HALTED", "REASON_HOLLOW", "REASON_MISSING",
|
||||
"REASON_RETRIED", "REASON_SPLIT", "REASON_TIMEOUT", "REASON_TRUNCATED", "REASON_UNANSWERED", "SEG",
|
||||
"classify_failure", "model_gate", "parts_for", "run_model_draft"]
|
||||
|
||||
@@ -6,8 +6,9 @@
|
||||
- **建任务**（create_job）：同一个数据源同时只有一个排队或进行中的任务（DraftBusy；部分唯一索引兜底）。不指定表时
|
||||
按使用次数取前 20 张；表结构里没有的表照样建进任务、状态 skipped。起草要求（数据源级）和本次说明随任务记下快照。
|
||||
- **在后台起草**（manager.start）：进程内的 asyncio 任务，引擎是 data/catalog_draft.py（分段、按失败原因处理）。
|
||||
每张表一结束就并入目录（catalog.write_draft，规则和同步起草一样：人工确认、驳回过的不动，乐观锁冲突重读重并）；
|
||||
逐表的状态、到第几段、计数、用量随时落库，接口读的是库里的状态。
|
||||
每张表一结束就并入目录（catalog.write_draft，规则和同步起草一样：人工确认、驳回过的不动，模型这次没给的项不删
|
||||
以前起草的，乐观锁冲突重读重并）；逐表的状态、到第几段、计数、模型留空的列，任务的用量（含超时、没写出来的调用
|
||||
次数）、当前的列数上限、提醒随时落库，接口读的是库里的状态。
|
||||
- **停止**（cancel）：排队的直接 cancelled；进行中的停掉正在进行的模型调用，没开始和没写完的表记 cancelled，已经写进
|
||||
目录的不回滚，任务最后是 cancelled。
|
||||
- **重试**（retry_tables + create_job）：新任务，只含原任务里 failed、partial、cancelled 的表；用不用模型、本次说明沿用，
|
||||
@@ -55,6 +56,8 @@ GUIDANCE_MAX = 2000
|
||||
NOTE_MAX = 1000
|
||||
#: 目录清单上「这张表最近一次起草」只看最近这么多个任务
|
||||
LAST_DRAFT_JOBS = 100
|
||||
#: 逐表最多记几个模型留空的列名
|
||||
BLANK_SAMPLE = 5
|
||||
#: 停止进行中的任务时最多等它收尾多久（秒）；关停服务时另给（shutdown 的 grace）
|
||||
STOP_GRACE_S = 10.0
|
||||
|
||||
@@ -164,6 +167,9 @@ def task_dict(task: CatalogDraftTask) -> dict[str, Any]:
|
||||
"removed": task.removed,
|
||||
"version": task.version,
|
||||
"model_items": task.model_items,
|
||||
# 模型拿不准、留空了的列（中文名和含义都没给）：几列、最多 5 个列名（按表结构的顺序）。表的状态不因此变
|
||||
"blank_columns": task.blank_columns or 0,
|
||||
"blank_sample": list(task.blank_sample or []),
|
||||
"error": task.error,
|
||||
"model_error": task.model_error,
|
||||
"started_at": _iso(task.started_at),
|
||||
@@ -200,8 +206,14 @@ def job_dict(job: CatalogDraftJob, tasks: list[CatalogDraftTask]) -> dict[str, A
|
||||
"added": sum(t.added or 0 for t in tasks),
|
||||
"updated": sum(t.updated or 0 for t in tasks),
|
||||
"removed": sum(t.removed or 0 for t in tasks),
|
||||
"blank_columns": sum(t.blank_columns or 0 for t in tasks),
|
||||
# 超时的调用拿不到用量，token 按 0 算（网关那边可能已经生成了）；failed_calls 含额度用完、只给空壳、超时、出错
|
||||
"usage": {"calls": job.calls or 0, "input_tokens": job.input_tokens or 0,
|
||||
"output_tokens": job.output_tokens or 0},
|
||||
"output_tokens": job.output_tokens or 0, "timeouts": job.timeouts or 0,
|
||||
"failed_calls": job.failed_calls or 0},
|
||||
# 新开始的段每段最多几列（不用模型、还没开始时为 null）；提醒：最近几次调用多数没写出来时的一句话，没有为 null
|
||||
"part_columns": job.part_columns,
|
||||
"notice": job.notice,
|
||||
"error": job.error,
|
||||
"tasks": [task_dict(t) for t in tasks],
|
||||
}
|
||||
@@ -225,20 +237,22 @@ async def job_brief(session: AsyncSession, job: CatalogDraftJob | None) -> dict[
|
||||
|
||||
|
||||
async def last_drafts(session: AsyncSession, source_id: str) -> dict[str, dict[str, Any]]:
|
||||
"""每张表最近一次起草任务里的那一项：{表名: {status, job_id, finished_at, error, model_error}}。任务还在进行，
|
||||
就是进行中的状态。只看最近 LAST_DRAFT_JOBS 个任务；每张表取哪一项在库里挑好（清单页起草时会反复取）。"""
|
||||
"""每张表最近一次起草任务里的那一项：{表名: {status, job_id, finished_at, error, model_error, blank_columns}}。
|
||||
任务还在进行，就是进行中的状态。只看最近 LAST_DRAFT_JOBS 个任务；每张表取哪一项在库里挑好（清单页起草时会反复取）。"""
|
||||
recent = (select(CatalogDraftJob.id).where(CatalogDraftJob.source_id == source_id)
|
||||
.order_by(CatalogDraftJob.created_at.desc()).limit(LAST_DRAFT_JOBS)).scalar_subquery()
|
||||
rank = func.row_number().over(partition_by=CatalogDraftTask.table_name,
|
||||
order_by=(CatalogDraftJob.created_at.desc(), CatalogDraftTask.seq)).label("rank")
|
||||
inner = (select(CatalogDraftTask.table_name, CatalogDraftTask.status, CatalogDraftTask.job_id,
|
||||
CatalogDraftTask.finished_at, CatalogDraftTask.error, CatalogDraftTask.model_error, rank)
|
||||
CatalogDraftTask.finished_at, CatalogDraftTask.error, CatalogDraftTask.model_error,
|
||||
CatalogDraftTask.blank_columns, rank)
|
||||
.join(CatalogDraftJob, CatalogDraftJob.id == CatalogDraftTask.job_id)
|
||||
.where(CatalogDraftJob.id.in_(recent))).subquery()
|
||||
rows = (await session.execute(select(inner).where(inner.c.rank == 1))).mappings()
|
||||
return {row["table_name"]: {"status": row["status"], "job_id": row["job_id"],
|
||||
"finished_at": _iso(row["finished_at"]), "error": row["error"],
|
||||
"model_error": row["model_error"]} for row in rows}
|
||||
"model_error": row["model_error"], "blank_columns": row["blank_columns"] or 0}
|
||||
for row in rows}
|
||||
|
||||
|
||||
# ==========================================================================
|
||||
@@ -463,7 +477,7 @@ class _JobRun:
|
||||
await self._write(self._skip, gone)
|
||||
if use_model and names:
|
||||
await self._write(self._set_model, model_id)
|
||||
hooks = catalog_draft.Hooks(call_started=self._call_started, usage=self._usage,
|
||||
hooks = catalog_draft.Hooks(call_started=self._call_started, call_done=self._call_done,
|
||||
part_done=self._part_done, table_done=self._table_done)
|
||||
await catalog_draft.run_model_draft(model, source, names, existing=existing, background=background,
|
||||
hooks=hooks)
|
||||
@@ -482,8 +496,8 @@ class _JobRun:
|
||||
self._started.update(fresh)
|
||||
await self._write(self._mark_running, fresh)
|
||||
|
||||
async def _usage(self, input_tokens: int, output_tokens: int) -> None:
|
||||
await self._write(self._add_usage, input_tokens, output_tokens)
|
||||
async def _call_done(self, report: catalog_draft.CallReport) -> None:
|
||||
await self._write(self._add_call, report)
|
||||
|
||||
async def _part_done(self, table: str) -> None:
|
||||
await self._write(self._bump_part, table)
|
||||
@@ -506,7 +520,8 @@ class _JobRun:
|
||||
|
||||
async def _set_model(self, session: AsyncSession, model_id: str) -> None:
|
||||
await session.execute(update(CatalogDraftJob).where(CatalogDraftJob.id == self.job_id)
|
||||
.values(model=model_id or None).execution_options(synchronize_session=False))
|
||||
.values(model=model_id or None, part_columns=max(1, int(catalog.MODEL_BATCH_COLUMNS)))
|
||||
.execution_options(synchronize_session=False))
|
||||
await session.commit()
|
||||
|
||||
async def _skip(self, session: AsyncSession, names: list[str]) -> None:
|
||||
@@ -521,11 +536,16 @@ class _JobRun:
|
||||
.values(status="running", started_at=utcnow()))
|
||||
await session.commit()
|
||||
|
||||
async def _add_usage(self, session: AsyncSession, input_tokens: int, output_tokens: int) -> None:
|
||||
async def _add_call(self, session: AsyncSession, report: catalog_draft.CallReport) -> None:
|
||||
"""一次调用：用量、超时、没写出来的次数累加，当前的列数上限和提醒照这次调用之后的写（按调用的先后写进来）。"""
|
||||
await session.execute(
|
||||
update(CatalogDraftJob).where(CatalogDraftJob.id == self.job_id)
|
||||
.values(calls=CatalogDraftJob.calls + 1, input_tokens=CatalogDraftJob.input_tokens + int(input_tokens),
|
||||
output_tokens=CatalogDraftJob.output_tokens + int(output_tokens))
|
||||
.values(calls=CatalogDraftJob.calls + 1,
|
||||
input_tokens=CatalogDraftJob.input_tokens + int(report.input_tokens),
|
||||
output_tokens=CatalogDraftJob.output_tokens + int(report.output_tokens),
|
||||
timeouts=func.coalesce(CatalogDraftJob.timeouts, 0) + int(report.timeout),
|
||||
failed_calls=func.coalesce(CatalogDraftJob.failed_calls, 0) + int(report.failure is not None),
|
||||
part_columns=report.part_columns, notice=report.notice)
|
||||
.execution_options(synchronize_session=False))
|
||||
await session.commit()
|
||||
|
||||
@@ -551,6 +571,7 @@ class _JobRun:
|
||||
"status": result.status, "added": result.added, "updated": result.updated, "removed": result.removed,
|
||||
"version": result.version, "model_items": result.model_items, "error": result.error,
|
||||
"model_error": result.model_error, "finished_at": utcnow(),
|
||||
"blank_columns": len(result.blank), "blank_sample": result.blank[:BLANK_SAMPLE] or None,
|
||||
}
|
||||
if outcome is not None:
|
||||
values["parts_done"] = outcome.parts_done
|
||||
@@ -638,8 +659,9 @@ class DraftJobs:
|
||||
manager = DraftJobs()
|
||||
|
||||
|
||||
__all__ = ["ACTIVE", "DraftBusy", "DraftFinished", "DraftJobs", "FINISHED", "GUIDANCE_MAX", "INTERRUPTED_ERROR",
|
||||
"MAX_TABLES", "NOTE_MAX", "NothingToRetry", "RETRYABLE", "TASK_DONE", "TASK_STATUSES", "TextTooLong",
|
||||
__all__ = ["ACTIVE", "BLANK_SAMPLE", "DraftBusy", "DraftFinished", "DraftJobs", "FINISHED", "GUIDANCE_MAX",
|
||||
"INTERRUPTED_ERROR", "MAX_TABLES", "NOTE_MAX", "NothingToRetry", "RETRYABLE", "TASK_DONE", "TASK_STATUSES",
|
||||
"TextTooLong",
|
||||
"active_job", "cancel", "check_model", "clean_text", "create_job", "ensure_idle", "guidance_of", "job_brief",
|
||||
"job_dict", "job_out", "last_drafts", "latest_job", "load", "manager", "retry_tables", "save_guidance",
|
||||
"startup", "tasks_of"]
|
||||
|
||||
@@ -204,6 +204,14 @@ _COLUMN_MIGRATIONS: list[tuple[str, str, str]] = [
|
||||
# 没有选项，也不是定时任务发起的
|
||||
("runs", "options", "JSON DEFAULT '{}'"),
|
||||
("runs", "schedule_id", "VARCHAR(32)"),
|
||||
# 数据目录起草任务（验收后修正）：超时、没写出来的调用次数，当前的列数上限，提醒；逐表的模型留空的列。
|
||||
# 存量任务回填 0 和 NULL：升级前没记这些
|
||||
("catalog_draft_jobs", "timeouts", "INTEGER DEFAULT 0"),
|
||||
("catalog_draft_jobs", "failed_calls", "INTEGER DEFAULT 0"),
|
||||
("catalog_draft_jobs", "part_columns", "INTEGER"),
|
||||
("catalog_draft_jobs", "notice", "TEXT"),
|
||||
("catalog_draft_tasks", "blank_columns", "INTEGER DEFAULT 0"),
|
||||
("catalog_draft_tasks", "blank_sample", "JSON"),
|
||||
]
|
||||
|
||||
#: 迁移补上的列要建的索引：(索引名, 表, 列)。create_all 不碰已有的表，表上新列的索引也就不会建。
|
||||
|
||||
@@ -379,10 +379,17 @@ class CatalogDraftJob(Base, TimestampMixin):
|
||||
created_by: Mapped[str | None] = mapped_column(String(100), default=None)
|
||||
started_at: Mapped[datetime | None] = mapped_column(UTCDateTime, default=None)
|
||||
finished_at: Mapped[datetime | None] = mapped_column(UTCDateTime, default=None)
|
||||
# 用量：调用次数、输入和输出 token（拿不到用量的调用按 0 算）
|
||||
# 用量：调用次数、输入和输出 token（拿不到用量的调用按 0 算，超时的调用也是 0）、超时几次、没写出来几次（额度用完、
|
||||
# 只给空壳、超时、出错）
|
||||
calls: Mapped[int] = mapped_column(Integer, default=0)
|
||||
input_tokens: Mapped[int] = mapped_column(Integer, default=0)
|
||||
output_tokens: Mapped[int] = mapped_column(Integer, default=0)
|
||||
timeouts: Mapped[int] = mapped_column(Integer, default=0)
|
||||
failed_calls: Mapped[int] = mapped_column(Integer, default=0)
|
||||
# 新开始的段每段最多几列（额度用完、超时多了会减半；用模型时开始起草后写上）
|
||||
part_columns: Mapped[int | None] = mapped_column(Integer, default=None)
|
||||
# 提醒：最近几次调用多数没写出来时的一句话（建议换起草模型、减少表或停止），少了清掉
|
||||
notice: Mapped[str | None] = mapped_column(Text, default=None)
|
||||
# 任务整体出错的原因（逐表的原因在 catalog_draft_tasks）
|
||||
error: Mapped[str | None] = mapped_column(Text, default=None)
|
||||
# 重试建的任务记下是重试哪一个
|
||||
@@ -411,6 +418,9 @@ class CatalogDraftTask(Base):
|
||||
removed: Mapped[int] = mapped_column(Integer, default=0)
|
||||
version: Mapped[int] = mapped_column(Integer, default=0)
|
||||
model_items: Mapped[int | None] = mapped_column(Integer, default=None)
|
||||
# 模型拿不准、留空了的列（中文名和含义都没给）：几列，最多记 5 个列名。表的状态不因此变
|
||||
blank_columns: Mapped[int] = mapped_column(Integer, default=0)
|
||||
blank_sample: Mapped[list[Any] | None] = mapped_column(default=None)
|
||||
# error：整张表没起草（表结构里没有、写入一直冲突、服务重启中断）；model_error：模型那部分没写出来的原因和办法
|
||||
error: Mapped[str | None] = mapped_column(Text, default=None)
|
||||
model_error: Mapped[str | None] = mapped_column(Text, default=None)
|
||||
|
||||
+14
-3
@@ -10,6 +10,7 @@
|
||||
不整列漏掉。list_all=False 时只给 replies 里有的列（模拟漏列的模型：漏掉的列会被补问一次，补问后仍没给的算没写出来）；
|
||||
- 直接 ainvoke 是纯文本那条路：同样的内容写成一段正文（AIMessage），前面带一句话；
|
||||
- fail 里的表所在的那一段抛异常，用来验证「一段失败只影响这一段」；
|
||||
- hang=N：头 N 次调用一直不回（等起草引擎按时限掐掉，当作超时）；
|
||||
- calls 记下每次收到的提示词全文，用来核对交给模型的内容（只有结构，没有数据）；modes 记下每次走的是
|
||||
structured 还是 text。
|
||||
|
||||
@@ -22,6 +23,7 @@ structured、text 模拟实测见过的几种回复（两条路各自设）：
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
from typing import Any
|
||||
|
||||
@@ -105,10 +107,11 @@ def with_every_column(entry: dict, wanted: list[str]) -> dict:
|
||||
|
||||
class FakeModel:
|
||||
def __init__(self, replies: dict[str, dict], fail: set[str] = frozenset(), *,
|
||||
structured: str = "tables", text: str = "tables", list_all: bool = True) -> None:
|
||||
structured: str = "tables", text: str = "tables", list_all: bool = True, hang: int = 0) -> None:
|
||||
self.replies, self.fail, self.calls = replies, set(fail), []
|
||||
self.structured, self.text = structured, text
|
||||
self.list_all = list_all
|
||||
self.hang = hang
|
||||
self.modes: list[str] = []
|
||||
|
||||
def with_structured_output(self, schema: Any, **_: Any) -> "_Structured":
|
||||
@@ -118,9 +121,17 @@ class FakeModel:
|
||||
async def ainvoke(self, messages: list[Any]) -> Any:
|
||||
from langchain_core.messages import AIMessage
|
||||
|
||||
reply = self._reply(messages, "text")
|
||||
reply = await self._answer(messages, "text")
|
||||
return AIMessage(content="草稿如下：" + json.dumps(reply, ensure_ascii=False))
|
||||
|
||||
async def _answer(self, messages: list[Any], mode: str) -> dict:
|
||||
if self.hang > 0:
|
||||
self.hang -= 1
|
||||
self.calls.append("\n".join(str(getattr(m, "content", m)) for m in messages))
|
||||
self.modes.append(mode)
|
||||
await asyncio.sleep(3600)
|
||||
return self._reply(messages, mode)
|
||||
|
||||
def _reply(self, messages: list[Any], mode: str) -> dict:
|
||||
text = "\n".join(str(getattr(m, "content", m)) for m in messages)
|
||||
self.calls.append(text)
|
||||
@@ -147,4 +158,4 @@ class _Structured:
|
||||
self.owner = owner
|
||||
|
||||
async def ainvoke(self, messages: list[Any]) -> dict:
|
||||
return self.owner._reply(messages, "structured")
|
||||
return await self.owner._answer(messages, "structured")
|
||||
|
||||
@@ -324,7 +324,7 @@ async def test_hollow_structured_output_retries_with_text():
|
||||
got = await catalog.draft_with_model(model, _source(_scenic_tables()), ["visits"])
|
||||
assert model.modes == ["structured", "text", "text"]
|
||||
assert got["visits"].notes == {"label": make_item("入园记录", "llm"), "kind": make_item("fact", "llm")}
|
||||
assert got["visits"].covered == frozenset() and "有 5 列没写出来" in got["visits"].error
|
||||
assert got["visits"].covered == frozenset() and "有 5 列（id 到 status）没写出来" in got["visits"].error
|
||||
|
||||
|
||||
async def test_hollow_structured_then_single_text_object():
|
||||
|
||||
@@ -6,7 +6,11 @@
|
||||
- 被截断的 JSON：已经完整的表、列对象捡回来，只补剩下的；
|
||||
- 只回一部分表：只补缺的，补一次；
|
||||
- 服务方嫌额度太大 → 退回 8192，后面都用它；结构化输出不支持 → 后面改走纯文本；
|
||||
- 网络错误、5xx、超时 → 退避重试，最多 2 次，报错写明重试过几次。
|
||||
- 网络错误、5xx → 退避重试，最多 2 次，报错写明重试过几次；超时当作这一段太大，拆小再问，不原样重试。
|
||||
|
||||
验收后修正的几条（用假模型复现）：宽表的列分段均分、没有很小的尾段；拆不开的空壳纯文本也再问一次；额度用完或超时
|
||||
累计 2 次后新开始的段列数减半；一段最多调用 8 次；最近 6 次调用 4 次没写出来时提醒、少于 2 次时清掉；记超时和没写
|
||||
出来的调用；模型留空的列计数；窄表不被宽表的列分段插队；原因写清是哪几段、哪几列。
|
||||
|
||||
表名、列名、说明文字都是虚构的。
|
||||
"""
|
||||
@@ -163,6 +167,34 @@ def _status(outcomes: dict) -> dict:
|
||||
return {t: o.status for t, o in outcomes.items()}
|
||||
|
||||
|
||||
def _hollow(call: Call) -> Reply:
|
||||
return Reply(obj={"tables": [{"table": t} for t in call.asked]})
|
||||
|
||||
|
||||
def _exhausted(seg: str = "这一段", budget: int = 32768) -> str:
|
||||
return D.REASON_EXHAUSTED.format(budget=budget, seg=seg)
|
||||
|
||||
|
||||
EXHAUSTED = Reply(text="", finish="length", thinking="……")
|
||||
|
||||
|
||||
class Reports:
|
||||
"""进度回调 call_done 收到的每次调用的结果。"""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.items: list[D.CallReport] = []
|
||||
|
||||
async def __call__(self, report: D.CallReport) -> None:
|
||||
self.items.append(report)
|
||||
|
||||
def hooks(self) -> D.Hooks:
|
||||
return D.Hooks(call_done=self)
|
||||
|
||||
@property
|
||||
def failures(self) -> list[str | None]:
|
||||
return [r.failure for r in self.items]
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 思考用完额度、空壳：拆开再问
|
||||
# --------------------------------------------------------------------------
|
||||
@@ -220,9 +252,10 @@ async def test_hollow_shell_splits_the_pack_by_table():
|
||||
assert all(c.asked[t][0] == "full" for c in model.calls[1:] for t in c.asked)
|
||||
|
||||
|
||||
async def test_splitting_stops_at_three_levels_and_five_columns():
|
||||
async def test_splitting_stops_at_three_levels_and_five_columns(monkeypatch):
|
||||
monkeypatch.setattr(catalog, "MODEL_PART_MAX_CALLS", 99)
|
||||
tables = {"wide40": _table(40)}
|
||||
model = Scripted(lambda call: Reply(text="", finish="length", thinking="……"))
|
||||
model = Scripted(lambda call: EXHAUSTED)
|
||||
got = await D.run_model_draft(model, _source(tables), ["wide40"])
|
||||
o = got["wide40"]
|
||||
assert len(model.calls) == 1 + 2 + 4 + 8, "40 → 20 → 10 → 5 列，最多拆三层"
|
||||
@@ -236,6 +269,30 @@ async def test_splitting_stops_at_three_levels_and_five_columns():
|
||||
assert got["t9"].error == "模型的思考和输出用完了额度（32768 个 token），这一段没写出来"
|
||||
|
||||
|
||||
async def test_one_part_makes_at_most_eight_calls(monkeypatch):
|
||||
"""一段（连同它拆出来的、补问的、重试的）最多调用 8 次：40 列一直用完额度，拆到第 8 次就停，剩下没问的记没写出来、
|
||||
原因写清。以前一段 40 列拆到底要 15 次。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_GIVE_UP_AFTER", 99)
|
||||
assert catalog.MODEL_PART_MAX_CALLS == 8
|
||||
model = Scripted(lambda call: EXHAUSTED)
|
||||
got = await D.run_model_draft(model, _source({"t40": _table(40)}), ["t40"])
|
||||
assert len(model.calls) == 8
|
||||
o = got["t40"]
|
||||
assert o.status == "failed"
|
||||
assert o.error == (_exhausted() + "；每段最多调用 8 次模型，这一段到了次数仍没写完，已停止重试"
|
||||
+ D.REASON_SPLIT)
|
||||
|
||||
# 网络错误的重试也算在这 8 次里
|
||||
def flaky(call: Call) -> Reply:
|
||||
raise ConnectError("All connection attempts failed")
|
||||
|
||||
monkeypatch.setattr(catalog, "MODEL_PART_MAX_CALLS", 2)
|
||||
model = Scripted(flaky)
|
||||
got = await D.run_model_draft(model, _source({"t4": _table(4)}), ["t4"])
|
||||
assert len(model.calls) == 2
|
||||
assert got["t4"].error == "每段最多调用 2 次模型，这一段到了次数仍没写完，已停止重试"
|
||||
|
||||
|
||||
async def test_an_unsplittable_hollow_shell_tries_text_once_then_stays_on_text():
|
||||
"""有的网关的结构化输出只给空壳：拆不开时换纯文本问一次，写出来了，这次起草后面直接走纯文本。
|
||||
|
||||
@@ -255,6 +312,31 @@ async def test_an_unsplittable_hollow_shell_tries_text_once_then_stays_on_text()
|
||||
assert catalog.CATALOG_DRAFT_TEXT_PROMPT in model.calls[1].text, "纯文本那条路要说清 JSON 的格式"
|
||||
|
||||
|
||||
def _text_only(handle_text: Callable[[Call], Reply]) -> Callable[[Call], Reply]:
|
||||
"""结构化输出不支持（这次起草一开始就改走纯文本），纯文本交给 handle_text。"""
|
||||
def handle(call: Call) -> Reply:
|
||||
if call.mode == "structured":
|
||||
raise _Status(400, "This response_format type is unavailable now")
|
||||
return handle_text(call)
|
||||
return handle
|
||||
|
||||
|
||||
async def test_an_unsplittable_hollow_shell_is_asked_again_in_text_mode_too():
|
||||
"""已经走纯文本时，拆不开的段（不到 10 列）只给空壳：原样再问一次，写出来就用；仍是空壳才记没写出来。
|
||||
以前纯文本模式下不补问，整张表直接记没写出来。"""
|
||||
tables = {"t8": _table(8)}
|
||||
ok = answering(tables)
|
||||
model = Scripted(_text_only(lambda call: _hollow(call) if call.n == 1 else ok(call)))
|
||||
got = await D.run_model_draft(model, _source(tables), ["t8"])
|
||||
assert [c.mode for c in model.calls] == ["structured", "text", "text"]
|
||||
assert got["t8"].status == "ok" and len(got["t8"].notes["columns"]) == 8
|
||||
|
||||
model = Scripted(_text_only(_hollow))
|
||||
got = await D.run_model_draft(model, _source(tables), ["t8"])
|
||||
assert [c.mode for c in model.calls] == ["structured", "text", "text"], "再问一次，不多问"
|
||||
assert got["t8"].status == "failed" and got["t8"].error == catalog.MODEL_EMPTY_REASON
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 截断、缺表、缺列
|
||||
# --------------------------------------------------------------------------
|
||||
@@ -358,12 +440,35 @@ async def test_columns_still_missing_after_the_followup_are_not_written():
|
||||
assert len(model.calls) == 2 and model.calls[1].asked == {"t8": ("columns", ["c007", "c008"])}
|
||||
o = got["t8"]
|
||||
assert o.status == "partial"
|
||||
assert o.error == "这张表 8 列，有 2 列没写出来：模型的回复里漏了这些列，补问一次仍没有给出。可以只重试这张表"
|
||||
assert o.error == ("这张表 8 列，有 2 列（c007、c008）没写出来：模型的回复里漏了这些列，补问一次仍没有给出。"
|
||||
"可以只重试这张表")
|
||||
assert set(o.notes["columns"]) == {f"c{i:03d}" for i in range(1, 7)}
|
||||
assert o.blank == [], "没写出来的列不算留空"
|
||||
compat = (await catalog.draft_with_model(Scripted(handle), _source(tables), ["t8"]))["t8"]
|
||||
assert compat.covered == frozenset(), "没写全的不算完整起草过"
|
||||
|
||||
|
||||
async def test_columns_listed_blank_in_the_followup_are_left_blank_not_failed():
|
||||
"""补问时模型把漏的列列出来了、字段却留空：是拿不准（记留空），不算「漏了这些列」。"""
|
||||
tables = {"t8": _table(8)}
|
||||
ok = answering(tables)
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
reply = ok(call)
|
||||
entry = reply.obj["tables"][0]
|
||||
if call.n == 0:
|
||||
entry["columns"] = [c for c in entry["columns"] if c["name"] not in ("c003", "c007")]
|
||||
else:
|
||||
entry["columns"] = [{"name": c["name"]} for c in entry["columns"]]
|
||||
return reply
|
||||
|
||||
model = Scripted(handle)
|
||||
got = await D.run_model_draft(model, _source(tables), ["t8"])
|
||||
assert model.calls[1].asked == {"t8": ("columns", ["c003", "c007"])}
|
||||
o = got["t8"]
|
||||
assert o.status == "ok" and o.error is None and o.blank == ["c003", "c007"]
|
||||
|
||||
|
||||
async def test_a_followup_that_comes_back_as_an_empty_shell_leaves_the_table_partial():
|
||||
"""40 列的表第一次只回了 20 列，补问只给一个空壳：没写出来的 20 列照实记下，表是 partial。"""
|
||||
tables = {"t40": _table(40)}
|
||||
@@ -382,11 +487,12 @@ async def test_a_followup_that_comes_back_as_an_empty_shell_leaves_the_table_par
|
||||
assert len(model.calls) == 2
|
||||
assert model.calls[1].asked == {"t40": ("columns", [f"c{i:03d}" for i in range(21, 41)])}
|
||||
assert o.status == "partial" and len(o.notes["columns"]) == 20
|
||||
assert o.error == "这张表 40 列，有 20 列没写出来：模型的回复里漏了这些列，补问一次仍没有给出。可以只重试这张表"
|
||||
assert o.error == ("这张表 40 列，有 20 列（c021 到 c040）没写出来：模型的回复里漏了这些列，补问一次仍没有给出。"
|
||||
"可以只重试这张表")
|
||||
|
||||
|
||||
async def test_columns_listed_with_empty_fields_are_left_to_the_model():
|
||||
"""列出了、字段留空的列是模型拿不准：不补问，也不算没写出来。"""
|
||||
"""列出了、字段留空的列是模型拿不准：不补问，也不算没写出来；记成留空（表的状态不因此变）。"""
|
||||
tables = {"t8": _table(8)}
|
||||
ok = answering(tables)
|
||||
|
||||
@@ -402,6 +508,30 @@ async def test_columns_listed_with_empty_fields_are_left_to_the_model():
|
||||
assert len(model.calls) == 1
|
||||
assert got["t8"].status == "ok" and got["t8"].error is None
|
||||
assert set(got["t8"].notes["columns"]) == {f"c{i:03d}" for i in range(1, 7)}
|
||||
assert got["t8"].blank == ["c007", "c008"]
|
||||
|
||||
|
||||
async def test_blank_columns_are_the_ones_without_a_label_or_meaning():
|
||||
"""留空的列：请模型写了、中文名和含义却都没给（只给了度量类型、单位，或者给的都没通过校验）。已确认的、库注释已经说清
|
||||
的不算：给模型的提示词里它们以此为准，模型不写也不缺。"""
|
||||
meta = _table(9)
|
||||
meta["columns"][5]["comment"] = "会员等级" # c006：库注释就是中文名
|
||||
meta["columns"][8]["comment"] = "C009" # c009：注释就是列名，说明不了什么
|
||||
tables = {"t9": meta}
|
||||
existing = {"t9": {"columns": {"c005": {"label": catalog.make_item("已确认的名字", "human")}}}}
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
return Reply(obj={"tables": [{"table": "t9", "label": "某表", "columns": [
|
||||
{"name": "c001", "label": "第一列"},
|
||||
{"name": "c002", "meaning": "第二列的含义"},
|
||||
{"name": "c003", "measure": "attribute", "unit": "元"}, # 中文名、含义都没给
|
||||
{"name": "c004", "label": "=c001+c002"}, # 像公式：校验丢掉
|
||||
{"name": "c005"}, {"name": "c006"}, {"name": "c007"}, {"name": "c008"}, {"name": "c009"}]}]})
|
||||
|
||||
got = await D.run_model_draft(Scripted(handle), _source(tables), ["t9"], existing=existing)
|
||||
o = got["t9"]
|
||||
assert o.status == "ok" and o.error is None
|
||||
assert o.blank == ["c003", "c004", "c007", "c008", "c009"]
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
@@ -521,7 +651,29 @@ async def test_other_errors_are_not_retried():
|
||||
assert len(model.calls) == 1 and got["t4"].error == "调用模型出错：网关拒绝了这次请求"
|
||||
|
||||
|
||||
async def test_a_call_past_the_time_limit_is_retried_then_reported(monkeypatch):
|
||||
async def test_a_call_past_the_time_limit_is_split_not_retried(monkeypatch, _no_wait):
|
||||
"""超时当作这一段太大：不原样重试同样大小的段，拆成两半再问（同额度用完）。超时的调用拿不到用量，记 0。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_TIMEOUT_S", 0.05)
|
||||
tables = {"t24": _table(24)}
|
||||
ok = answering(tables, usage=(100, 50))
|
||||
|
||||
async def handle(call: Call) -> Reply:
|
||||
if call.n == 0:
|
||||
await asyncio.sleep(1)
|
||||
return ok(call)
|
||||
|
||||
model, reports = Scripted(handle), Reports()
|
||||
got = await D.run_model_draft(model, _source(tables), ["t24"], hooks=reports.hooks())
|
||||
cols = [f"c{i:03d}" for i in range(1, 25)]
|
||||
assert got["t24"].status == "ok"
|
||||
assert [c.asked for c in model.calls] == [{"t24": ("full", None)}, {"t24": ("table+columns", cols[:12])},
|
||||
{"t24": ("columns", cols[12:])}]
|
||||
assert reports.failures == ["timeout", None, None]
|
||||
assert [(r.input_tokens, r.output_tokens) for r in reports.items] == [(0, 0), (100, 50), (100, 50)]
|
||||
assert _no_wait == [], "不退避重试"
|
||||
|
||||
|
||||
async def test_a_small_part_past_the_time_limit_is_not_retried(monkeypatch):
|
||||
monkeypatch.setattr(catalog, "MODEL_TIMEOUT_S", 0.05)
|
||||
|
||||
async def slow(call: Call) -> Reply:
|
||||
@@ -530,8 +682,8 @@ async def test_a_call_past_the_time_limit_is_retried_then_reported(monkeypatch):
|
||||
|
||||
model = Scripted(slow)
|
||||
got = await D.run_model_draft(model, _source({"t4": _table(4)}), ["t4"])
|
||||
assert len(model.calls) == 3
|
||||
assert got["t4"].error == "模型 0.05 秒内没有写完这一段（已重试 2 次）"
|
||||
assert len(model.calls) == 1, "拆不开、也不原样重试"
|
||||
assert got["t4"].status == "failed" and got["t4"].error == "模型 0.05 秒内没有写完这一段"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
@@ -554,13 +706,30 @@ async def test_a_wide_table_is_split_into_a_table_part_and_column_parts():
|
||||
first, *rest = model.calls
|
||||
assert first.asked == {"wide95": ("table", None)}
|
||||
assert catalog.CATALOG_DRAFT_PROMPT_LABELS["brief_columns"].format(n=95) in first.text
|
||||
assert sorted(len(c.asked["wide95"][1]) for c in rest) == [15, 40, 40]
|
||||
assert sorted(len(c.asked["wide95"][1]) for c in rest) == [31, 32, 32]
|
||||
for call in rest:
|
||||
assert call.asked["wide95"][0] == "columns"
|
||||
assert "wide95的中文名" in call.text and "虚构的宽表" in call.text, "列分段带着表注释和表级起草结果"
|
||||
assert {c for call in rest for c in call.asked["wide95"][1]} == {f"c{i:03d}" for i in range(1, 96)}
|
||||
|
||||
|
||||
@pytest.mark.parametrize("n, sizes", [(164, [33, 33, 33, 33, 32]), (95, [32, 32, 31]), (41, [21, 20]),
|
||||
(80, [40, 40]), (81, [27, 27, 27])])
|
||||
async def test_column_parts_are_even_without_a_small_tail(monkeypatch, n, sizes):
|
||||
"""列分段数 = 列数 / 40 向上取整，各段列数平均（相差不超过 1）：不再留下几列的尾段（实测会思考的模型常对这种小段只给
|
||||
空壳）。表级那段照旧在最前面。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_CONCURRENCY", 1)
|
||||
tables = {"wide": _table(n)}
|
||||
model = Scripted(answering(tables))
|
||||
got = await D.run_model_draft(model, _source(tables), ["wide"])
|
||||
o = got["wide"]
|
||||
assert o.status == "ok" and o.parts_total == 1 + len(sizes) == D.parts_for(n)
|
||||
assert model.calls[0].asked == {"wide": ("table", None)}
|
||||
parts = [c.asked["wide"][1] for c in model.calls[1:]]
|
||||
assert [len(p) for p in parts] == sizes
|
||||
assert [c for p in parts for c in p] == [f"c{i:03d}" for i in range(1, n + 1)], "按表结构的顺序连续切开"
|
||||
|
||||
|
||||
async def test_a_wide_table_with_one_failed_part_is_partial_and_says_which():
|
||||
tables = {"wide95": _table(95)}
|
||||
ok = answering(tables)
|
||||
@@ -574,12 +743,69 @@ async def test_a_wide_table_with_one_failed_part_is_partial_and_says_which():
|
||||
model = Scripted(handle)
|
||||
got = await D.run_model_draft(model, _source(tables), ["wide95"])
|
||||
o = got["wide95"]
|
||||
assert o.status == "partial" and len(o.notes["columns"]) == 55
|
||||
assert o.error == "这张表 95 列，分 4 段起草，第 3 段没写出来（共 40 列）：调用模型出错：网关拒绝。可以只重试这张表"
|
||||
assert o.status == "partial" and len(o.notes["columns"]) == 63
|
||||
assert o.error == ("这张表 95 列，分 4 段起草，第 3 段（c033 到 c064，共 32 列）没写出来：调用模型出错：网关拒绝。"
|
||||
"可以只重试这张表")
|
||||
compat = (await catalog.draft_with_model(Scripted(handle), _source(tables), ["wide95"]))["wide95"]
|
||||
assert compat.covered == frozenset() and compat.error == o.error, "没写全的不算完整起草过，旧的模型项不删"
|
||||
|
||||
|
||||
async def test_reasons_name_each_part_and_say_these_parts(monkeypatch):
|
||||
"""几段因为同一个原因没写出来：原因里写「这几段」；每段写上列的范围（列名照表结构原样）；拆过的才补「已拆小重试」。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_SHRINK_AFTER", 99)
|
||||
tables = {"wide95": _table(95)}
|
||||
ok = answering(tables)
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
part, cols = call.asked["wide95"]
|
||||
if part == "columns" and cols[0] >= "c033": # 第 3、4 段（c033 起）一直用完额度
|
||||
return EXHAUSTED
|
||||
return ok(call)
|
||||
|
||||
got = await D.run_model_draft(Scripted(handle), _source(tables), ["wide95"])
|
||||
o = got["wide95"]
|
||||
assert o.status == "partial" and len(o.notes["columns"]) == 32
|
||||
assert o.error == ("这张表 95 列，分 4 段起草，第 3 段（c033 到 c064，共 32 列）、第 4 段（c065 到 c095，共 31 列）没写出来："
|
||||
+ _exhausted("这几段") + D.REASON_SPLIT + "。可以只重试这张表")
|
||||
|
||||
|
||||
async def test_a_part_that_was_never_halved_does_not_say_it_was():
|
||||
"""被截断后只补剩下的（不是拆小）：补的那次也用完了额度、又拆不开，原因里不写「已拆小重试」。"""
|
||||
tables = {"t9": _table(9)}
|
||||
full = json.dumps({"tables": [{"table": "t9", **_reply("t9", tables["t9"])}]}, ensure_ascii=False)
|
||||
cut = full[:full.index('"c005"') + 8]
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
return Reply(text=cut, finish="length") if call.n == 0 else EXHAUSTED
|
||||
|
||||
model = Scripted(handle)
|
||||
got = await D.run_model_draft(model, _source(tables), ["t9"])
|
||||
o = got["t9"]
|
||||
assert len(model.calls) == 2 and model.calls[1].asked == {"t9": ("columns", [f"c{i:03d}" for i in range(5, 10)])}
|
||||
assert o.status == "partial"
|
||||
assert o.error == "这张表 9 列，有 5 列（c005 到 c009）没写出来：" + _exhausted() + "。可以只重试这张表"
|
||||
|
||||
|
||||
@pytest.mark.parametrize("missing, said", [
|
||||
(["c004"], "有 1 列（c004）"),
|
||||
(["c003", "c007"], "有 2 列（c003、c007）"),
|
||||
(["c002", "c004", "c006", "c008"], "有 4 列（c002、c004、c006 等）"),
|
||||
(["c003", "c004", "c005", "c006"], "有 4 列（c003 到 c006）"),
|
||||
])
|
||||
async def test_columns_not_written_are_named(missing, said):
|
||||
tables = {"t8": _table(8)}
|
||||
ok = answering(tables)
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
reply = ok(call)
|
||||
for entry in reply.obj["tables"]:
|
||||
entry["columns"] = [c for c in entry.get("columns", []) if c["name"] not in missing]
|
||||
return reply
|
||||
|
||||
got = await D.run_model_draft(Scripted(handle), _source(tables), ["t8"])
|
||||
assert got["t8"].error == f"这张表 8 列，{said}没写出来：{D.REASON_UNANSWERED}。可以只重试这张表"
|
||||
|
||||
|
||||
async def test_tables_are_packed_up_to_40_columns_and_4_tables():
|
||||
widths = {"ta": 15, "tb": 15, "tc": 15, "td": 5, "te": 5, "tf": 5, "tg": 5, "th": 41}
|
||||
tables = {name: _table(n) for name, n in widths.items()}
|
||||
@@ -598,13 +824,10 @@ async def test_tables_are_packed_up_to_40_columns_and_4_tables():
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _hollow(call: Call) -> Reply:
|
||||
return Reply(obj={"tables": [{"table": t} for t in call.asked]})
|
||||
|
||||
|
||||
async def test_the_same_reason_is_said_once_even_when_some_parts_were_split(monkeypatch):
|
||||
"""宽表的表级那段拆不开、列分段拆过：原因只说一次，拆过的在最后补一句。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_GIVE_UP_AFTER", 99)
|
||||
monkeypatch.setattr(catalog, "MODEL_PART_MAX_CALLS", 99)
|
||||
model = Scripted(_hollow)
|
||||
got = await D.run_model_draft(model, _source({"wide45": _table(45)}), ["wide45"])
|
||||
o = got["wide45"]
|
||||
@@ -613,14 +836,19 @@ async def test_the_same_reason_is_said_once_even_when_some_parts_were_split(monk
|
||||
|
||||
|
||||
def test_reasons_are_joined_by_their_base():
|
||||
"""(原因, 第几段)：同样的原因只说一次；同一个原因落在几段上写「这几段」，同一段拆出来的几块仍是「这一段」。"""
|
||||
W = D._Why
|
||||
exhausted = D.REASON_EXHAUSTED.format(budget=32768)
|
||||
assert D._join([W(D.REASON_HOLLOW), W(D.REASON_HOLLOW, split=True)]) == D.REASON_HOLLOW + D.REASON_SPLIT
|
||||
assert D._join([W(exhausted, split=True), W(D.REASON_HOLLOW)]) == \
|
||||
f"{exhausted}；{D.REASON_HOLLOW}{D.REASON_SPLIT}"
|
||||
assert D._join([W("甲"), W("乙"), W("丙", split=True)]) == "甲；乙", "没说出来的原因不补后缀"
|
||||
assert D._join([W(D.REASON_HOLLOW, halted=True), W(D.REASON_HOLLOW)]) == \
|
||||
ex = D.REASON_EXHAUSTED.format(budget=32768, seg=D.SEG)
|
||||
assert D._join([(W(D.REASON_HOLLOW), 1), (W(D.REASON_HOLLOW, split=True), 2)]) == \
|
||||
D.REASON_HOLLOW + D.REASON_SPLIT
|
||||
assert D._join([(W(ex, split=True), 2), (W(D.REASON_HOLLOW), 1)]) == \
|
||||
f"{_exhausted()}；{D.REASON_HOLLOW}{D.REASON_SPLIT}"
|
||||
assert D._join([(W(ex), 2), (W(ex), 3)]) == _exhausted("这几段")
|
||||
assert D._join([(W(ex), 2), (W(ex, split=True), 2)]) == _exhausted() + D.REASON_SPLIT
|
||||
assert D._join([(W("甲"), 1), (W("乙"), 1), (W("丙", split=True), 1)]) == "甲；乙", "没说出来的原因不补后缀"
|
||||
assert D._join([(W(D.REASON_HOLLOW, halted=True), 1), (W(D.REASON_HOLLOW), 2)]) == \
|
||||
D.REASON_HOLLOW + D.REASON_HALTED.format(n=catalog.MODEL_GIVE_UP_AFTER)
|
||||
assert D._join([(W("调用模型出错：{坏的}"), 1)]) == "调用模型出错：{坏的}", "报错原文里的花括号原样留着"
|
||||
|
||||
|
||||
def _by_table(model: Scripted) -> dict[str, int]:
|
||||
@@ -643,7 +871,7 @@ async def test_after_two_segments_in_a_row_come_to_nothing_the_rest_are_not_spli
|
||||
model = Scripted(lambda call: exhausted)
|
||||
got = await D.run_model_draft(model, _source(tables), list(tables))
|
||||
assert _by_table(model) == {"t1": 3, "t2": 3, "t3": 1, "t4": 1}
|
||||
budget = D.REASON_EXHAUSTED.format(budget=32768)
|
||||
budget = _exhausted()
|
||||
assert got["t1"].error == budget + D.REASON_SPLIT
|
||||
assert got["t3"].error == budget + "；模型连续 2 段都没写出来，已停止拆小重试，建议更换起草模型"
|
||||
|
||||
@@ -657,23 +885,23 @@ async def test_after_two_segments_in_a_row_come_to_nothing_the_rest_are_not_spli
|
||||
async def test_the_column_parts_of_a_wide_table_stop_splitting_too(monkeypatch):
|
||||
"""宽表的列分段一段一段排队起草（一段拆到底再起草下一段），止损在宽表里同样起作用。
|
||||
|
||||
95 列：表级那段拆不开（结构化、纯文本各一次）；第 2 段 40 列拆到底（15 次结构化 + 8 次纯文本）；这时已经连续 2 段
|
||||
没写出来，第 3、4 段只问一次、换纯文本一次。"""
|
||||
95 列：表级那段拆不开（结构化、纯文本各一次）；第 2 段 32 列拆小、到 8 次封顶；这时已经连续 2 段没写出来，第 3、4 段
|
||||
只问一次、换纯文本一次。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_CONCURRENCY", 1)
|
||||
model = Scripted(_hollow)
|
||||
got = await D.run_model_draft(model, _source({"wide95": _table(95)}), ["wide95"])
|
||||
assert len(model.calls) == 2 + 23 + 2 + 2
|
||||
assert got["wide95"].error == (D.REASON_HOLLOW + D.REASON_SPLIT
|
||||
+ "；模型连续 2 段都没写出来，已停止拆小重试，建议更换起草模型")
|
||||
assert len(model.calls) == 2 + 8 + 2 + 2
|
||||
assert got["wide95"].error == (D.REASON_HOLLOW + "；每段最多调用 8 次模型，这一段到了次数仍没写完，已停止重试"
|
||||
+ D.REASON_SPLIT + "；模型连续 2 段都没写出来，已停止拆小重试，建议更换起草模型")
|
||||
|
||||
|
||||
async def test_a_wide_table_that_never_comes_back_is_bounded_by_the_first_wave():
|
||||
"""默认并发（3 段同时起草）：294 列一直只给空壳，最多是表级那段 2 次、头 3 个列分段拆到底（各 23 次）、其余 5 段
|
||||
各 2 次。以前所有列分段一下子都排上去、一层一层地拆，要 168 次。"""
|
||||
"""默认并发（3 段同时起草）：294 列（8 个列分段）一直只给空壳，最多是表级那段 2 次、头 3 个列分段各 8 次（封顶）、
|
||||
其余 5 段各 2 次。以前所有列分段一下子都排上去、一层一层地拆，要 168 次。"""
|
||||
model = Scripted(_hollow)
|
||||
got = await D.run_model_draft(model, _source({"wide294": _table(294)}), ["wide294"])
|
||||
assert got["wide294"].status == "failed"
|
||||
assert len(model.calls) <= 2 + 3 * 23 + 5 * 2
|
||||
assert got["wide294"].status == "failed" and got["wide294"].parts_total == 9
|
||||
assert len(model.calls) <= 2 + 3 * 8 + 5 * 2
|
||||
|
||||
|
||||
async def test_a_big_draft_does_not_hold_the_model_ahead_of_a_later_one(monkeypatch):
|
||||
@@ -707,6 +935,85 @@ async def test_a_big_draft_does_not_hold_the_model_ahead_of_a_later_one(monkeypa
|
||||
assert order == ["a1", "b1", "a2", "a3", "a4"]
|
||||
|
||||
|
||||
def _order(model: Scripted) -> list[tuple[str, str]]:
|
||||
return [(name, part) for call in model.calls for name, (part, _) in call.asked.items()]
|
||||
|
||||
|
||||
async def test_narrow_tables_are_not_cut_in_by_the_column_parts_of_a_wide_one(monkeypatch):
|
||||
"""一个任务里按表的先后排队：宽表的表级那段写完，它的列分段排到已经排着的窄表段后面（先来先服务），不插队。
|
||||
以前插到队首，窄表要等宽表的列分段全部派完。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_CONCURRENCY", 1)
|
||||
monkeypatch.setattr(catalog, "MODEL_BATCH_TABLES", 1)
|
||||
tables = {"wide": _table(45), "na": _table(4), "nb": _table(4)}
|
||||
model = Scripted(answering(tables))
|
||||
got = await D.run_model_draft(model, _source(tables), ["wide", "na", "nb"])
|
||||
assert set(_status(got).values()) == {"ok"}
|
||||
assert _order(model) == [("wide", "table"), ("na", "full"), ("nb", "full"), ("wide", "columns"),
|
||||
("wide", "columns")]
|
||||
|
||||
|
||||
async def test_after_two_exhausted_or_timed_out_calls_new_parts_are_halved(monkeypatch):
|
||||
"""额度用完或超时累计 2 次，之后新开始的段列数上限减半（40 → 20 → 10，最低 10）；已经开始的段照旧（按原因拆小）。
|
||||
每次调用的结果里带着当前的上限（part_columns）。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_CONCURRENCY", 1)
|
||||
monkeypatch.setattr(catalog, "MODEL_BATCH_TABLES", 1)
|
||||
tables = {f"t{i}": _table(40) for i in range(1, 4)}
|
||||
ok = answering(tables)
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
(_, (part, cols)), = call.asked.items()
|
||||
width = 40 if part == "full" else len(cols)
|
||||
return EXHAUSTED if width > 10 else ok(call) # 这个模型一次只写得出 10 列
|
||||
|
||||
model, reports = Scripted(handle), Reports()
|
||||
got = await D.run_model_draft(model, _source(tables), list(tables), hooks=reports.hooks())
|
||||
assert set(_status(got).values()) == {"ok"}
|
||||
asked = {name: [(part, len(cols or [])) for c in model.calls for n, (part, cols) in c.asked.items() if n == name]
|
||||
for name in tables}
|
||||
# t1：整张 40 列开始，用完额度拆成 20、20，又各拆成 10、10
|
||||
assert asked["t1"][0] == ("full", 0) and len(asked["t1"]) == 7
|
||||
# 第 2 次用完额度后上限是 20：t2 一开始就按 20 列切成两段（各自再拆一次）
|
||||
assert asked["t2"][0] == ("table+columns", 20) and ("columns", 20) in asked["t2"]
|
||||
# 第 4 次之后上限是 10：t3 一开始就切成 4 段 10 列，一次都没用完额度
|
||||
assert asked["t3"] == [("table+columns", 10), ("columns", 10), ("columns", 10), ("columns", 10)]
|
||||
caps = [r.part_columns for r in reports.items]
|
||||
assert list(dict.fromkeys(caps)) == [40, 20, 10]
|
||||
assert caps[:2] == [40, 20], "第 2 次用完额度的那次调用之后就是 20"
|
||||
assert min(caps) == 10 == catalog.MODEL_MIN_PART_COLUMNS
|
||||
|
||||
|
||||
async def test_the_notice_comes_when_four_of_six_calls_fail_and_goes_when_fewer_than_two(monkeypatch):
|
||||
"""最近 6 次调用有 4 次以上没写出来：提醒（写明次数和原因，建议换起草模型、减少表或停止）；之后最近 6 次里没写出来
|
||||
的少于 2 次就清掉。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_CONCURRENCY", 1)
|
||||
monkeypatch.setattr(catalog, "MODEL_BATCH_TABLES", 1)
|
||||
tables = {f"t{i}": _table(4) for i in range(1, 10)}
|
||||
ok = answering(tables)
|
||||
|
||||
def handle(call: Call) -> Reply:
|
||||
name = next(iter(call.asked))
|
||||
if name in ("t1", "t2"):
|
||||
raise RuntimeError("网关拒绝了这次请求")
|
||||
if name == "t3":
|
||||
return EXHAUSTED
|
||||
if name == "t4":
|
||||
return _hollow(call)
|
||||
return ok(call)
|
||||
|
||||
reports = Reports()
|
||||
got = await D.run_model_draft(Scripted(handle), _source(tables), list(tables), hooks=reports.hooks())
|
||||
assert reports.failures == ["error", "error", "exhausted", "hollow", "hollow"] + [None] * 5
|
||||
notices = [r.notice for r in reports.items]
|
||||
assert notices[:3] == [None] * 3
|
||||
assert notices[3] == ("最近 4 次调用有 4 次没写出来（额度用完 1 次、只给空壳 1 次、出错 2 次）。可以在「设置 → 偏好设置」"
|
||||
"的「起草模型」换一个模型、减少这次起草的表，或者停止这次起草")
|
||||
assert notices[4].startswith("最近 5 次调用有 5 次没写出来")
|
||||
assert notices[7].startswith("最近 6 次调用有 3 次没写出来"), "还有 2 次以上没写出来：留着，次数照实写"
|
||||
assert notices[8].startswith("最近 6 次调用有 2 次没写出来")
|
||||
assert notices[9] is None, "最近 6 次只有 1 次没写出来：清掉"
|
||||
assert got["t9"].status == "ok"
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 起草要求、用量
|
||||
# --------------------------------------------------------------------------
|
||||
@@ -747,13 +1054,11 @@ async def test_usage_is_added_up_through_the_hooks():
|
||||
return Reply(text="", finish="length", thinking="……", usage=(100, 32768))
|
||||
return ok(call)
|
||||
|
||||
seen: list[tuple[int, int]] = []
|
||||
|
||||
async def usage(i: int, o: int) -> None:
|
||||
seen.append((i, o))
|
||||
|
||||
await D.run_model_draft(Scripted(handle), _source(tables), ["orders_x"], hooks=D.Hooks(usage=usage))
|
||||
assert seen == [(100, 32768), (120, 60), (120, 60)]
|
||||
reports = Reports()
|
||||
await D.run_model_draft(Scripted(handle), _source(tables), ["orders_x"], hooks=reports.hooks())
|
||||
assert [(r.input_tokens, r.output_tokens, r.failure) for r in reports.items] == \
|
||||
[(100, 32768, "exhausted"), (120, 60, None), (120, 60, None)]
|
||||
assert {r.notice for r in reports.items} == {None} and {r.part_columns for r in reports.items} == {40}
|
||||
|
||||
|
||||
def test_budget_words_match_the_other_drafts():
|
||||
|
||||
@@ -6,7 +6,9 @@
|
||||
- 停止：排队的直接 cancelled；进行中的停掉正在进行的模型调用，没写完的表记 cancelled，已经写进目录的不回滚；已结束的 409
|
||||
- 服务启动时把没结束的任务标成 interrupted（没结束的表记 cancelled、写明服务重启），不自动续跑；关停时同样
|
||||
- 重试：只含原任务里 failed、partial、cancelled 的表，本次说明沿用、起草要求取数据源当前的；没有可重试的 422
|
||||
- 目录清单：每行 last_draft，顶层 guidance、active_job
|
||||
- 目录清单：每行 last_draft（带模型留空的列数），顶层 guidance、active_job
|
||||
- 重新起草时模型这次留空、给的值没通过校验：以前起草好的模型项留着，删除只在列已经不在表结构里时才有
|
||||
- 用量记超时次数、没写出来的调用次数；任务记当前的列数上限、提醒（最近 6 次调用 4 次没写出来）；逐表记模型留空的列
|
||||
- 起草要求存在数据源 options 里（上传表格的源存在它自己那一行），数据源接口不显示、整份回写 options 时留着，拼连接串时摘掉
|
||||
|
||||
表名、列名、说明文字都是虚构的。
|
||||
@@ -261,11 +263,136 @@ async def test_a_followup_that_comes_back_empty_keeps_the_earlier_model_items(cl
|
||||
task = job["tasks"][0]
|
||||
assert model.calls == 2
|
||||
assert job["status"] == "partial" and task["status"] == "partial" and task["removed"] == 0
|
||||
assert "有 4 列没写出来" in task["model_error"] and "可以只重试这张表" in task["model_error"]
|
||||
assert "有 4 列（c05 到 c08）没写出来" in task["model_error"] and "可以只重试这张表" in task["model_error"]
|
||||
after = (await client.get(f"{_base(sid)}/tickets")).json()["notes"]["columns"]
|
||||
assert {c for c, items in after.items() if "label" in items} == {f"c{i:02d}" for i in range(1, 9)}
|
||||
|
||||
|
||||
def _tickets(n: int = 6) -> dict:
|
||||
return {"tickets": {"qualified": "tickets", "columns": [_col(f"c{i:02d}") for i in range(1, n + 1)],
|
||||
"primary_key": ["c01"], "comment": None, "is_view": False, "foreign_keys": []}}
|
||||
|
||||
|
||||
async def _draft(client, sid: str, monkeypatch, model, tables: list[str]) -> dict:
|
||||
use(monkeypatch, model)
|
||||
r = await client.post(f"{_base(sid)}/draft-jobs", json={"tables": tables, "use_model": True})
|
||||
assert r.status_code == 202, r.text
|
||||
await finished(r.json()["job"]["id"])
|
||||
return await get_job(client, sid, r.json()["job"]["id"])
|
||||
|
||||
|
||||
async def test_a_redraft_keeps_what_the_model_left_blank_this_time(client, monkeypatch):
|
||||
"""重新起草（比如补了起草要求再起草一次）时模型这次留空、或者给的值没通过校验：以前起草好的模型项留着，不算删除；
|
||||
给了合格新值的才替换。以前整库重跑一次删掉几百项，表仍记「已完成」。"""
|
||||
sid = await _source(_tickets())
|
||||
full = {"label": "门票", "kind": "fact",
|
||||
"columns": [{"name": f"c{i:02d}", "label": f"第{i}列", "meaning": f"第{i}列的含义"} for i in range(1, 7)]}
|
||||
first = await _draft(client, sid, monkeypatch, FakeModel({"tickets": full}), ["tickets"])
|
||||
assert first["tasks"][0]["status"] == "ok" and first["blank_columns"] == 0
|
||||
before = (await client.get(f"{_base(sid)}/tickets")).json()["notes"]
|
||||
|
||||
# 这一次：表的中文名换了；c03 的中文名像公式（校验丢掉）、含义换了；c04 只给了中文名；其余列列出了但都留空
|
||||
again = {"label": "门票明细", "columns": [{"name": "c03", "label": "=c01+c02", "meaning": "新的含义"},
|
||||
{"name": "c04", "label": "第4列"}]}
|
||||
second = await _draft(client, sid, monkeypatch, FakeModel({"tickets": again}), ["tickets"])
|
||||
task = second["tasks"][0]
|
||||
assert task["status"] == "ok" and task["removed"] == 0 and second["removed"] == 0
|
||||
assert (task["added"], task["updated"]) == (0, 2)
|
||||
after = (await client.get(f"{_base(sid)}/tickets")).json()["notes"]
|
||||
assert after["label"]["value"] == "门票明细" and after["kind"] == before["kind"]
|
||||
assert after["columns"]["c03"]["label"] == before["columns"]["c03"]["label"]
|
||||
assert after["columns"]["c03"]["meaning"]["value"] == "新的含义"
|
||||
for col in ("c01", "c02", "c05", "c06"):
|
||||
assert after["columns"][col] == before["columns"][col], col
|
||||
# 模型这次留空的列照实记下（表的状态不因此变）
|
||||
assert task["blank_columns"] == 4 and task["blank_sample"] == ["c01", "c02", "c05", "c06"]
|
||||
|
||||
|
||||
async def test_model_items_of_a_column_gone_from_the_schema_are_removed(client, monkeypatch):
|
||||
sid = await _source(_tickets())
|
||||
full = {"label": "门票", "columns": [{"name": f"c{i:02d}", "label": f"第{i}列"} for i in range(1, 7)]}
|
||||
await _draft(client, sid, monkeypatch, FakeModel({"tickets": full}), ["tickets"])
|
||||
async with SessionLocal() as session:
|
||||
row = await session.get(DataSource, sid)
|
||||
row.schema_cache = {"tables": _tickets(5)} # 重新探查过：c06 已经没有了
|
||||
await session.commit()
|
||||
job = await _draft(client, sid, monkeypatch, FakeModel({"tickets": {"label": "门票"}}), ["tickets"])
|
||||
assert job["tasks"][0]["removed"] == 1
|
||||
after = (await client.get(f"{_base(sid)}/tickets")).json()["notes"]
|
||||
assert set(after["columns"]) == {"c01", "c02", "c03", "c04", "c05"}
|
||||
|
||||
|
||||
async def test_usage_counts_timeouts_and_failed_calls(client, monkeypatch):
|
||||
"""超时的调用拆小再问，用量里记超时几次、没写出来几次（超时的调用拿不到用量，token 记 0）；额度用完或超时累计 2 次，
|
||||
新开始的段列数上限减半，任务上记着当前的上限。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_TIMEOUT_S", 0.05)
|
||||
sid = await _source(_tickets(24))
|
||||
replies = {"tickets": {"label": "门票", "columns": [{"name": f"c{i:02d}", "label": f"第{i}列"}
|
||||
for i in range(1, 25)]}}
|
||||
model = FakeModel(replies, hang=2)
|
||||
job = await _draft(client, sid, monkeypatch, model, ["tickets"])
|
||||
assert job["status"] == "succeeded" and job["tasks"][0]["status"] == "ok"
|
||||
assert job["usage"] == {"calls": 5, "input_tokens": 0, "output_tokens": 0, "timeouts": 2, "failed_calls": 2}
|
||||
assert job["part_columns"] == 20 and job["notice"] is None
|
||||
# 24 列整张问超时 → 前 12 列（带表级的项）超时 → 拆成 6、6 → 后 12 列
|
||||
assert [len(requested(c)["tickets"][1] or []) for c in model.calls] == [0, 12, 6, 6, 12]
|
||||
|
||||
plain = await _draft(client, sid, monkeypatch, FakeModel(replies), ["tickets"])
|
||||
assert plain["part_columns"] == 40 and plain["usage"]["timeouts"] == plain["usage"]["failed_calls"] == 0
|
||||
|
||||
|
||||
async def test_the_notice_is_kept_on_the_job(client, monkeypatch):
|
||||
"""最近 6 次调用有 4 次以上没写出来：任务上写一句提醒（次数、原因、建议换起草模型）；不用模型的任务没有。"""
|
||||
monkeypatch.setattr(catalog, "MODEL_BATCH_TABLES", 1)
|
||||
tables = {f"t{i}": {"qualified": f"t{i}", "columns": [_col("id"), _col("v")], "primary_key": ["id"],
|
||||
"comment": None, "is_view": False, "foreign_keys": []} for i in range(1, 6)}
|
||||
sid = await _source(tables)
|
||||
replies = {name: {"label": f"表{name}"} for name in tables}
|
||||
job = await _draft(client, sid, monkeypatch, FakeModel(replies, fail={"t1", "t2", "t3", "t4"}), list(tables))
|
||||
assert job["usage"]["calls"] == 5 and job["usage"]["failed_calls"] == 4
|
||||
assert job["notice"] == ("最近 5 次调用有 4 次没写出来（出错 4 次）。可以在「设置 → 偏好设置」的「起草模型」换一个模型、"
|
||||
"减少这次起草的表，或者停止这次起草")
|
||||
plain = (await client.post(f"{_base(sid)}/draft-jobs", json={"tables": ["t1"]})).json()["job"]
|
||||
await finished(plain["id"])
|
||||
plain = await get_job(client, sid, plain["id"])
|
||||
assert plain["notice"] is None and plain["part_columns"] is None and plain["blank_columns"] == 0
|
||||
|
||||
|
||||
async def test_blank_columns_reach_the_task_the_job_and_the_list(client, monkeypatch):
|
||||
"""模型列出了、中文名和含义都没给的列：逐表记列数和最多 5 个列名，任务上是合计，目录清单的 last_draft 也带列数。"""
|
||||
tables = {**_tickets(8), "gates": _tables()["gates"]}
|
||||
sid = await _source(tables)
|
||||
replies = {"tickets": {"label": "门票", "columns": [{"name": "c01", "label": "编号"}]},
|
||||
"gates": {"label": "闸口", "columns": [{"name": "id", "label": "编号"}]}}
|
||||
job = await _draft(client, sid, monkeypatch, FakeModel(replies), ["tickets", "gates"])
|
||||
by = {t["table_name"]: t for t in job["tasks"]}
|
||||
assert by["tickets"]["status"] == "ok" and by["tickets"]["blank_columns"] == 7
|
||||
assert by["tickets"]["blank_sample"] == ["c02", "c03", "c04", "c05", "c06"]
|
||||
assert by["gates"]["blank_columns"] == 1 and by["gates"]["blank_sample"] == ["gate_code"]
|
||||
assert job["blank_columns"] == 8
|
||||
rows = {r["table_name"]: r for r in (await client.get(_base(sid))).json()["tables"]}
|
||||
assert rows["tickets"]["last_draft"]["blank_columns"] == 7 and rows["gates"]["last_draft"]["blank_columns"] == 1
|
||||
|
||||
|
||||
def test_old_job_tables_get_the_new_columns():
|
||||
"""升级前建的起草任务表：新列补上，存量行的计数是 0、提醒和列数上限为空。"""
|
||||
from sqlalchemy import create_engine, text
|
||||
|
||||
from app.db.base import _migrate
|
||||
|
||||
db = create_engine("sqlite://")
|
||||
with db.begin() as conn:
|
||||
conn.execute(text("CREATE TABLE catalog_draft_jobs (id VARCHAR(32) PRIMARY KEY, calls INTEGER)"))
|
||||
conn.execute(text("CREATE TABLE catalog_draft_tasks (id VARCHAR(32) PRIMARY KEY, removed INTEGER)"))
|
||||
conn.execute(text("INSERT INTO catalog_draft_jobs (id, calls) VALUES ('j1', 3)"))
|
||||
conn.execute(text("INSERT INTO catalog_draft_tasks (id, removed) VALUES ('t1', 0)"))
|
||||
_migrate(conn)
|
||||
_migrate(conn)
|
||||
assert conn.execute(text("SELECT timeouts, failed_calls, part_columns, notice FROM catalog_draft_jobs")) \
|
||||
.one() == (0, 0, None, None)
|
||||
assert conn.execute(text("SELECT blank_columns, blank_sample FROM catalog_draft_tasks")).one() == (0, None)
|
||||
|
||||
|
||||
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()
|
||||
|
||||
Reference in New Issue
Block a user