From b940b27e242b78e87329e03b69d28b89bd267b41 Mon Sep 17 00:00:00 2001 From: ThearchyHelios Date: Tue, 13 Oct 2026 15:00:38 +0800 Subject: [PATCH 01/29] =?UTF-8?q?=E8=B5=B7=E8=8D=89=E6=A8=A1=E5=9E=8B?= =?UTF-8?q?=E8=AE=BE=E7=BD=AE=EF=BC=9A=E6=96=B0=E5=A2=9E=E8=AE=BE=E7=BD=AE?= =?UTF-8?q?=E7=BB=84=20catalog=5Fdraft=EF=BC=88=E6=8E=A5=E5=85=A5=E3=80=81?= =?UTF-8?q?=E6=A8=A1=E5=9E=8B=EF=BC=89=EF=BC=8C=E7=82=B9=E5=90=8D=E7=9A=84?= =?UTF-8?q?=E6=8E=A5=E5=85=A5=E4=B8=8D=E5=AD=98=E5=9C=A8=E6=88=96=E5=B7=B2?= =?UTF-8?q?=E5=81=9C=E7=94=A8=E6=97=B6=E4=BF=9D=E5=AD=98=E5=9B=9E=20422?= =?UTF-8?q?=EF=BC=9B=E6=95=B0=E6=8D=AE=E7=9B=AE=E5=BD=95=E8=B5=B7=E8=8D=89?= =?UTF-8?q?=E5=92=8C=E6=8C=89=E9=85=8D=E6=96=B9=E5=AF=BC=E5=85=A5=E7=9A=84?= =?UTF-8?q?=20AI=20=E8=B5=B7=E8=8D=89=E5=85=B1=E7=94=A8=E4=B8=80=E5=A4=84?= =?UTF-8?q?=E5=8F=96=E6=A8=A1=E5=9E=8B=EF=BC=88=E8=B5=B7=E8=8D=89=E6=A8=A1?= =?UTF-8?q?=E5=9E=8B=20>=20=E5=8A=A9=E6=89=8B=E7=9A=84=E6=A8=A1=E5=9E=8B?= =?UTF-8?q?=20>=20=E9=BB=98=E8=AE=A4=E6=8E=A5=E5=85=A5=EF=BC=89=EF=BC=8C?= =?UTF-8?q?=E5=8F=96=E8=87=AA=E8=B5=B7=E8=8D=89=E6=A8=A1=E5=9E=8B=E6=97=B6?= =?UTF-8?q?=E6=8A=A5=E9=94=99=E7=85=A7=E5=AE=9E=E8=AF=B4=E8=B5=B7=E8=8D=89?= =?UTF-8?q?=E6=A8=A1=E5=9E=8B=E9=82=A3=E4=B8=80=E9=A1=B9=EF=BC=9B=E6=95=B0?= =?UTF-8?q?=E6=8D=AE=E7=9B=AE=E5=BD=95=E8=B5=B7=E8=8D=89=E4=B8=8D=E5=86=8D?= =?UTF-8?q?=E5=85=B3=E6=80=9D=E8=80=83=EF=BC=8C=E9=A2=9D=E5=BA=A6=2032768?= =?UTF-8?q?=EF=BC=8C=E6=9C=8D=E5=8A=A1=E6=96=B9=E5=AB=8C=E5=A4=A7=E6=97=B6?= =?UTF-8?q?=E9=80=80=E5=9B=9E=208192=EF=BC=88=E4=B9=8B=E5=90=8E=E9=83=BD?= =?UTF-8?q?=E7=94=A8=E5=B0=8F=E9=A2=9D=E5=BA=A6=EF=BC=89=EF=BC=8C=E5=8D=95?= =?UTF-8?q?=E6=AC=A1=E8=B0=83=E7=94=A8=E6=97=B6=E9=99=90=20420=20=E7=A7=92?= =?UTF-8?q?=E4=BA=A4=E7=BB=99=20SDK=EF=BC=9B=E9=85=8D=E6=96=B9=20AI=20?= =?UTF-8?q?=E8=B5=B7=E8=8D=89=E7=9A=84=E9=A2=9D=E5=BA=A6=E3=80=81=E6=80=9D?= =?UTF-8?q?=E8=80=83=E8=AE=BE=E7=BD=AE=E4=B8=8D=E5=8A=A8?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- backend/app/api/settings.py | 45 +++- backend/app/data/catalog.py | 94 ++++--- backend/app/data/draft_model.py | 114 +++++++++ backend/app/data/recipe_ai.py | 47 ++-- backend/tests/test_catalog_draft_settings.py | 247 +++++++++++++++++++ 5 files changed, 487 insertions(+), 60 deletions(-) create mode 100644 backend/app/data/draft_model.py create mode 100644 backend/tests/test_catalog_draft_settings.py diff --git a/backend/app/api/settings.py b/backend/app/api/settings.py index 9a4ee8ad..e1fdcafb 100644 --- a/backend/app/api/settings.py +++ b/backend/app/api/settings.py @@ -14,6 +14,7 @@ from app.api.coded import CodedHTTPException from app.core.errors import explain, not_configured, raw from app.core.config import settings as app_settings from app.core.crypto import encrypt, mask +from app.data.draft_model import CATALOG_DRAFT_SETTING_KEY, setting_pair from app.db.base import get_session from app.db.models import Provider, Setting from app.engine.judge import JUDGE_DEFAULTS, MAX_TOKENS as JUDGE_MAX_TOKENS @@ -569,6 +570,39 @@ async def verify_model_spec(session: AsyncSession, node_config: dict[str, Any] | return ModelSpec(provider=provider, model=model, max_tokens=JUDGE_MAX_TOKENS, thinking="off") +# -------------------------------------------------------------------------- +# 起草模型：数据目录的「用模型起草」和按配方导入的 AI 起草用哪个模型 +# -------------------------------------------------------------------------- + +#: 设置里 catalog_draft 这一组：{provider, model}。都留空时沿用助手的模型，再没有就是默认接入（取值顺序见 +#: data/draft_model.py，两处起草共用）。只放接入和模型：额度、思考由两处起草各自定 +CATALOG_DRAFT_GROUP = CATALOG_DRAFT_SETTING_KEY +CATALOG_DRAFT_DEFAULTS: dict[str, Any] = {"provider": None, "model": None} +#: 设置页「起草模型」一节各项的叫法：报错时写它，不写键名 +CATALOG_DRAFT_LABEL = {"provider": "起草模型 · 接入", "model": "起草模型 · 模型"} + + +def catalog_draft_problem(value: Any) -> str | None: + """设置页写进来的 catalog_draft 这一组有什么问题，没问题返回 None。点名的接入在不在由 _provider_problem 查。""" + if not isinstance(value, dict): + return f"「起草模型」设置格式有误，需要是一个对象，当前为「{value}」" + for key, item in value.items(): + if key not in CATALOG_DRAFT_DEFAULTS: + return (f"「起草模型」设置中不支持「{key}」，可设置的项：" + + "、".join(f"「{CATALOG_DRAFT_LABEL[k]}」" for k in CATALOG_DRAFT_DEFAULTS)) + if item is not None and not isinstance(item, str): + # 「跟随助手的模型」是设置页接入下拉框空选项的原话 + what = "模型接入的名称" if key == "provider" else "模型 ID" + return f"「{CATALOG_DRAFT_LABEL[key]}」需要填写{what}，留空表示跟随助手的模型，当前为「{item}」" + return None + + +def catalog_draft_values(stored: Any) -> dict[str, Any]: + """库里存的 catalog_draft 一组 → 补齐默认、规范化（去掉首尾空白，空串和存坏了的当没填）。""" + provider, model = setting_pair(stored) + return {"provider": provider, "model": model} + + # -------------------------------------------------------------------------- # 通知：定时任务推送 Webhook 时拼运行链接用的对外访问地址 # -------------------------------------------------------------------------- @@ -661,6 +695,8 @@ DEFAULT_SETTINGS: dict[str, Any] = { VERIFY_GROUP: verify_values(None), # 通知：定时任务的 Webhook 里拼运行链接用的对外访问地址 NOTIFY_GROUP: notify_values(None), + # 起草模型：数据目录起草、按配方导入的 AI 起草用的模型，留空沿用助手的模型 + CATALOG_DRAFT_GROUP: catalog_draft_values(None), } @@ -717,6 +753,7 @@ async def get_settings(session: AsyncSession = Depends(get_session)) -> dict[str merged[JUDGE_GROUP] = judge_values(stored.get(JUDGE_GROUP)) merged[VERIFY_GROUP] = verify_values(stored.get(VERIFY_GROUP)) merged[NOTIFY_GROUP] = notify_values(stored.get(NOTIFY_GROUP)) + merged[CATALOG_DRAFT_GROUP] = catalog_draft_values(stored.get(CATALOG_DRAFT_GROUP)) for key, value in stored.items(): merged.setdefault(key, value) return merged @@ -731,8 +768,8 @@ PROVIDER_INVALID = "settings_provider_invalid" def _provider_fields() -> dict[tuple[str, str], tuple[str, str | None]]: - """设置里点名模型接入的项：(组, 键) → (这一项在界面上的叫法, 同一组模型那一项的叫法)。门控、裁判、核查的叫法和 - 设置页一致；助手的模型由「编排」页助手栏的「助手使用的模型」单独设定（PUT /api/copilot/model），只有这一个下拉框 + """设置里点名模型接入的项：(组, 键) → (这一项在界面上的叫法, 同一组模型那一项的叫法)。门控、裁判、核查、起草模型的 + 叫法和设置页一致；助手的模型由「编排」页助手栏的「助手使用的模型」单独设定（PUT /api/copilot/model），只有这一个下拉框 （模型那一项为 None），这里只防经设置接口写进去的那一份。""" from app.api.copilot import COPILOT_SETTING_KEY # 在这里导入：copilot 模块很大，设置模块被引擎各处导入 from app.providers.factory import ASSISTANT_MODEL_FIELD @@ -741,6 +778,7 @@ def _provider_fields() -> dict[tuple[str, str], tuple[str, str | None]]: ("run", "tool_gate_provider"): ("门控模型 · 接入", "门控模型 · 模型"), (JUDGE_GROUP, "provider"): (JUDGE_LABEL["provider"], JUDGE_LABEL["model"]), (VERIFY_GROUP, "provider"): (VERIFY_SETTING_LABEL["provider"], VERIFY_SETTING_LABEL["model"]), + (CATALOG_DRAFT_GROUP, "provider"): (CATALOG_DRAFT_LABEL["provider"], CATALOG_DRAFT_LABEL["model"]), (COPILOT_SETTING_KEY, "provider"): (ASSISTANT_MODEL_FIELD, None), } @@ -798,6 +836,9 @@ async def put_settings( raise HTTPException(422, problem) if VERIFY_GROUP in payload.values and (problem := verify_problem(payload.values[VERIFY_GROUP])): raise HTTPException(422, problem) + if CATALOG_DRAFT_GROUP in payload.values and ( + problem := catalog_draft_problem(payload.values[CATALOG_DRAFT_GROUP])): + raise HTTPException(422, problem) if NOTIFY_GROUP in payload.values: if problem := notify_problem(payload.values[NOTIFY_GROUP]): raise HTTPException(422, problem) diff --git a/backend/app/data/catalog.py b/backend/app/data/catalog.py index fb7e7063..281633a7 100644 --- a/backend/app/data/catalog.py +++ b/backend/app/data/catalog.py @@ -1401,15 +1401,25 @@ def draft_structure(source: Any, table: str) -> TableDraft: # 起草：模型（可选，按表分批，结构化输出） # ========================================================================== -#: 每批最多几张表、合计多少列。一批太大模型容易漏表、输出被截断；太小则调用次数多 +#: 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 MODEL_BATCH_TABLES = 4 -MODEL_BATCH_COLUMNS = 120 -#: 同时发出的批数 +MODEL_BATCH_COLUMNS = 40 +#: 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 -#: 一批最多等多久 -MODEL_TIMEOUT_S = 120 -#: 输出额度。一批四张表的目录远用不完；给太大的额度有的服务方会直接拒收 -MODEL_MAX_TOKENS = 8192 +#: 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 +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） +MODEL_MAX_TOKENS = 32768 +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 +#: 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） +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 DRAFT_DEFAULT_TABLES = 20 @@ -1502,38 +1512,60 @@ CATALOG_DRAFT_PROMPT_SCHEMA: dict[str, Any] = { _P = CATALOG_DRAFT_PROMPT_LABELS +class DraftModel: + """起草用的模型：先按 MODEL_MAX_TOKENS 的额度调；服务方嫌额度太大（budget_rejected）时换成 MODEL_FALLBACK_MAX_TOKENS + 的那一个，之后的调用都用它（shrink）。一次起草（一个起草任务、一次同步起草）共用一个，退回之后不再先失败一次。 + + 只转发起草用到的两个方法：with_structured_output 每次都取当前那个模型的，ainvoke 同理。""" + + def __init__(self, big: Any, small: Any | None, model_name: str = "") -> None: + self.model, self._small = big, small + self.model_name = model_name + self.budget = MODEL_MAX_TOKENS + + def shrink(self) -> bool: + """换成小额度的模型。已经换过、没有小的，返回 False。""" + if self._small is None or self.model is self._small: + return False + self.model, self.budget = self._small, MODEL_FALLBACK_MAX_TOKENS + return True + + def with_structured_output(self, schema: Any, **kwargs: Any) -> Any: + return self.model.with_structured_output(schema, **kwargs) + + async def ainvoke(self, messages: list[Any]) -> Any: + return await self.model.ainvoke(messages) + + async def resolve_draft_model(session: AsyncSession) -> tuple[Any, str]: - """(模型, 模型 id)：用设置里助手的模型，关掉思考、温度 0。没有可用的模型抛 CatalogModelUnavailable。 + """(模型, 模型 id)：用哪个接入、哪个模型按 data/draft_model.py 的顺序（设置里的起草模型 > 助手的模型 > 默认接入）， + 和按配方导入的 AI 起草共用。没有可用的模型抛 CatalogModelUnavailable。 + + 不关思考（spec 里不传 thinking，按接入默认）、温度 0、额度 MODEL_MAX_TOKENS，服务方嫌大时退回 + MODEL_FALLBACK_MAX_TOKENS（DraftModel）。单次调用的时限 MODEL_TIMEOUT_S 也交给 SDK，免得它先按全局的 5 分钟断开。 + 只建模型对象，不发请求。 和 recipe_ai.resolve_draft_model 一样不用 get_chat_model：它在没有接入时静默退回演示模型，拿演示模型 - 起草出来的目录会被当成真的写进库里。数据层不导入接口层，设置键取 recipe_ai 里那份（与助手的同值， - 那边有测试核对）。 + 起草出来的目录会被当成真的写进库里。 """ - from app.data.recipe_ai import COPILOT_SETTING_KEY - from app.db.models import Setting - from app.providers.factory import ( - ModelSpec, ProviderNotConfigured, assistant_unavailable, build_chat_model, resolve_provider, - ) + from app.data.draft_model import DraftModelUnavailable, draft_unavailable, resolve_draft_provider + from app.providers.factory import ModelSpec, ProviderNotConfigured, build_chat_model - row = await session.get(Setting, COPILOT_SETTING_KEY) - saved = (row.value if row else None) or {} - spec = ModelSpec(provider=saved.get("provider") or None, model=saved.get("model") or None, - max_tokens=MODEL_MAX_TOKENS, temperature=0, thinking="off") try: - # 存的接入名不存在、已停用，只存了模型名、提供它的接入已停用：resolve_provider 都会报出来 - provider = await resolve_provider(session, spec.provider, spec.model) - except ProviderNotConfigured as e: - raise CatalogModelUnavailable(assistant_unavailable(e)) from e - if provider is None: - raise CatalogModelUnavailable("未配置模型接入，无法由助手起草数据目录。请到「设置 → 模型接入」添加") - if provider.kind == "mock": - raise CatalogModelUnavailable("当前只有演示用的模型接入，无法用于起草数据目录。" - "请到「设置 → 模型接入」添加真实的模型接入") + provider, choice = await resolve_draft_provider( + session, none_reason="未配置模型接入，无法由模型起草数据目录。请到「设置 → 模型接入」添加", + mock_reason="当前只有演示用的模型接入，无法用于起草数据目录。请到「设置 → 模型接入」添加真实的模型接入") + except DraftModelUnavailable as e: + raise CatalogModelUnavailable(str(e)) from e + spec = ModelSpec(provider=choice.provider, model=choice.model, max_tokens=MODEL_MAX_TOKENS, temperature=0, + timeout=MODEL_TIMEOUT_S) try: - model = build_chat_model(provider, spec) + big = build_chat_model(provider, spec) + small = build_chat_model(provider, spec.model_copy(update={"max_tokens": MODEL_FALLBACK_MAX_TOKENS})) except ProviderNotConfigured as e: - raise CatalogModelUnavailable(assistant_unavailable(e)) from e - return model, spec.model or provider.default_model or "" + raise CatalogModelUnavailable(draft_unavailable(e, choice.origin)) from e + model_id = spec.model or provider.default_model or "" + return DraftModel(big, small, model_id), model_id def _confirmed_only(notes: Mapping[str, Any] | None) -> dict[str, Any]: diff --git a/backend/app/data/draft_model.py b/backend/app/data/draft_model.py new file mode 100644 index 00000000..da3be8be --- /dev/null +++ b/backend/app/data/draft_model.py @@ -0,0 +1,114 @@ +"""起草用哪个模型：数据目录的「用模型起草」和按配方导入的 AI 起草共用这一处取值。 + +顺序（哪一级写了就整组用哪一级，不跨级拼接入和模型）： + +1. 设置里的起草模型（catalog_draft 一组：{provider, model}），填了接入或模型就用它； +2. 否则沿用助手的模型设置（copilot 一组，「编排」页助手栏的「助手使用的模型」）； +3. 再没有就是默认接入（providers/factory.resolve_provider 的回退）。 + +这里只决定「用哪个接入、哪个模型」。额度、思考、温度由两边各自定：数据目录起草保留思考、额度 32768 +（data/catalog.resolve_draft_model），配方 AI 起草照旧关思考（data/recipe_ai.resolve_draft_model）。 + +**为什么不用 get_chat_model。** 它在没有接入时静默退回演示模型，拿演示模型起草出来的东西会被当成真的。 +两边都自己解析接入：没有、停用、只有演示接入、缺 key，都在调用模型之前判定为不可用，说清是哪一级的设置出了问题。 + +数据层不导入接口层：设置键在这里定义，api/settings.py 从这里取（助手那一项和 api/copilot.py 同值，有测试核对）。 +""" +from __future__ import annotations + +from dataclasses import dataclass +from typing import Any + +from sqlalchemy.ext.asyncio import AsyncSession + +#: 设置里起草模型那一组的键 +CATALOG_DRAFT_SETTING_KEY = "catalog_draft" +#: 助手的模型设置键（与 api/copilot.py 的 COPILOT_SETTING_KEY 同值） +COPILOT_SETTING_KEY = "copilot" +#: 起草模型在设置页上的位置。报错里让人去这里改；和前端设置页放这一组的地方一致 +DRAFT_MODEL_PLACE = "「设置 → 偏好设置」的「起草模型」" + +#: 取自哪一级：draft 设置里的起草模型、assistant 助手的模型、default 默认接入 +ORIGINS = ("draft", "assistant", "default") + + +@dataclass(frozen=True) +class DraftModelChoice: + """起草用的接入和模型（都可以为 None：接入为 None 时按模型名找、再退回默认接入；模型为 None 时用接入的默认模型）。""" + + provider: str | None + model: str | None + origin: str + + +def _text(value: Any) -> str | None: + """设置里的一项：字符串去掉首尾空白，空串和别的类型（存坏了的）都当没填。""" + return value.strip() or None if isinstance(value, str) else None + + +def setting_pair(stored: Any) -> tuple[str | None, str | None]: + """库里存的一组 → (接入, 模型)，规范化同 _text。""" + stored = stored if isinstance(stored, dict) else {} + return _text(stored.get("provider")), _text(stored.get("model")) + + +async def draft_model_choice(session: AsyncSession) -> DraftModelChoice: + """按上面的顺序取。只读设置，不查接入在不在（resolve_provider 去查）。""" + from app.db.models import Setting + + for key, origin in ((CATALOG_DRAFT_SETTING_KEY, "draft"), (COPILOT_SETTING_KEY, "assistant")): + row = await session.get(Setting, key) + provider, model = setting_pair(row.value if row is not None else None) + if provider or model: + return DraftModelChoice(provider, model, origin) + return DraftModelChoice(None, None, "default") + + +def draft_unavailable(e: BaseException, origin: str) -> str: + """取到的接入用不了（点名的不存在、已停用，缺 key、没有模型……）时给人看的那句话。 + + 沿用助手的模型（或默认接入）时，和助手取不到模型时同一句（factory.assistant_unavailable）；用的是设置里的起草 + 模型时，照实说是起草模型那一项、去设置页哪里改。""" + from app.core.errors import not_configured + from app.providers.factory import NAMED_FIX, assistant_unavailable, model_fact + + if origin != "draft": + return assistant_unavailable(e) + code = getattr(e, "code", "") + if code in NAMED_FIX: + said = f"起草模型使用的{getattr(e, 'reason', '') or e}。" + if owner := getattr(e, "model_owner", ""): + return (f"{said}{NAMED_FIX[code]}；{model_fact(getattr(e, 'name', ''), owner)}，" + f"也可在{DRAFT_MODEL_PLACE}中把它填进模型一栏") + return f"{said}{NAMED_FIX[code]}，或在{DRAFT_MODEL_PLACE}中重新选择" + return (f"起草模型尚未配置完成：{not_configured(e, with_hint=False)}。请到「设置 → 模型接入」检查，" + f"或在{DRAFT_MODEL_PLACE}中重新选择") + + +class DraftModelUnavailable(RuntimeError): + """没有可用于起草的模型。message 给人看；在调用模型之前判定。""" + + +async def resolve_draft_provider(session: AsyncSession, *, none_reason: str, mock_reason: str) -> tuple[Any, DraftModelChoice]: + """(接入, 取值)。用不了时抛 DraftModelUnavailable：没有任何启用的接入说 none_reason，只有演示接入说 mock_reason + （两边各有各的说法），点名的接入不存在、已停用按取自哪一级说（draft_unavailable）。 + + 只解析接入，不建模型对象：建的时候缺 key 之类的问题由调用方经 draft_unavailable 说。""" + from app.providers.factory import ProviderNotConfigured, resolve_provider + + choice = await draft_model_choice(session) + try: + # 存的接入名不存在、已停用，只存了模型名、提供它的接入已停用：resolve_provider 都会报出来 + provider = await resolve_provider(session, choice.provider, choice.model) + except ProviderNotConfigured as e: + raise DraftModelUnavailable(draft_unavailable(e, choice.origin)) from e + if provider is None: + raise DraftModelUnavailable(none_reason) + if provider.kind == "mock": + raise DraftModelUnavailable(mock_reason) + return provider, choice + + +__all__ = ["CATALOG_DRAFT_SETTING_KEY", "COPILOT_SETTING_KEY", "DRAFT_MODEL_PLACE", "DraftModelChoice", + "DraftModelUnavailable", "draft_model_choice", "draft_unavailable", "resolve_draft_provider", + "setting_pair"] diff --git a/backend/app/data/recipe_ai.py b/backend/app/data/recipe_ai.py index 44ec1f92..5cb8b42f 100644 --- a/backend/app/data/recipe_ai.py +++ b/backend/app/data/recipe_ai.py @@ -6,7 +6,8 @@ 必须人工确认之后才能启用。每月重传按已确认的配方重放，不再调用模型。 **为什么不用 get_chat_model。** 它在没有接入时会静默退回演示模型；拿演示模型起草等于假装起草了。这里 -自己按助手的模型设置解析接入，没有、停用、只有演示接入、缺 key 都在调用模型之前判定为不可用。 +自己解析接入（取值顺序和数据目录起草共用 data/draft_model.py：设置里的起草模型 > 助手的模型 > 默认接入）， +没有、停用、只有演示接入、缺 key 都在调用模型之前判定为不可用。 **调用参数照 c837cdb 修过的问题来**：关思考（thinking="off"）、额度 32768、服务方嫌额度大时退回 8192； 正文一律经 message_text 抽取（回复可能是块列表）；输出被截断单独记一种失败，不当成「不是 JSON」去修订。 @@ -42,6 +43,7 @@ from app.data.recipe_types import ( GridCell, Recipe, ) +from app.data.draft_model import COPILOT_SETTING_KEY as _DRAFT_COPILOT_KEY from app.data.tabular import EXCEL_ERRORS from app.data.xlsx_scan import WorkbookScan, cell_ref, col_letter @@ -58,8 +60,9 @@ class AiTooLarge(ValueError): # 没有 AiFailed：调用之后的任何失败都体现在返回值里（draft=None、error、usage 照记） -#: 助手的模型设置键（与 app/api/copilot.py 的 COPILOT_SETTING_KEY 同值；数据层不导入接口层，测试断言两者相等） -COPILOT_SETTING_KEY = "copilot" +#: 助手的模型设置键（与 app/api/copilot.py 的 COPILOT_SETTING_KEY 同值；数据层不导入接口层，测试断言两者相等）。 +#: 取模型时由 data/draft_model.py 读，这里留着这个名字给测试和老代码用 +COPILOT_SETTING_KEY = _DRAFT_COPILOT_KEY #: 1 次起草 + 最多 2 次修订 AI_MAX_ATTEMPTS = 3 #: 单次调用的超时（秒）；测试 monkeypatch 调小 @@ -252,37 +255,27 @@ class _BudgetFallback: return await self.model.ainvoke(messages) -def _unconfigured(e: BaseException) -> str: - """和 copilot._unconfigured 同一句（providers/factory.assistant_unavailable）。""" - from app.providers.factory import assistant_unavailable - - return assistant_unavailable(e) - - async def resolve_draft_model(session: AsyncSession) -> tuple[Any | None, str, AiAvailability]: - """(可调用的模型, 模型 id, 可用性)；不可用时模型为 None。只建模型对象，不发请求。""" - from app.db.models import Setting - from app.providers.factory import ModelSpec, ProviderNotConfigured, build_chat_model, resolve_provider + """(可调用的模型, 模型 id, 可用性)；不可用时模型为 None。只建模型对象，不发请求。 + + 用哪个接入、哪个模型和数据目录起草共用一处取值（data/draft_model.py：起草模型 > 助手的模型 > 默认接入）； + 额度、关思考照旧是这里自己的。""" + from app.data.draft_model import DraftModelUnavailable, draft_unavailable, resolve_draft_provider + from app.providers.factory import ModelSpec, ProviderNotConfigured, build_chat_model - row = await session.get(Setting, COPILOT_SETTING_KEY) - saved = (row.value if row else None) or {} - spec = ModelSpec(provider=saved.get("provider") or None, model=saved.get("model") or None, - max_tokens=AI_MAX_TOKENS, temperature=0, thinking="off") try: - provider = await resolve_provider(session, spec.provider, spec.model) - except ProviderNotConfigured as e: - # 存的接入名不存在、已停用，只存了模型名、提供它的接入已停用：resolve_provider 都会报出来 - return None, "", AiAvailability(False, reason=_unconfigured(e)) - if provider is None: - return None, "", AiAvailability(False, reason="未配置模型接入，无法使用 AI 起草。请到「设置 → 模型接入」添加") - if provider.kind == "mock": - return None, "", AiAvailability(False, reason="当前只有演示用的模型接入，无法用于 AI 起草。" - "请到「设置 → 模型接入」添加真实的模型接入") + provider, choice = await resolve_draft_provider( + session, none_reason="未配置模型接入，无法使用 AI 起草。请到「设置 → 模型接入」添加", + mock_reason="当前只有演示用的模型接入，无法用于 AI 起草。请到「设置 → 模型接入」添加真实的模型接入") + except DraftModelUnavailable as e: + return None, "", AiAvailability(False, reason=str(e)) + spec = ModelSpec(provider=choice.provider, model=choice.model, max_tokens=AI_MAX_TOKENS, temperature=0, + thinking="off") try: big = build_chat_model(provider, spec) small = build_chat_model(provider, spec.model_copy(update={"max_tokens": AI_FALLBACK_MAX_TOKENS})) except ProviderNotConfigured as e: - return None, "", AiAvailability(False, reason=_unconfigured(e)) + return None, "", AiAvailability(False, reason=draft_unavailable(e, choice.origin)) model_id = spec.model or provider.default_model or "" return (_BudgetFallback(big, small, model_id), model_id, AiAvailability(True, model=model_id, provider=provider.name)) diff --git a/backend/tests/test_catalog_draft_settings.py b/backend/tests/test_catalog_draft_settings.py new file mode 100644 index 00000000..c4cbbec4 --- /dev/null +++ b/backend/tests/test_catalog_draft_settings.py @@ -0,0 +1,247 @@ +"""起草模型（设置里的 catalog_draft 一组）与起草取模型的顺序。 + +- 设置组 catalog_draft：{provider, model}，经 PUT /api/settings 写、GET 读，没存过补 {provider: null, model: null}； + provider 不是字符串、点名的接入不存在或已停用 → 422，说法写设置页上的叫法「起草模型 · 接入」「起草模型 · 模型」 +- 取模型的顺序（数据目录起草和按配方导入的 AI 起草共用 draft_model.draft_model_choice）：起草模型填了接入或模型就整组 + 用它；否则沿用助手的模型设置；再没有就是默认接入。不跨级拼接入和模型 +- 数据目录起草不关思考（按接入默认）、额度 32768，服务方嫌大退回 8192；配方 AI 起草只改取模型的顺序，额度、思考不动 + +接入名、模型名都是虚构的。 +""" +from __future__ import annotations + +from typing import Any + +import pytest +from httpx import ASGITransport, AsyncClient +from sqlalchemy import delete + +from app.data import catalog, draft_model +from app.data import recipe_ai as A +from app.db.base import SessionLocal +from app.db.models import Provider, Setting +from app.main import app + +LIVE, OFF, GONE = "起草用接入", "起草用停用接入", "起草用不存在的接入" + + +async def _clean() -> None: + async with SessionLocal() as session: + await session.execute(delete(Setting).where(Setting.key.in_([draft_model.CATALOG_DRAFT_SETTING_KEY]))) + await session.execute(delete(Provider).where(Provider.name.in_([LIVE, OFF]))) + await session.commit() + + +@pytest.fixture +async def providers(): + """一个启用、一个停用的演示类接入（不联网）；用完删掉。""" + await _clean() + async with SessionLocal() as session: + session.add_all([ + Provider(name=LIVE, kind="mock", models=[{"id": "起草模型甲"}], default_model="起草模型甲", enabled=True, + extra={}), + Provider(name=OFF, kind="mock", models=[{"id": "起草模型乙"}], default_model="起草模型乙", enabled=False, + extra={}), + ]) + await session.commit() + yield + await _clean() + + +@pytest.fixture +async def client(): + async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c: + yield c + + +# -------------------------------------------------------------------------- +# 设置组的读写与校验 +# -------------------------------------------------------------------------- + + +async def test_the_group_is_exposed_with_empty_defaults(client, providers): + got = (await client.get("/api/settings")).json() + assert got["catalog_draft"] == {"provider": None, "model": None} + + +async def test_the_group_round_trips(client, providers): + r = await client.put("/api/settings", json={"values": {"catalog_draft": {"provider": LIVE, "model": "起草模型甲"}}}) + assert r.status_code == 200, r.text + assert r.json()["catalog_draft"] == {"provider": LIVE, "model": "起草模型甲"} + got = (await client.get("/api/settings")).json()["catalog_draft"] + assert got == {"provider": LIVE, "model": "起草模型甲"} + # 只写模型、接入留空：照存，接入是 null + r = await client.put("/api/settings", json={"values": {"catalog_draft": {"provider": " ", "model": "x"}}}) + assert r.status_code == 200, r.text + assert r.json()["catalog_draft"] == {"provider": None, "model": "x"} + + +@pytest.mark.parametrize("value,label", [ + ({"provider": 3}, "起草模型 · 接入"), + ({"model": ["a"]}, "起草模型 · 模型"), + ({"temperature": 0}, "temperature"), + ("起草模型甲", "起草模型"), +]) +async def test_bad_values_are_refused_with_the_settings_page_wording(client, providers, value, label): + r = await client.put("/api/settings", json={"values": {"catalog_draft": value, "ui": {"theme": "dark"}}}) + assert r.status_code == 422, r.text + assert label in r.json()["detail"] + got = (await client.get("/api/settings")).json() + assert got["catalog_draft"] == {"provider": None, "model": None} and got["ui"]["theme"] != "dark" + + +@pytest.mark.parametrize("name,reason,fix", [ + (GONE, f"模型接入「{GONE}」不存在", "请在「设置 → 模型接入」中核对名称"), + (OFF, f"模型接入「{OFF}」已停用", "请在「设置 → 模型接入」中启用它"), +]) +async def test_a_missing_or_disabled_provider_is_refused_with_the_field(client, providers, name, reason, fix): + r = await client.put("/api/settings", json={"values": {"catalog_draft": {"provider": name}}}) + assert r.status_code == 422, r.text + body = r.json() + assert body["code"] == "settings_provider_invalid" and body["field"] == "catalog_draft.provider" + assert body["detail"] == f"「起草模型 · 接入」指定的{reason}。{fix}，或换一个接入" + async with SessionLocal() as session: + assert await session.get(Setting, "catalog_draft") is None + + +# -------------------------------------------------------------------------- +# 取模型的顺序 +# -------------------------------------------------------------------------- + + +class _MemDB: + """独立的内存库：取模型只看这里的接入和设置，不受别的测试建的接入影响。""" + + async def __aenter__(self): + from sqlalchemy.ext.asyncio import AsyncSession, async_sessionmaker, create_async_engine + from sqlalchemy.pool import StaticPool + + import app.db.models # noqa: F401 + from app.db.base import Base + + self.engine = create_async_engine("sqlite+aiosqlite:///:memory:", poolclass=StaticPool) + async with self.engine.begin() as conn: + await conn.run_sync(Base.metadata.create_all) + self.session = async_sessionmaker(self.engine, expire_on_commit=False, class_=AsyncSession)() + return self.session + + async def __aexit__(self, *exc): + await self.session.close() + await self.engine.dispose() + + +@pytest.fixture +async def mem(): + async with _MemDB() as session: + yield session + + +@pytest.fixture +def no_env_keys(monkeypatch): + for k in ("ANTHROPIC_API_KEY", "ANTHROPIC_AUTH_TOKEN", "OPENAI_API_KEY", "ANTHROPIC_BASE_URL"): + monkeypatch.delenv(k, raising=False) + + +async def _setup(session, *, draft: dict | None = None, copilot: dict | None = None) -> None: + from app.core.crypto import encrypt + + for name, model in (("甲接入", "模型甲"), ("乙接入", "模型乙"), ("丙接入", "模型丙")): + session.add(Provider(name=name, kind="anthropic", api_key=encrypt("sk-test"), models=[{"id": model}], + default_model=model, enabled=True, extra={})) + if draft is not None: + session.add(Setting(key=draft_model.CATALOG_DRAFT_SETTING_KEY, value=draft)) + if copilot is not None: + session.add(Setting(key=draft_model.COPILOT_SETTING_KEY, value=copilot)) + await session.commit() + + +@pytest.mark.parametrize("draft,copilot,expected", [ + (None, None, (None, None, "default")), + (None, {"provider": "乙接入", "model": ""}, ("乙接入", None, "assistant")), + ({"provider": None, "model": None}, {"provider": "", "model": "模型乙"}, (None, "模型乙", "assistant")), + ({"provider": "", "model": "模型丙"}, {"provider": "乙接入", "model": "模型乙"}, (None, "模型丙", "draft")), + ({"provider": "甲接入", "model": None}, {"provider": "乙接入", "model": "模型乙"}, ("甲接入", None, "draft")), + ({"provider": 7}, {"provider": "乙接入"}, ("乙接入", None, "assistant")), # 存坏了的按没填 +]) +async def test_choice_order(mem, draft, copilot, expected): + await _setup(mem, draft=draft, copilot=copilot) + got = await draft_model.draft_model_choice(mem) + assert (got.provider, got.model, got.origin) == expected + + +def _spy(monkeypatch) -> list[Any]: + from app.providers import factory + + specs: list[Any] = [] + + class Built: + def __init__(self, provider, spec): + self.provider, self.spec = provider.name, spec + + def fake_build(provider, spec): + specs.append((provider.name, spec)) + return Built(provider, spec) + + monkeypatch.setattr(factory, "build_chat_model", fake_build) + return specs + + +async def test_catalog_draft_uses_the_draft_setting_keeps_thinking_and_a_big_budget(mem, no_env_keys, monkeypatch): + await _setup(mem, draft={"provider": "甲接入", "model": None}, copilot={"provider": "乙接入", "model": "模型乙"}) + specs = _spy(monkeypatch) + model, model_id = await catalog.resolve_draft_model(mem) + assert model_id == "模型甲" + assert [p for p, _ in specs] == ["甲接入", "甲接入"] + assert [s.max_tokens for _, s in specs] == [catalog.MODEL_MAX_TOKENS, catalog.MODEL_FALLBACK_MAX_TOKENS] + assert catalog.MODEL_MAX_TOKENS == 32768 and catalog.MODEL_FALLBACK_MAX_TOKENS == 8192 + assert all(s.thinking is None and s.temperature == 0 for _, s in specs), "不关思考，按接入默认" + assert all(s.timeout == catalog.MODEL_TIMEOUT_S == 420 for _, s in specs) + assert model.budget == 32768 and model.model_name == "模型甲" + + +async def test_catalog_draft_falls_back_to_the_assistant_model(mem, no_env_keys, monkeypatch): + await _setup(mem, copilot={"provider": "", "model": "模型乙"}) + specs = _spy(monkeypatch) + _, model_id = await catalog.resolve_draft_model(mem) + assert model_id == "模型乙" and specs[0][0] == "乙接入" + + +async def test_recipe_ai_draft_follows_the_same_order_but_keeps_its_own_spec(mem, no_env_keys, monkeypatch): + await _setup(mem, draft={"provider": "", "model": "模型丙"}, copilot={"provider": "乙接入", "model": "模型乙"}) + specs = _spy(monkeypatch) + model, model_id, avail = await A.resolve_draft_model(mem) + assert avail.available and model_id == "模型丙" and avail.provider == "丙接入" + assert [s.max_tokens for _, s in specs] == [A.AI_MAX_TOKENS, A.AI_FALLBACK_MAX_TOKENS] + assert all(s.thinking == "off" and s.temperature == 0 for _, s in specs), "配方 AI 起草的思考设置不动" + + +async def test_a_broken_draft_setting_says_where_to_fix_it(mem, no_env_keys): + """起草模型点名的接入不存在：报错说的是起草模型，不是助手的模型。沿用助手模型时照旧说助手。""" + from app.core.crypto import encrypt + + mem.add(Provider(name="停用的接入", kind="anthropic", api_key=encrypt("k"), models=[{"id": "模型丁"}], + default_model="模型丁", enabled=False, extra={})) + mem.add(Setting(key=draft_model.CATALOG_DRAFT_SETTING_KEY, value={"provider": "停用的接入"})) + await mem.commit() + with pytest.raises(catalog.CatalogModelUnavailable) as e: + await catalog.resolve_draft_model(mem) + assert "起草模型" in str(e.value) and "停用的接入" in str(e.value) and "已停用" in str(e.value) + assert "助手" not in str(e.value) + _, _, avail = await A.resolve_draft_model(mem) + assert not avail.available and "起草模型" in avail.reason and "助手" not in avail.reason + + await mem.execute(delete(Setting)) + mem.add(Setting(key=draft_model.COPILOT_SETTING_KEY, value={"provider": "停用的接入"})) + await mem.commit() + with pytest.raises(catalog.CatalogModelUnavailable) as e: + await catalog.resolve_draft_model(mem) + assert "助手使用的" in str(e.value) + + +async def test_only_a_demo_provider_is_refused(mem, no_env_keys): + mem.add(Provider(name="演示", kind="mock", models=[{"id": "mock-fast"}], default_model="mock-fast", enabled=True, + extra={})) + await mem.commit() + with pytest.raises(catalog.CatalogModelUnavailable) as e: + await catalog.resolve_draft_model(mem) + assert "演示" in str(e.value) From 6b1f63a9dcf3cbbbd3043f6c9256c2c8f683c08e Mon Sep 17 00:00:00 2001 From: ThearchyHelios Date: Tue, 13 Oct 2026 15:19:06 +0800 Subject: [PATCH 02/29] =?UTF-8?q?=E6=95=B0=E6=8D=AE=E7=9B=AE=E5=BD=95?= =?UTF-8?q?=E6=A8=A1=E5=9E=8B=E8=B5=B7=E8=8D=89=E6=8D=A2=E6=88=90=E5=88=86?= =?UTF-8?q?=E6=AE=B5=E3=80=81=E6=8C=89=E5=A4=B1=E8=B4=A5=E5=8E=9F=E5=9B=A0?= =?UTF-8?q?=E5=A4=84=E7=90=86=E7=9A=84=E5=BC=95=E6=93=8E=EF=BC=88data/cata?= =?UTF-8?q?log=5Fdraft.py=EF=BC=89=EF=BC=9A=E6=AF=8F=E6=AC=A1=E6=9C=80?= =?UTF-8?q?=E5=A4=9A=204=20=E5=BC=A0=E8=A1=A8=E3=80=8140=20=E5=88=97?= =?UTF-8?q?=EF=BC=8C=E7=AA=84=E8=A1=A8=E6=8C=89=E9=A1=BA=E5=BA=8F=E6=8B=BC?= =?UTF-8?q?=E6=AE=B5=EF=BC=8C=E5=AE=BD=E8=A1=A8=E6=8B=86=E6=88=90=E3=80=8C?= =?UTF-8?q?=E8=A1=A8=E7=BA=A7=E4=B8=80=E6=AE=B5=20+=20=E6=AF=8F=2040=20?= =?UTF-8?q?=E5=88=97=E4=B8=80=E6=AE=B5=E3=80=8D=EF=BC=8C=E5=88=97=E5=88=86?= =?UTF-8?q?=E6=AE=B5=E5=B8=A6=E7=9D=80=E8=A1=A8=E6=B3=A8=E9=87=8A=E5=92=8C?= =?UTF-8?q?=E8=A1=A8=E7=BA=A7=E8=B5=B7=E8=8D=89=E7=BB=93=E6=9E=9C=E5=81=9A?= =?UTF-8?q?=E8=83=8C=E6=99=AF=EF=BC=8C=E6=9C=80=E5=90=8E=E5=90=88=E5=B9=B6?= =?UTF-8?q?=E7=85=A7=E6=97=A7=E9=80=90=E9=A1=B9=E6=A0=A1=E9=AA=8C=EF=BC=9B?= =?UTF-8?q?=E6=80=9D=E8=80=83=E5=92=8C=E8=BE=93=E5=87=BA=E7=94=A8=E5=AE=8C?= =?UTF-8?q?=E9=A2=9D=E5=BA=A6=E3=80=81=E5=8F=AA=E7=BB=99=E7=A9=BA=E5=A3=B3?= =?UTF-8?q?=E6=97=B6=E6=8B=86=E6=88=90=E4=B8=A4=E5=8D=8A=E5=86=8D=E9=97=AE?= =?UTF-8?q?=EF=BC=88=E6=9C=80=E5=B0=91=205=20=E5=88=97=E3=80=81=E6=9C=80?= =?UTF-8?q?=E5=A4=9A=203=20=E5=B1=82=EF=BC=89=EF=BC=8C=E6=8B=86=E4=B8=8D?= =?UTF-8?q?=E5=BC=80=E7=9A=84=E7=A9=BA=E5=A3=B3=E6=8D=A2=E7=BA=AF=E6=96=87?= =?UTF-8?q?=E6=9C=AC=E9=97=AE=E4=B8=80=E6=AC=A1=EF=BC=8C=E5=86=99=E5=87=BA?= =?UTF-8?q?=E6=9D=A5=E4=BA=86=E5=90=8E=E9=9D=A2=E9=83=BD=E8=B5=B0=E7=BA=AF?= =?UTF-8?q?=E6=96=87=E6=9C=AC=EF=BC=9B=E8=A2=AB=E6=88=AA=E6=96=AD=E7=9A=84?= =?UTF-8?q?=20JSON=20=E5=85=88=E6=8D=A1=E5=9B=9E=E5=B7=B2=E5=AE=8C?= =?UTF-8?q?=E6=95=B4=E7=9A=84=E8=A1=A8=E5=92=8C=E5=88=97=E3=80=81=E5=8F=AA?= =?UTF-8?q?=E8=A1=A5=E5=89=A9=E4=B8=8B=E7=9A=84=EF=BC=9B=E7=BC=BA=E4=BA=86?= =?UTF-8?q?=E7=9A=84=E8=A1=A8=E5=92=8C=E5=88=97=E5=8F=AA=E8=A1=A5=E7=BC=BA?= =?UTF-8?q?=E7=9A=84=E3=80=81=E8=A1=A5=E4=B8=80=E6=AC=A1=EF=BC=9B=E6=9C=8D?= =?UTF-8?q?=E5=8A=A1=E6=96=B9=E5=AB=8C=E9=A2=9D=E5=BA=A6=E5=A4=A7=E6=97=B6?= =?UTF-8?q?=E9=80=80=E5=9B=9E=208192=E3=80=81=E7=BB=93=E6=9E=84=E5=8C=96?= =?UTF-8?q?=E8=BE=93=E5=87=BA=E4=B8=8D=E6=94=AF=E6=8C=81=E6=97=B6=E6=94=B9?= =?UTF-8?q?=E8=B5=B0=E7=BA=AF=E6=96=87=E6=9C=AC=EF=BC=8C=E9=83=BD=E7=AE=A1?= =?UTF-8?q?=E5=88=B0=E8=BF=99=E6=AC=A1=E8=B5=B7=E8=8D=89=E7=BB=93=E6=9D=9F?= =?UTF-8?q?=EF=BC=9B=E7=BD=91=E7=BB=9C=E9=94=99=E8=AF=AF=E3=80=81429?= =?UTF-8?q?=E3=80=815xx=E3=80=81=E8=B6=85=E6=97=B6=E9=80=80=E9=81=BF?= =?UTF-8?q?=E9=87=8D=E8=AF=95=E6=9C=80=E5=A4=9A=202=20=E6=AC=A1=EF=BC=9B?= =?UTF-8?q?=E5=8D=95=E6=AC=A1=E8=B0=83=E7=94=A8=20420=20=E7=A7=92=EF=BC=9B?= =?UTF-8?q?=E5=85=A8=E8=BF=9B=E7=A8=8B=E6=9C=80=E5=A4=9A=203=20=E4=B8=AA?= =?UTF-8?q?=E8=B0=83=E7=94=A8=E5=90=8C=E6=97=B6=E8=BF=9B=E8=A1=8C=EF=BC=9B?= =?UTF-8?q?=E6=8A=A5=E9=94=99=E5=86=99=E5=8E=9F=E5=9B=A0=E5=92=8C=E5=8A=9E?= =?UTF-8?q?=E6=B3=95=EF=BC=88=E7=AC=AC=E5=87=A0=E6=AE=B5=E3=80=81=E5=85=B1?= =?UTF-8?q?=E5=87=A0=E5=88=97=E6=B2=A1=E5=86=99=E5=87=BA=E6=9D=A5=E3=80=81?= =?UTF-8?q?=E5=8F=AF=E4=BB=A5=E5=8F=AA=E9=87=8D=E8=AF=95=E8=BF=99=E5=BC=A0?= =?UTF-8?q?=E8=A1=A8=EF=BC=89=EF=BC=9B=E6=8F=90=E7=A4=BA=E8=AF=8D=E5=8A=A0?= =?UTF-8?q?=E4=B8=8A=E8=B5=B7=E8=8D=89=E8=A6=81=E6=B1=82=E4=B8=80=E6=AE=B5?= =?UTF-8?q?=EF=BC=88=E6=95=B0=E6=8D=AE=E6=BA=90=E8=AF=B4=E6=98=8E=E3=80=81?= =?UTF-8?q?=E8=B5=B7=E8=8D=89=E8=A6=81=E6=B1=82=E3=80=81=E6=9C=AC=E6=AC=A1?= =?UTF-8?q?=E8=AF=B4=E6=98=8E=EF=BC=8C=E8=A7=84=E7=9F=A9=E5=9C=A8=E5=89=8D?= =?UTF-8?q?=E3=80=81=E5=86=B2=E7=AA=81=E4=BB=A5=E8=A7=84=E7=9F=A9=E4=B8=BA?= =?UTF-8?q?=E5=87=86=EF=BC=89=E5=92=8C=E3=80=8C=E6=80=9D=E8=80=83=E6=97=B6?= =?UTF-8?q?=E4=B8=8D=E5=BF=85=E9=80=90=E5=88=97=E8=8D=89=E6=8B=9F=E3=80=8D?= =?UTF-8?q?=EF=BC=9B=E5=90=8C=E6=AD=A5=E8=B5=B7=E8=8D=89=E6=94=B9=E7=94=A8?= =?UTF-8?q?=E6=96=B0=E5=BC=95=E6=93=8E=EF=BC=8C=E5=86=99=E5=85=A5=E6=8A=BD?= =?UTF-8?q?=E6=88=90=E6=AF=8F=E5=BC=A0=E8=A1=A8=E4=B8=80=E7=BB=93=E6=9D=9F?= =?UTF-8?q?=E5=B0=B1=E8=83=BD=E8=B0=83=E7=9A=84=20write=5Fdraft?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- backend/app/data/catalog.py | 357 +++--- backend/app/data/catalog_draft.py | 1048 ++++++++++++++++++ backend/tests/fixtures/catalog/fake_model.py | 62 +- backend/tests/test_catalog_draft.py | 5 +- backend/tests/test_catalog_draft_engine.py | 555 ++++++++++ 5 files changed, 1819 insertions(+), 208 deletions(-) create mode 100644 backend/app/data/catalog_draft.py create mode 100644 backend/tests/test_catalog_draft_engine.py diff --git a/backend/app/data/catalog.py b/backend/app/data/catalog.py index 281633a7..c48b99bc 100644 --- a/backend/app/data/catalog.py +++ b/backend/app/data/catalog.py @@ -1130,10 +1130,14 @@ async def save_patch(session: AsyncSession, source_id: str, table: str, changes: @dataclass(frozen=True) class TableDraft: - """一张表的起草结果。covered：这一轮完整算过的来源（merge_notes 据此删掉这些来源里已经没有的旧项）。""" + """一张表的起草结果。covered：这一轮完整算过的来源（merge_notes 据此删掉这些来源里已经没有的旧项）。 + + error：模型起草只写出了一部分（有的段没写出来）时，没写出来的是哪部分、为什么。这时 covered 不含 llm：没写全 + 就不算完整起草过，已有的模型起草项不会因此被删。""" notes: dict[str, Any] covered: frozenset[str] = frozenset() + error: str | None = None def system_notes_source(source: Any) -> bool: @@ -1446,10 +1450,25 @@ status 状态；attribute 属性）。 - 只写数据事实，不写计算公式，不写指标怎么算。 - 拿不准的字段留空，不要猜。列名、表名照原样写，不要翻译或改写。 - 标着「系统说明」的内容是导入时按核对结果生成的，以它为准：不要重复，也不要与之矛盾。 -- 标着「已确认」的目录项由人工确认过，以它为准。""" +- 标着「已确认」的目录项由人工确认过，以它为准。 +- 表名下一行写着「这次只写……」的表，只写它说的那部分：只写表级的项时不写列，只写列出的几列时不写表级的项、\ +也不写别的列。 +- 使用者补充的背景和要求帮你理解业务、统一用词；它和这里的规矩冲突时，以这里的规矩为准。 +- 思考时不必把要输出的内容逐列草拟一遍，想清楚后直接按格式输出：思考和输出共用同一份长度额度，额度用完时还没输出的\ +内容就全丢了。""" CATALOG_DRAFT_PROMPT = "请为下面 {n} 张表编写数据目录，每张表输出一项，table 写表名。\n\n{tables}" +#: 使用者补充的背景和要求（数据源说明、起草要求、本次说明），放在表结构前面单独一段；都没有就不出现这一段 +CATALOG_DRAFT_BACKGROUND_PROMPT = "使用者补充的背景和要求：\n{lines}" + +#: 纯文本那条路（结构化输出用不了时）：没有 JSON Schema 约束，要把格式说清 +CATALOG_DRAFT_TEXT_PROMPT = """只输出一个 JSON 对象，不要输出别的文字，格式如下（拿不准的字段不写）： +{"tables": [{"table": "表名", "label": "中文名", "grain": "一行代表什么", "kind": "fact", "keys": ["列名"], \ +"business_date": {"column": "列名", "rule": "一句规则", "timezone": "时区"}, \ +"columns": [{"name": "列名", "label": "中文名", "meaning": "含义", "unit": "单位", "measure": "flow"}]}]} +kind 取 fact、dimension、snapshot、log、config 之一；measure 取 flow、stock、ratio、identifier、status、attribute 之一。""" + #: 拼表结构描述用的固定片段（写给模型） CATALOG_DRAFT_PROMPT_LABELS = { "table": "表 {name}", @@ -1465,6 +1484,16 @@ CATALOG_DRAFT_PROMPT_LABELS = { "col_system": "系统说明：{text}", "unique": "唯一约束：{cols}", "confirmed": "已确认的目录项：", + # 分段起草：表名下一行说这次只写哪部分（宽表的表级一段、列分段，拆小重问时的半段） + "only_table": "这次只写表级的项（中文名、粒度、表类型、业务主键、业务日期），不写列", + "only_columns": "这次只写下面列出的 {n} 列（这张表共 {total} 列），不写表级的项", + "table_and_columns": "这次只写表级的项和这 {n} 列：{cols}", + "brief_columns": "列（共 {n} 列，这里只列名字和类型）：", + "drafted": "已起草的表级内容：", + # 使用者补充的背景和要求 + "description": "数据源说明：{text}", + "guidance": "起草要求：{text}", + "note": "本次说明：{text}", } CATALOG_DRAFT_PROMPT_SCHEMA: dict[str, Any] = { @@ -1509,9 +1538,6 @@ CATALOG_DRAFT_PROMPT_SCHEMA: dict[str, Any] = { "required": ["tables"], } -_P = CATALOG_DRAFT_PROMPT_LABELS - - class DraftModel: """起草用的模型：先按 MODEL_MAX_TOKENS 的额度调；服务方嫌额度太大（budget_rejected）时换成 MODEL_FALLBACK_MAX_TOKENS 的那一个，之后的调用都用它（shrink）。一次起草（一个起草任务、一次同步起草）共用一个，退回之后不再先失败一次。 @@ -1572,58 +1598,6 @@ def _confirmed_only(notes: Mapping[str, Any] | None) -> dict[str, Any]: return _assemble({k: v for k, v in _slots(notes).items() if v.get("status") in ("confirmed", "verified")}) -def _describe_for_model(name: str, meta: Mapping[str, Any], *, system: bool, existing: Mapping[str, Any] | None) -> str: - """一张表交给模型的结构描述：列、类型、主键、外键、唯一约束、注释；不含任何数据值。""" - lines = [(_P["view"] if meta.get("is_view") else _P["table"]).format(name=name)] - comment = (meta.get("comment") or "").strip() - if comment: - lines.append((_P["system"] if system else _P["comment"]).format(text=comment)) - pk = set(meta.get("primary_key") or []) - fk_of: dict[str, str] = {} - for fk in meta.get("foreign_keys") or []: - if isinstance(fk, dict) and len(fk.get("columns") or []) == 1: - to_cols = fk.get("to_columns") or [] - fk_of[fk["columns"][0]] = f"{fk.get('to_table')}.{to_cols[0]}" if to_cols else str(fk.get("to_table")) - lines.append(_P["columns"]) - for col in meta.get("columns") or []: - marks = [] - if col["name"] in pk: - marks.append(_P["pk"]) - if not col.get("nullable", True): - marks.append(_P["not_null"]) - if col["name"] in fk_of: - marks.append(_P["fk"].format(target=fk_of[col["name"]])) - if text := (col.get("comment") or "").strip(): - marks.append((_P["col_system"] if system else _P["col_comment"]).format(text=text)) - line = _P["column"].format(name=col["name"], type=col.get("type") or "") - lines.append(line + ("，" + "，".join(marks) if marks else "")) - for group in meta.get("unique") or []: - if isinstance(group, list) and group: - lines.append(_P["unique"].format(cols="、".join(group))) - confirmed = _notes_lines(_confirmed_only(existing), meta=meta, system_notes=False) - if confirmed: - lines.append(_P["confirmed"]) - lines += confirmed - return "\n".join(lines) - - -def _batches(names: list[str], tables: Mapping[str, Any]) -> list[list[str]]: - """按表分批：每批不超过 MODEL_BATCH_TABLES 张、合计不超过 MODEL_BATCH_COLUMNS 列（一张表超了就自成一批）。""" - out: list[list[str]] = [] - batch: list[str] = [] - width = 0 - for name in names: - n = len((tables.get(name) or {}).get("columns") or []) - if batch and (len(batch) >= MODEL_BATCH_TABLES or width + n > MODEL_BATCH_COLUMNS): - out.append(batch) - batch, width = [], 0 - batch.append(name) - width += n - if batch: - out.append(batch) - return out - - #: 模型给了这张表、但一项可用内容都没有（只有表名的空壳，或者给的值全不合规）：这张表记模型失败，不算起草成功。 #: 不然界面只看到「新增 0 项」，会说成「没有变化」 MODEL_EMPTY_REASON = "模型没有给出这张表的可用内容" @@ -1653,39 +1627,6 @@ def _model_tables(raw: Any) -> list[dict[str, Any]] | None: return [t for t in raw if isinstance(t, dict)] -async def _ask_model(model: Any, messages: list[Any], - usable: Callable[[list[dict[str, Any]]], bool] | None = None) -> list[dict[str, Any]]: - """问一批。先走结构化输出；模型不支持、调用报错、给回来的不是约定的结构，或者 usable 说它不够用（实测见过 - 结构化输出只给 {"tables":[{"table":"x"}]} 这样的空壳），再退回纯文本、从正文里取 JSON。 - - 结构化输出报错也退回（同 judge._ask）：有的网关不支持工具调用，一调就报错，纯文本却能用。代价是真的 - 网络故障会多试一次，两次加起来仍受 MODEL_TIMEOUT_S 约束。 - - 两条路都拿到了东西时，结构化的在前、纯文本的在后一起交回（同一张表各有一份，调用方挑能用的那份）；纯文本那次 - 失败了就只交回结构化的那份（调用方据此给空壳的表记失败）。 - """ - try: - parsed = _model_tables(await model.with_structured_output(CATALOG_DRAFT_PROMPT_SCHEMA).ainvoke(messages)) - except Exception: # noqa: BLE001 - 退回纯文本 - parsed = None - if parsed is not None and (usable is None or usable(parsed)): - return parsed - from app.engine.state import message_text - - try: - reply = await model.ainvoke(messages) - except Exception: # noqa: BLE001 - 结构化那份还在：交回它，空壳的表由调用方记失败 - if parsed is not None: - return parsed - raise - text = _model_tables(message_text(reply)) - if text is None: - if parsed is not None: - return parsed - raise ValueError("模型的回复不是约定的格式") - return [*(parsed or []), *text] - - def _clean_text(value: Any, limit: int = _TEXT_MAX) -> str | None: if not isinstance(value, str): return None @@ -1740,65 +1681,30 @@ def _from_model(entry: Mapping[str, Any], meta: Mapping[str, Any], *, system: bo async def draft_with_model(model: Any, source: Any, tables: list[str], *, - existing: Mapping[str, Mapping[str, Any]] | None = None) -> dict[str, TableDraft | str]: + existing: Mapping[str, Mapping[str, Any]] | None = None, + guidance: str | None = None, note: str | None = None) -> dict[str, TableDraft | str]: """用模型起草这些表的中文名、粒度、表类型、业务主键、业务日期和每列的中文名、含义、单位、度量类型。 - model 要有 with_structured_output(schema) 或 ainvoke(messages)（测试注入假模型）。按表分批（_batches）， - 最多 MODEL_CONCURRENCY 批同时发。返回 {表名: TableDraft 或失败原因}：一批失败只影响这一批的表， - 其余照常；失败的表不算 covered，已有的模型起草项不会因此被删。 + 实现在 data/catalog_draft.py（分段、按失败原因处理、捡回被截断的部分）；这里是给老调用方的形状：返回 + {表名: TableDraft 或失败原因}。全部写出来的 covered 是 {"llm"}；只写出一部分的仍是 TableDraft，covered 为空、 + error 写明哪部分没写出来；一项都没写出来的是失败原因。表结构里没有的表不在结果里。 - 交给模型的只有表结构（列、类型、主键、外键、唯一约束、注释），不发任何数据值；existing 里已确认、 - 有确证的项作为「以此为准」的背景一起交过去。 + model 要有 with_structured_output(schema) 或 ainvoke(messages)（测试注入假模型）。交给模型的只有表结构和使用者 + 补充的背景（数据源说明、guidance 起草要求、note 本次说明），不发任何数据值；existing 里已确认、有确证的项作为 + 「以此为准」的背景一起交过去。 """ - from langchain_core.messages import HumanMessage, SystemMessage + from app.data import catalog_draft - all_tables = (getattr(source, "schema_cache", None) or {}).get("tables") or {} - system = system_notes_source(source) - names = [t for t in tables if isinstance(all_tables.get(t), dict)] - results: dict[str, TableDraft | str] = {} - gate = asyncio.Semaphore(MODEL_CONCURRENCY) - - async def run(batch: list[str]) -> None: - described = "\n\n".join(_describe_for_model(n, all_tables[n], system=system, - existing=(existing or {}).get(n)) for n in batch) - messages = [SystemMessage(content=CATALOG_DRAFT_SYSTEM), - HumanMessage(content=CATALOG_DRAFT_PROMPT.format(n=len(batch), tables=described))] - def drafted(entries: list[dict[str, Any]]) -> dict[str, dict[str, Any]]: - """这批里每张表能用的那份：同一张表给了几份（结构化、纯文本各一份）时取第一份有可用内容的。""" - out: dict[str, dict[str, Any]] = {} - for entry in entries: - name = str(entry.get("table")) - if name in batch and name not in out: - notes = _from_model(entry, all_tables[name], system=system) - if notes: - out[name] = notes - return out - - try: - async with gate: - got = await asyncio.wait_for( - _ask_model(model, messages, usable=lambda entries: len(drafted(entries)) == len(batch)), - timeout=MODEL_TIMEOUT_S) - except asyncio.TimeoutError: - for n in batch: - results[n] = f"模型起草超时：{MODEL_TIMEOUT_S} 秒内没有收到回复" - return - except Exception as e: # noqa: BLE001 - 一批失败不影响其他批和其他来源 - from app.core.errors import describe_exception, first_line - - for n in batch: - results[n] = f"模型起草失败：{first_line(e) or describe_exception(e)}" - return - named = {str(t.get("table")) for t in got} - usable = drafted(got) - for n in batch: - if n in usable: - results[n] = TableDraft(notes=usable[n], covered=frozenset({"llm"})) - else: - results[n] = MODEL_EMPTY_REASON if n in named else "模型的回复里没有这张表" - - await asyncio.gather(*(run(b) for b in _batches(names, all_tables))) - return {n: results[n] for n in names if n in results} + outcomes = await catalog_draft.run_model_draft( + model, source, tables, existing=existing, background=catalog_draft.Background(guidance=guidance, note=note)) + out: dict[str, TableDraft | str] = {} + for name, outcome in outcomes.items(): + if outcome.status == "failed": + out[name] = outcome.error or MODEL_EMPTY_REASON + else: + out[name] = TableDraft(notes=outcome.notes, covered=frozenset({"llm"}) if outcome.status == "ok" + else frozenset(), error=outcome.error) + return out # ========================================================================== @@ -1808,7 +1714,8 @@ async def draft_with_model(model: Any, source: Any, tables: list[str], *, @dataclass class TableDraftResult: - """一张表这次起草的结果。error：这张表没起草（表结构里没有、写入一直冲突）；model_error：只是模型那部分失败。""" + """一张表这次起草的结果。error：这张表没起草（表结构里没有、写入一直冲突）；model_error：只是模型那部分失败， + 或者只写出了一部分（哪部分没写出来、为什么）。""" table: str added: int = 0 @@ -1817,9 +1724,14 @@ class TableDraftResult: version: int = 0 error: str | None = None model_error: str | None = None - #: 模型给了这张表几项可用内容（并入目录之前，按项数）。没用模型、模型这部分失败时为 None。 + #: 模型给了这张表几项可用内容（并入目录之前，按项数）。没用模型、模型这部分一项都没写出来时为 None。 #: 大于 0 而新增、更新、删除都是 0，才是「模型给了内容、只是和现有目录一致」 model_items: int | None = None + #: ok 写入了（用模型时各段都写出来了）、partial 模型只写出一部分、failed 没写入或模型一段都没写出来、 + #: skipped 表结构里没有这张表 + status: str = "ok" + #: 模型分几段起草（不用模型时为 None） + parts_total: int | None = None @dataclass @@ -1834,35 +1746,102 @@ class DraftReport: model_error: str | None = None -async def draft_catalog(session: AsyncSession, source: Any, *, tables: list[str] | None = None, - use_model: bool = False, model: Any = None, actor: str | None = None) -> DraftReport: - """起草这些表的目录并写入，同步完成。 +#: 表结构里没有要起草的表（被删了、改名了，或者名字写错了） +def missing_table_error(name: str) -> str: + return f"表结构里没有 {name}，请先重新探查结构" + + +#: 写入一直撞上别人刚改过（乐观锁重读重并三次仍冲突） +WRITE_CONFLICT_ERROR = "这张表的数据目录正被其他人修改，请稍后再起草" + + +async def write_draft(session: AsyncSession, source: Any, table: str, modeled: Any = None, *, at: str, + actor: str | None) -> TableDraftResult: + """一张表的起草并入目录并写入：数据库注释、外键约束、命名推断照做，modeled（catalog_draft.ModelOutcome， + 没用模型时为 None）有可用内容的再并上。同步起草和后台起草任务（每张表一结束就调）共用这一处。 - - source：手工源的 DataSource，或上传源绑定当前快照的视图（table_versions.resolve_source）；按 source.id 存。 - - tables 为空时按使用次数（table_usage）取前 DRAFT_DEFAULT_TABLES 张，次数相同按表结构里的顺序。 - - 每张表：数据库注释、外键约束、命名推断照做；use_model 时再加模型起草。model 可注入（测试用假模型）， - 不给就按设置解析（resolve_draft_model）；模型用不了、某一批失败，都不影响其余来源照常写入。 - 并入已有目录按 merge_notes 的规则：人工确认、驳回过的项不动。内容没变不写、不升版本。 + - 模型只写出一部分的，模型那份不算完整起草过（covered 不含 llm），已有的模型起草项不会因此被删。 - 写入撞上别人刚改过（乐观锁）就重读重并，最多三次。 """ - cache = getattr(source, "schema_cache", None) or {} - all_tables = cache.get("tables") or {} - report = DraftReport() - if tables: - wanted = list(dict.fromkeys(str(t) for t in tables)) + result = TableDraftResult(table=table) + structure = draft_structure(source, table) + model_draft: TableDraft | None = None + if modeled is not None: + result.parts_total = modeled.parts_total + result.model_error = modeled.error + if modeled.items: + result.model_items = modeled.items + model_draft = TableDraft(notes=modeled.notes, covered=frozenset({"llm"}) if modeled.status == "ok" + else frozenset()) + for _attempt in range(3): + entry = await read_entry(session, source.id, table) + version = entry.version if entry else 0 + merged, stats = merge_notes(entry.notes if entry else {}, structure.notes, covered=structure.covered, at=at) + if model_draft is not None: + merged, more = merge_notes(merged, model_draft.notes, covered=model_draft.covered, at=at) + stats += more + result.added, result.updated, result.removed = stats.added, stats.updated, stats.removed + result.version = version + if entry is None and not merged: + break # 什么也没起草出来，不建空目录 + try: + written = await write_entry(session, source.id, table, merged, if_version=version, actor=actor) + except CatalogConflict: + continue + result.version = written.version + break else: - usage = await table_usage(session, source) - order = {name: i for i, name in enumerate(all_tables)} - wanted = sorted(all_tables, key=lambda n: (-usage.get(n, 0), order[n]))[:DRAFT_DEFAULT_TABLES] + result.error = WRITE_CONFLICT_ERROR + result.added = result.updated = result.removed = 0 + if result.error or (modeled is not None and modeled.status == "failed"): + result.status = "failed" + elif modeled is not None and modeled.status == "partial": + result.status = "partial" + return result + + +def pick_tables(cache: Mapping[str, Any], wanted: Iterable[str]) -> tuple[list[str], list[str]]: + """要起草的表 → (表结构里的键，去重、按给的顺序；表结构里没有的原样名字)。""" picked: list[str] = [] - for name in wanted: + unknown: list[str] = [] + for name in dict.fromkeys(str(t) for t in wanted): key = resolve_table_name(cache, name) if key is None: - report.tables.append(TableDraftResult(table=name, error=f"表结构里没有 {name}，请先重新探查结构")) + unknown.append(name) elif key not in picked: picked.append(key) + return picked, unknown - llm: dict[str, TableDraft | str] = {} + +async def default_tables(session: AsyncSession, source: Any) -> list[str]: + """不指定表时起草哪些：按使用次数（table_usage）取前 DRAFT_DEFAULT_TABLES 张，次数相同按表结构里的顺序。""" + all_tables = (getattr(source, "schema_cache", None) or {}).get("tables") or {} + usage = await table_usage(session, source) + order = {name: i for i, name in enumerate(all_tables)} + return sorted(all_tables, key=lambda n: (-usage.get(n, 0), order[n]))[:DRAFT_DEFAULT_TABLES] + + +async def draft_catalog(session: AsyncSession, source: Any, *, tables: list[str] | None = None, + use_model: bool = False, model: Any = None, actor: str | None = None, + guidance: str | None = None, note: str | None = None) -> DraftReport: + """起草这些表的目录并写入，同步完成（POST /catalog/draft；界面改用后台起草任务，见 data/catalog_jobs.py）。 + + - source：手工源的 DataSource，或上传源绑定当前快照的视图（table_versions.resolve_source）；按 source.id 存。 + - tables 为空时按使用次数取前 DRAFT_DEFAULT_TABLES 张（default_tables）。 + - 每张表：数据库注释、外键约束、命名推断照做；use_model 时再加模型起草（data/catalog_draft.py：分段、按失败 + 原因处理）。model 可注入（测试用假模型），不给就按设置解析（resolve_draft_model）；模型用不了、某一段失败， + 都不影响其余来源照常写入。guidance 是数据源级的起草要求，note 是本次说明，和数据源说明一起交给模型。 + - 并入、写入见 write_draft。 + """ + cache = getattr(source, "schema_cache", None) or {} + report = DraftReport() + wanted = list(tables) if tables else await default_tables(session, source) + picked, unknown = pick_tables(cache, wanted) + for name in unknown: + report.tables.append(TableDraftResult(table=name, error=missing_table_error(name), status="skipped")) + + outcomes: dict[str, Any] = {} if use_model and picked: try: if model is None: @@ -1872,41 +1851,17 @@ async def draft_catalog(session: AsyncSession, source: Any, *, tables: list[str] except CatalogModelUnavailable as e: report.model_error = str(e) else: + from app.data import catalog_draft + report.model_used = True existing = await read_catalog(session, source.id) - llm = await draft_with_model(model, source, picked, existing={t: e.notes for t, e in existing.items()}) + outcomes = await catalog_draft.run_model_draft( + model, source, picked, existing={t: e.notes for t, e in existing.items()}, + background=catalog_draft.Background(guidance=guidance, note=note)) at = now_iso() for table in picked: - result = TableDraftResult(table=table) - structure = draft_structure(source, table) - modeled = llm.get(table) - if isinstance(modeled, str): - result.model_error = modeled - elif isinstance(modeled, TableDraft): - result.model_items = len(_slots(modeled.notes)) - for _attempt in range(3): - entry = await read_entry(session, source.id, table) - version = entry.version if entry else 0 - merged, stats = merge_notes(entry.notes if entry else {}, structure.notes, covered=structure.covered, - at=at) - if isinstance(modeled, TableDraft): - merged, more = merge_notes(merged, modeled.notes, covered=modeled.covered, at=at) - stats += more - result.added, result.updated, result.removed = stats.added, stats.updated, stats.removed - result.version = version - if entry is None and not merged: - break # 什么也没起草出来，不建空目录 - try: - written = await write_entry(session, source.id, table, merged, if_version=version, actor=actor) - except CatalogConflict: - continue - result.version = written.version - break - else: - result.error = "这张表的数据目录正被其他人修改，请稍后再起草" - result.added = result.updated = result.removed = 0 - report.tables.append(result) + report.tables.append(await write_draft(session, source, table, outcomes.get(table), at=at, actor=actor)) return report @@ -2271,12 +2226,16 @@ def frozen_catalog(entries: Mapping[str, CatalogEntry], tables: Iterable[str]) - __all__ = [ - "CARDINALITIES", "CARDINALITY_LABEL", "CATALOG_DRAFT_PROMPT", "CATALOG_DRAFT_PROMPT_SCHEMA", - "CATALOG_DRAFT_SYSTEM", "COLUMN_FIELDS", "COLUMN_FIELD_LABEL", "COLUMN_MEASURES", "CatalogConflict", + "CARDINALITIES", "CARDINALITY_LABEL", "CATALOG_DRAFT_BACKGROUND_PROMPT", "CATALOG_DRAFT_PROMPT", + "CATALOG_DRAFT_PROMPT_LABELS", "CATALOG_DRAFT_PROMPT_SCHEMA", "CATALOG_DRAFT_SYSTEM", "CATALOG_DRAFT_TEXT_PROMPT", + "COLUMN_FIELDS", "COLUMN_FIELD_LABEL", "COLUMN_MEASURES", "CatalogConflict", "CatalogEntry", "CatalogInvalid", "CatalogModelUnavailable", "CatalogPathError", "DRAFT_DEFAULT_TABLES", - "DraftReport", "INFERRED_MARK", "ITEM_SOURCES", "ITEM_STATUSES", "JoinEdge", "KIND_LABEL", "MEASURE_LABEL", - "MODEL_EMPTY_REASON", "MergeStats", "NEW_RELATION_PATH", "PATCH_MAX_CHANGES", "PROFILE_NOTE_PREFIX", - "PROFILE_VERIFY_COVERAGE", + "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", + "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", "TABLE_KINDS", "TableDraft", "TableDraftResult", "UI_KIND_LABEL", "UI_MEASURE_LABEL", "apply_human_edit", "apply_patch", "codes_complete", "describe_slot", "draft_catalog", "draft_structure", "draft_with_model", diff --git a/backend/app/data/catalog_draft.py b/backend/app/data/catalog_draft.py new file mode 100644 index 00000000..2b1d71bf --- /dev/null +++ b/backend/app/data/catalog_draft.py @@ -0,0 +1,1048 @@ +"""数据目录的模型起草：分段、按失败原因处理、捡回被截断的部分、记用量。 + +对外的名字留在 data/catalog.py（draft_with_model、draft_catalog、write_draft、CATALOG_DRAFT_SYSTEM、各个 MODEL_* +常量……），这里是实现。常量一律在调用时按 catalog.X 读：测试改 catalog 上的常量就能生效。 + +**为什么要分段、按原因处理。** 实测（经 OpenAI 兼容网关的会思考的模型）：思考算在输出额度里，一段上百列时思考就把 +额度全用掉，正文 0 字或只给一个空壳；同一批表有时成功有时失败，看模型这次想多久。以前失败了退回纯文本、发同一个提示词， +同样失败、耗时翻倍；超时把第一次拿到的部分也丢掉；一批里有一张表是空的就整批重问。 + +**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）。 + +**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。 +- 被截断的 JSON：先把已经完整的表、列对象捡回来，只补剩下的。 +- 一段里缺了的表、列只补缺的，补一次。补过仍然没给的列算模型拿不准、留空，不算没写出来；整张表一项都没有才算。 +- 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。 +- 别的错误：这一段没写出来，原因照实写。 + +**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）。 + +**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）。 + +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）。 +""" +from __future__ import annotations + +import asyncio +import copy +import json +import logging +import re +import weakref +from collections.abc import Awaitable, Callable, Mapping, Sequence +from dataclasses import dataclass, field +from typing import Any + +from app.data import catalog as C +from app.data.names import name_key + +logger = logging.getLogger(__name__) + +#: 退避等待用的 sleep（测试里换掉，不真等） +_sleep = asyncio.sleep + +#: 服务方嫌额度太大时报错里会出现的字样（同 copilot._BUDGET_WORDS、recipe_ai._BUDGET_WORDS，有测试核对） +BUDGET_WORDS = ("max_tokens", "max_completion_tokens", "max_output_tokens", "maximum context length") +#: 结构化输出这个接口不支持时报错里会出现的字样：json_schema 的 response_format（有的服务只认 json_object）、 +#: 工具调用（有的网关不支持）、Anthropic 的 output_config +STRUCTURED_WORDS = ("response_format", "json_schema", "json_object", "tool", "function", "output_config", + "output_format", "structured output") +#: 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" +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_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_TIMEOUT = "模型 {seconds} 秒内没有写完这一段" +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}" + + +# ========================================================================== +# 对外的形状 +# ========================================================================== + + +@dataclass(frozen=True) +class Background: + """使用者补充的背景和要求：guidance 是数据源级的起草要求，note 是这一次的补充说明。数据源说明（description） + 从源上读。都没有就不出现那一段。""" + + guidance: str | None = None + note: str | None = None + + +@dataclass +class ModelOutcome: + """一张表的模型起草结果。 + + 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。""" + + status: str + notes: dict[str, Any] + items: int + error: str | None + parts_total: int + parts_done: int + + +@dataclass +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； + - 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 + 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。""" + width = max(1, int(C.MODEL_BATCH_COLUMNS)) + return 1 if columns <= width else 1 + -(-columns // width) + + +async def run_model_draft(model: Any, source: Any, tables: Sequence[str], *, + existing: Mapping[str, Mapping[str, Any]] | None = None, + background: Background | None = None, + hooks: Hooks | None = None) -> dict[str, ModelOutcome]: + """用模型起草这些表（schema_cache["tables"] 的键；表结构里没有的跳过，不在结果里）。 + + model 要有 with_structured_output(schema, include_raw=…) 和 ainvoke(messages)；有 shrink() 的（catalog.DraftModel） + 在服务方嫌额度太大时换小额度。existing 是 {表名: 目录}：已确认、有确证的项作为「以此为准」的背景交给模型。 + """ + run = _Run(model, source, tables, existing or {}, background or Background(), hooks or Hooks()) + return await run.run() + + +# ========================================================================== +# 并发名额 +# ========================================================================== + +_GATES: "weakref.WeakKeyDictionary[asyncio.AbstractEventLoop, tuple[int, asyncio.Semaphore]]" = \ + weakref.WeakKeyDictionary() + + +def model_gate() -> asyncio.Semaphore: + """全进程的模型调用名额（MODEL_CONCURRENCY 个）。按事件循环各一个：asyncio 的信号量一旦在某个循环里等过， + 就不能换循环用（测试每个用例一个循环）。""" + loop = asyncio.get_running_loop() + limit = max(1, int(C.MODEL_CONCURRENCY)) + held = _GATES.get(loop) + if held is None or held[0] != limit: + held = (limit, asyncio.Semaphore(limit)) + _GATES[loop] = held + return held[1] + + +# ========================================================================== +# 失败的种类 +# ========================================================================== + + +def _status_of(e: BaseException) -> int | None: + for obj in (e, getattr(e, "response", None)): + code = getattr(obj, "status_code", None) + if isinstance(code, int): + return code + return None + + +def classify_failure(e: BaseException) -> str: + """调用模型抛出来的异常属于哪一种：length 额度用完（LengthFinishReasonError）、budget 服务方嫌额度太大、 + unsupported 结构化输出不支持、timeout 超时、transient 网络错误 / 429 / 5xx、error 别的。""" + from app.core.errors import NETWORK, RATE_LIMIT, SERVER, TIMEOUT, classify + + names = {c.__name__ for c in type(e).__mro__} + if "LengthFinishReasonError" in names: + return "length" + status = _status_of(e) + text = str(e).lower() + if status in (400, 422) and any(w in text for w in BUDGET_WORDS): + return "budget" + # thinking 打开时 Anthropic 不让强制调用工具，langchain 改成不强制，模型没调用工具就抛 OutputParserException + if isinstance(e, NotImplementedError) or "OutputParserException" in names: + return "unsupported" + if status in (400, 422) and any(w in text for w in STRUCTURED_WORDS): + return "unsupported" + if isinstance(e, TimeoutError): + return "timeout" + kind = classify(e) + if kind == TIMEOUT: + return "timeout" + if kind in (RATE_LIMIT, SERVER, NETWORK): + return "transient" + return "error" + + +def _seconds(value: Any) -> Any: + number = float(value) + return int(number) if number.is_integer() else number + + +# ========================================================================== +# 读回复：结构化输出、正文里的 JSON、被截断的 JSON +# ========================================================================== + + +@dataclass +class _Reply: + #: 回复里的表对象（截断的只含已经完整的字段和列） + entries: list[dict[str, Any]] = field(default_factory=list) + #: JSON 没写完（捡回了一部分） + truncated: bool = False + #: 思考加输出用完了额度 + exhausted: bool = False + #: 这一段最终没写出来的原因（重试用完、别的错误、不是约定的格式） + error: str | None = None + + +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。""" + + __slots__ = ("value",) + + def __init__(self, value: Any) -> None: + self.value = value + + +class _Bad(ValueError): + """不是 JSON（不是写到一半，是写错了）。""" + + +_NUMBER = re.compile(r"-?(?:0|[1-9]\d*)(?:\.\d+)?(?:[eE][+-]?\d+)?") +_LITERALS = (("true", True), ("false", False), ("null", None)) +_MAX_DEPTH = 64 + + +def _skip(text: str, i: int) -> int: + n = len(text) + while i < n and text[i] in " \t\r\n": + i += 1 + return i + + +def _string(text: str, i: int) -> tuple[Any, int, bool]: + n = len(text) + j = i + 1 + while j < n: + ch = text[j] + if ch == "\\": + j += 2 + continue + if ch == '"': + try: + return json.loads(text[i:j + 1]), j + 1, True + except ValueError as e: + raise _Bad from e + j += 1 + return None, n, False + + +def _value(text: str, i: int, depth: int) -> tuple[Any, int, bool]: + """从 text[i] 起读一个 JSON 值：(值, 读到哪, 写完了没有)。读到结尾还没写完的容器返回 _Cut。""" + if depth > _MAX_DEPTH: + raise _Bad + n = len(text) + i = _skip(text, i) + if i >= n: + return None, i, False + ch = text[i] + if ch == "{": + return _object(text, i + 1, depth + 1) + if ch == "[": + return _array(text, i + 1, depth + 1) + if ch == '"': + return _string(text, i) + for word, value in _LITERALS: + if text.startswith(word, i): + return value, i + len(word), True + if word.startswith(text[i:]): + return None, n, False + m = _NUMBER.match(text, i) + if not m: + raise _Bad + j = m.end() + if j >= n: + return None, j, False # 数字写到了结尾，可能还没写完 + number = m.group() + return (float(number) if any(c in number for c in ".eE") else int(number)), j, True + + +def _object(text: str, i: int, depth: int) -> tuple[Any, int, bool]: + out: dict[str, Any] = {} + n = len(text) + while True: + i = _skip(text, i) + if i >= n: + return _Cut(out), i, False + ch = text[i] + if ch == "}": + return out, i + 1, True + if ch == ",": + i += 1 + continue + if ch != '"': + raise _Bad + key, i, ok = _string(text, i) + if not ok: + return _Cut(out), i, False + i = _skip(text, i) + if i >= n: + return _Cut(out), i, False + if text[i] != ":": + raise _Bad + value, i, ok = _value(text, i + 1, depth) + if not ok: + if isinstance(value, _Cut): + out[key] = value + return _Cut(out), i, False + out[key] = value + + +def _array(text: str, i: int, depth: int) -> tuple[Any, int, bool]: + out: list[Any] = [] + n = len(text) + while True: + i = _skip(text, i) + if i >= n: + return _Cut(out), i, False + ch = text[i] + if ch == "]": + return out, i + 1, True + if ch == ",": + i += 1 + continue + value, i, ok = _value(text, i, depth) + if not ok: + if isinstance(value, _Cut): + out.append(value) + return _Cut(out), i, False + out.append(value) + + +def _json_prefix(text: str) -> tuple[Any, bool] | None: + """正文里的第一个 JSON 值：(值, 写完了没有)。前面可以有一句话、围栏，后面可以有别的字；不是 JSON 返回 None。""" + for start in sorted(k for k in (text.find("{"), text.find("[")) if k >= 0): + try: + value, _, complete = _value(text, start, 0) + except (_Bad, RecursionError): + continue + return value, complete + return None + + +def _entry(item: Any) -> dict[str, Any] | None: + """一个表对象：只留写完了的字段；columns 留写完了的列对象。没有表名的不要。""" + if isinstance(item, _Cut): + item = item.value + if not isinstance(item, dict) or not isinstance(item.get("table"), str): + return None + out: dict[str, Any] = {} + for key, value in item.items(): + if key == "columns": + cols = value.value if isinstance(value, _Cut) else value + if isinstance(cols, list): + out["columns"] = [c for c in cols if isinstance(c, dict)] + elif not isinstance(value, _Cut): + out[key] = value + return out + + +def _tables_of(root: Any) -> list[dict[str, Any]] | None: + """解析出来的根 → 表对象列表。{"tables": […]}、[…]、单个表对象（实测见过不带 tables 外壳的）都认。""" + if isinstance(root, _Cut): + root = root.value + if isinstance(root, dict): + if "tables" in root: + items = root["tables"] + elif isinstance(root.get("table"), str): + items = [root] + elif not root: + return [] + else: + return None + else: + items = root + if isinstance(items, _Cut): + items = items.value + if not isinstance(items, list): + return None + return [e for e in (_entry(item) for item in items) if e is not None] + + +def _parse_text(text: str, cut: bool) -> _Reply | None: + found = _json_prefix(text) + if found is None: + return None + root, complete = found + entries = _tables_of(root) + if entries is None: + return None + if complete: + return _Reply(entries=entries) + return _Reply(entries=entries, truncated=True, exhausted=cut) + + +def _cut_off(raw: Any) -> bool: + """额度用完停下的：OpenAI 系是 finish_reason=length，Anthropic 是 stop_reason=max_tokens。""" + meta = getattr(raw, "response_metadata", None) or {} + return (meta.get("finish_reason") or meta.get("stop_reason")) in ("length", "max_tokens") + + +def _thought(raw: Any) -> bool: + """这次回复有没有思考：思考内容（Anthropic 的 thinking 块、OpenAI 兼容系的 reasoning_content），或者用量里有推理 token。""" + from app.engine.state import thinking_text + + if raw is None: + return False + if thinking_text(raw): + return True + details = (getattr(raw, "usage_metadata", None) or {}).get("output_token_details") or {} + return bool(details.get("reasoning")) + + +def _text_of(raw: Any) -> str: + from app.engine.state import message_text + + return message_text(raw) if raw is not None else "" + + +def _raw_texts(raw: Any) -> list[str]: + """结构化输出没解析出来时，原始回复里可能还有 JSON 的地方：工具调用的参数（截断的在 invalid_tool_calls 里，是一段 + 字符串）、正文。""" + out: list[str] = [] + for call in getattr(raw, "tool_calls", None) or []: + args = call.get("args") if isinstance(call, dict) else None + if isinstance(args, dict) and args: + out.append(json.dumps(args, ensure_ascii=False)) + for call in getattr(raw, "invalid_tool_calls", None) or []: + args = call.get("args") if isinstance(call, dict) else None + if isinstance(args, str) and args.strip(): + out.append(args) + text = _text_of(raw) + if text.strip(): + out.append(text) + return out + + +def _entries_of(parsed: Any) -> list[dict[str, Any]] | None: + if parsed is None: + return None + if isinstance(parsed, dict) and not parsed: + return [] + return C._model_tables(parsed) + + +def _structured_reply(raw: Any, parsed: Any) -> _Reply: + entries = _entries_of(parsed) + if entries is not None: + return _Reply(entries=entries) + cut = _cut_off(raw) + texts = _raw_texts(raw) + for text in texts: + reply = _parse_text(text, cut) + if reply is not None: + return reply + if cut or _thought(raw): + return _Reply(exhausted=True) + if not texts: + return _Reply() # 什么都没写：按空壳处理 + return _Reply(error=REASON_FORMAT) + + +def _text_reply(raw: Any) -> _Reply: + text = _text_of(raw) + cut = _cut_off(raw) + if text.strip(): + reply = _parse_text(text, cut) + if reply is not None: + return reply + if cut: + return _Reply(exhausted=True) # 写了字、额度就用完了，一个 JSON 对象都没捡回来 + return _Reply(error=REASON_FORMAT) + return _Reply(exhausted=cut or _thought(raw)) + + +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): + return out.get("raw"), out.get("parsed") + return None, out + + +# ========================================================================== +# 一次起草 +# ========================================================================== + + +@dataclass(frozen=True) +class _Piece: + """一段里的一张表：要写哪些列、要不要表级的项、属于这张表的第几段（从 1 起）。""" + + table: str + columns: tuple[str, ...] + table_level: bool + part: int + + +@dataclass +class _Gain: + """一次回复里某张表拿到了什么。""" + + named: bool = False + items: int = 0 + table_items: int = 0 + answered: set[str] = field(default_factory=set) + + +@dataclass +class _Table: + name: str + meta: Mapping[str, Any] + columns: list[str] + existing: Mapping[str, Any] | None + parts: list[_Piece] = field(default_factory=list) + #: 已经拿到的模型起草（逐段合并） + notes: dict[str, Any] = field(default_factory=dict) + #: 模型的回复里出现过这张表 + named: bool = False + got_table: bool = False + got_columns: set[str] = field(default_factory=set) + #: 没写出来的列（表级的项记在 _TABLE）→ 原因 + failed: dict[str, str] = field(default_factory=dict) + parts_done: int = 0 + finished: bool = False + + def whole(self, piece: _Piece) -> bool: + return piece.table_level and len(piece.columns) == len(self.columns) + + +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。""" + out: dict[str, Any] = {} + if piece.table_level: + out.update({k: v for k, v in notes.items() if k != "columns"}) + wanted = set(piece.columns) + cols = {c: items for c, items in (notes.get("columns") or {}).items() if c in wanted} + if cols: + out["columns"] = cols + return out + + +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: + k = (len(pieces) + 1) // 2 + return pieces[:k], pieces[k:] + piece = pieces[0] + n = len(piece.columns) + if n < 2 * max(1, int(C.MODEL_SPLIT_MIN_COLUMNS)): + return None + h = (n + 1) // 2 + return ([_Piece(piece.table, piece.columns[:h], piece.table_level, piece.part)], + [_Piece(piece.table, piece.columns[h:], False, piece.part)]) + + +class _Run: + def __init__(self, model: Any, source: Any, tables: Sequence[str], existing: Mapping[str, Mapping[str, Any]], + background: Background, hooks: Hooks) -> None: + self.model = model + self.hooks = hooks + #: structured 结构化输出；text 纯文本、从正文取 JSON。结构化输出用不了时这次起草后面都改走纯文本 + self.mode = "structured" + self.system = C.system_notes_source(source) + all_tables = (getattr(source, "schema_cache", None) or {}).get("tables") or {} + self.order = [t for t in dict.fromkeys(str(t) for t in tables) if isinstance(all_tables.get(t), dict)] + self.tables = { + name: _Table(name=name, meta=all_tables[name], existing=existing.get(name), + columns=[str(c["name"]) for c in all_tables[name].get("columns") or [] + if isinstance(c, dict) and c.get("name") is not None]) + for name in self.order + } + self.background = self._background(getattr(source, "description", None), background) + + # ---------------------------------------------------------------- 总流程 + + async def run(self) -> dict[str, ModelOutcome]: + await asyncio.gather(*(self._unit(unit) for unit in self._plan())) + 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。""" + width = max(1, int(C.MODEL_BATCH_COLUMNS)) + most = max(1, int(C.MODEL_BATCH_TABLES)) + units: list[list[str] | str] = [] + pack: list[str] | None = None + used = 0 + for name in self.order: + st = self.tables[name] + 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))] + units.append(name) + continue + st.parts = [_Piece(name, tuple(st.columns), True, 1)] + if pack is None or len(pack) >= most or used + n > width: + pack, used = [], 0 + units.append(pack) + pack.append(name) + used += n + return units + + async def _unit(self, unit: list[str] | str) -> None: + names = unit if isinstance(unit, list) else [unit] + if isinstance(unit, list): + await self._guarded([self.tables[n].parts[0] for n in unit]) + for name in unit: + await self._part_done(name) + else: + head, *rest = self.tables[unit].parts + await self._guarded([head]) # 表级那段先写：列分段拿它做背景 + await self._part_done(unit) + + async def column_part(piece: _Piece) -> None: + await self._guarded([piece]) + await self._part_done(unit) + + await asyncio.gather(*(column_part(p) for p in rest)) + for name in names: + await self._table_done(name) + + 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) + except asyncio.CancelledError: + raise + except Exception as e: # noqa: BLE001 + from app.core.errors import describe_exception, first_line + + logger.exception("起草数据目录时出错：%s", "、".join(p.table for p in pieces)) + self._fail(pieces, REASON_CRASH.format(line=first_line(e) or describe_exception(e))) + + # ---------------------------------------------------------------- 一段 + + async def _segment(self, pieces: list[_Piece], depth: int, followup: bool, *, 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，空着的算模型拿不准）。 + mode / last：拆不开的空壳换纯文本问的那一次（再空就记没写出来）。""" + reply = await self._ask(pieces, mode) + if reply.error is not None: + self._fail(pieces, reply.error) + return + gains = self._take(pieces, reply.entries) + anything = any(g.items for g in gains) + 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) + if not rest: + return + kind = "exhausted" if reply.exhausted else "truncated" + if not anything: + await self._split(rest, depth, kind, followup) + 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 + else: + self._fail(rest, self._reason(kind, split=True)) + return + if not anything: + if followup: + return + if last: + self._fail(pieces, lambda p: self._reason("hollow", split=depth > 0, piece=p)) + return + await self._split(pieces, depth, "hollow", followup) + return + if followup: + return + rest = self._remainder(pieces, gains, truncated=False) + if rest: + await self._segment(rest, depth, True) + + async def _split(self, pieces: list[_Piece], depth: int, kind: str, followup: bool) -> None: + halves = _halves(pieces) if depth < int(C.MODEL_SPLIT_DEPTH) else None + if halves is not None: + 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) + return + results = await asyncio.gather(self._segment(first, depth + 1, followup), + self._segment(second, depth + 1, followup), 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) + return + self._fail(pieces, lambda p: self._reason(kind, split=depth > 0, piece=p)) + + def _remainder(self, pieces: list[_Piece], gains: list[_Gain], *, truncated: 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；有内容的只问没给的列。 + 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）。""" + out: list[_Piece] = [] + for piece, gain in zip(pieces, gains): + if not gain.named or not gain.items: + out.append(piece) + continue + cols = tuple(c for c in piece.columns if c not in gain.answered) + table_level = piece.table_level and truncated and not gain.table_items + if cols or table_level: + out.append(_Piece(piece.table, cols, table_level, piece.part)) + return out + + # ---------------------------------------------------------------- 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。""" + mode = mode or self.mode + retried = 0 + while True: + used = getattr(self.model, "budget", None) + try: + async with model_gate(): + await self._hook(self.hooks.call_started, [p.table for p in pieces]) + return await asyncio.wait_for(self._call(pieces, mode), timeout=C.MODEL_TIMEOUT_S) + 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) + if kind == "budget" and (getattr(self.model, "budget", None) != used or self._shrink()): + 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" + continue + if kind in ("transient", "timeout") and retried < int(C.MODEL_RETRIES): + 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)) + + async def _call(self, pieces: list[_Piece], mode: str) -> _Reply: + messages = self._messages(pieces, mode) + if mode == "structured": + runnable = self.model.with_structured_output(C.CATALOG_DRAFT_PROMPT_SCHEMA, include_raw=True) + raw, parsed = _unpack(await runnable.ainvoke(messages)) + await self._usage(raw) + return _structured_reply(raw, parsed) + raw = await self.model.ainvoke(messages) + await self._usage(raw) + return _text_reply(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。""" + 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)) + try: + text = completion.choices[0].message.content or "" # type: ignore[union-attr] + except (AttributeError, IndexError, TypeError): + text = "" + 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) + + def _shrink(self) -> bool: + shrink = getattr(self.model, "shrink", None) + if callable(shrink) and shrink(): + logger.info("起草数据目录：服务方不接受这么大的输出额度，这次起草改用 %s", getattr(self.model, "budget", "")) + return True + return False + + 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: + 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)) + + # ---------------------------------------------------------------- 拿到的内容 + + def _match(self, pieces: list[_Piece], entries: list[dict[str, Any]]) -> dict[str, list[dict[str, Any]]]: + """回复里的表对象按表名对到这一段的表上。表名照原样写的直接对上；带了 schema 前缀、大小写不一致的，只在恰好 + 对上一张时认。""" + names = [p.table for p in pieces] + out: dict[str, list[dict[str, Any]]] = {} + for entry in entries: + raw = entry.get("table") + if not isinstance(raw, str): + continue + name = raw if raw in names else None + if name is None: + short = raw.strip().rsplit(".", 1)[-1].strip('"`[] ') + found = [n for n in names if n == short or name_key(n) == name_key(short)] + name = found[0] if len(found) == 1 else None + if name is not None: + out.setdefault(name, []).append(entry) + return out + + def _take(self, pieces: list[_Piece], entries: list[dict[str, Any]]) -> list[_Gain]: + """回复里能用的并进各表（逐项校验，只留这一段请写的部分），返回每张表这次拿到了什么。""" + by_table = self._match(pieces, entries) + gains: list[_Gain] = [] + for piece in pieces: + st = self.tables[piece.table] + gain = _Gain() + for entry in by_table.get(piece.table, []): + gain.named = st.named = True + notes = _restrict(C._from_model(entry, st.meta, system=self.system), piece) + cols = entry.get("columns") if isinstance(entry.get("columns"), list) else [] + gain.answered |= {c.get("name") for c in cols if isinstance(c, dict) and c.get("name") in piece.columns} + slots = C._slots(notes) + gain.items += len(slots) + gain.table_items += sum(1 for key in slots if key[0] == "t") + self._merge(st, notes) + gains.append(gain) + return gains + + @staticmethod + def _merge(st: _Table, notes: Mapping[str, Any]) -> None: + for key, value in notes.items(): + if key == "columns": + cols = st.notes.setdefault("columns", {}) + for col, items in value.items(): + target = cols.setdefault(col, {}) + for name, item in items.items(): + target.setdefault(name, item) + st.got_columns.add(col) + else: + st.notes.setdefault(key, value) + st.got_table = True + + def _fail(self, pieces: list[_Piece], reason: str | Callable[[_Piece], str]) -> None: + """这几段没写出来：还没拿到内容的列（和表级的项）记下原因。""" + for piece in pieces: + st = self.tables[piece.table] + why = reason(piece) if callable(reason) else reason + for col in piece.columns: + if col not in st.got_columns: + st.failed.setdefault(col, why) + if piece.table_level and not st.got_table: + st.failed.setdefault(_TABLE, why) + + def _reason(self, kind: str, *, split: bool, piece: _Piece | None = None) -> str: + if kind == "exhausted": + base = REASON_EXHAUSTED.format(budget=self._budget()) + elif kind == "truncated": + base = REASON_TRUNCATED + else: + st = self.tables[piece.table] if piece is not None else None + if st is not None and st.whole(piece): + base = C.MODEL_EMPTY_REASON if st.named else REASON_MISSING + else: + base = REASON_HOLLOW + return base + (REASON_SPLIT if split else "") + + # ---------------------------------------------------------------- 收尾 + + def _outcome(self, st: _Table) -> ModelOutcome: + items = len(C._slots(st.notes)) + 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 + if not items: + reasons = list(dict.fromkeys(st.failed.values())) + reason = "；".join(reasons[:2]) if reasons 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) + 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) + + @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。""" + keys = ([_TABLE] if failed_table else []) + failed_cols + reason = "；".join(list(dict.fromkeys(st.failed[k] for k in keys))[:2]) + 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）" + 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" + elif failed_cols: + head = f"这张表 {n} 列，有 {len(failed_cols)} 列没写出来" + 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" + + async def _part_done(self, name: str) -> None: + self.tables[name].parts_done += 1 + await self._hook(self.hooks.part_done, name) + + async def _table_done(self, name: str) -> None: + st = self.tables[name] + if st.finished: + return + st.finished = True + await self._hook(self.hooks.table_done, name, self._outcome(st)) + + @staticmethod + async def _hook(fn: Callable[..., Awaitable[None]] | None, *args: Any) -> None: + if fn is None: + return + try: + await fn(*args) + except asyncio.CancelledError: + raise + except Exception: # noqa: BLE001 - 进度回调出错不影响起草 + logger.exception("起草数据目录的进度回调出错") + + # ---------------------------------------------------------------- 提示词 + + @staticmethod + def _background(description: Any, background: Background) -> str: + labels = C.CATALOG_DRAFT_PROMPT_LABELS + lines = [] + for key, text in (("description", description), ("guidance", background.guidance), ("note", background.note)): + text = text.strip() if isinstance(text, str) else "" + if text: + lines.append(labels[key].format(text=text)) + return C.CATALOG_DRAFT_BACKGROUND_PROMPT.format(lines="\n".join(lines)) if lines else "" + + def _messages(self, pieces: list[_Piece], mode: str) -> list[Any]: + from langchain_core.messages import HumanMessage, SystemMessage + + text = C.CATALOG_DRAFT_PROMPT.format(n=len(pieces), tables="\n\n".join(self._describe(p) for p in pieces)) + if self.background: + text = f"{self.background}\n\n{text}" + if mode == "text": + text = f"{text}\n\n{C.CATALOG_DRAFT_TEXT_PROMPT}" + return [SystemMessage(content=C.CATALOG_DRAFT_SYSTEM), HumanMessage(content=text)] + + def _describe(self, piece: _Piece) -> str: + """一段里一张表的结构描述：列、类型、主键、外键、唯一约束、注释；不含任何数据值。 + + 整表：照旧的完整描述。只写表级的项：全部列只给名字和类型。表级的项加几列：完整描述，表名下一行点名这几列。 + 只写几列：表注释、已起草的表级内容做背景，只列这几列。""" + labels = C.CATALOG_DRAFT_PROMPT_LABELS + st = self.tables[piece.table] + meta = st.meta + lines = [(labels["view"] if meta.get("is_view") else labels["table"]).format(name=st.name)] + brief = piece.table_level and not piece.columns + if not st.whole(piece): + if brief: + lines.append(labels["only_table"]) + elif piece.table_level: + lines.append(labels["table_and_columns"].format(n=len(piece.columns), cols="、".join(piece.columns))) + else: + lines.append(labels["only_columns"].format(n=len(piece.columns), total=len(st.columns))) + comment = (meta.get("comment") or "").strip() + if comment: + lines.append((labels["system"] if self.system else labels["comment"]).format(text=comment)) + pk = set(meta.get("primary_key") or []) + fk_of: dict[str, str] = {} + for fk in meta.get("foreign_keys") or []: + if isinstance(fk, dict) and len(fk.get("columns") or []) == 1: + to_cols = fk.get("to_columns") or [] + fk_of[fk["columns"][0]] = f"{fk.get('to_table')}.{to_cols[0]}" if to_cols else str(fk.get("to_table")) + if brief: + lines.append(labels["brief_columns"].format(n=len(st.columns))) + wanted = set(st.columns) + else: + if not piece.table_level: + drafted = C._notes_lines({k: v for k, v in st.notes.items() if k in C.TABLE_FIELDS}, meta=meta, + system_notes=False) + if drafted: + lines.append(labels["drafted"]) + lines += drafted + lines.append(labels["columns"]) + wanted = set(st.columns) if piece.table_level else set(piece.columns) + for col in meta.get("columns") or []: + if isinstance(col, dict) and col.get("name") in wanted: + lines.append(self._column_line(col, pk, fk_of, brief=brief)) + if piece.table_level: + for group in meta.get("unique") or []: + if isinstance(group, list) and group: + lines.append(labels["unique"].format(cols="、".join(group))) + confirmed = self._confirmed_lines(st, piece) + if confirmed: + lines.append(labels["confirmed"]) + lines += confirmed + return "\n".join(lines) + + def _column_line(self, col: Mapping[str, Any], pk: set[str], fk_of: Mapping[str, str], *, brief: bool) -> str: + labels = C.CATALOG_DRAFT_PROMPT_LABELS + marks = [] + if col["name"] in pk: + marks.append(labels["pk"]) + if not brief and not col.get("nullable", True): + marks.append(labels["not_null"]) + if col["name"] in fk_of: + marks.append(labels["fk"].format(target=fk_of[col["name"]])) + if not brief and (text := (col.get("comment") or "").strip()): + marks.append((labels["col_system"] if self.system else labels["col_comment"]).format(text=text)) + line = labels["column"].format(name=col["name"], type=col.get("type") or "") + return line + ("，" + "，".join(marks) if marks else "") + + @staticmethod + def _confirmed_lines(st: _Table, piece: _Piece) -> list[str]: + """已确认、有确证的项（以此为准的背景）：整表全给；其余给表级的项和这一段的列，关系只在要写表级的项时给。""" + confirmed = C._confirmed_only(st.existing) + if not st.whole(piece): + wanted = set(piece.columns) + kept = {key: item for key, item in C._slots(confirmed).items() + if key[0] == "t" or (key[0] == "c" and key[1] in wanted) or (key[0] == "r" and piece.table_level)} + confirmed = C._assemble(kept) + return C._notes_lines(confirmed, meta=st.meta, system_notes=False) + + +__all__ = ["BUDGET_WORDS", "Background", "Hooks", "ModelOutcome", "REASON_CALL", "REASON_EXHAUSTED", "REASON_FORMAT", + "REASON_HOLLOW", "REASON_MISSING", "REASON_RETRIED", "REASON_SPLIT", "REASON_TIMEOUT", "REASON_TRUNCATED", + "classify_failure", "model_gate", "parts_for", "run_model_draft"] diff --git a/backend/tests/fixtures/catalog/fake_model.py b/backend/tests/fixtures/catalog/fake_model.py index 42bfdec5..1ffd796d 100644 --- a/backend/tests/fixtures/catalog/fake_model.py +++ b/backend/tests/fixtures/catalog/fake_model.py @@ -3,10 +3,11 @@ model = FakeModel({"visits": {"label": "入园记录", "kind": "fact"}}, fail={"orders"}) await catalog.draft_with_model(model, source, ["visits"]) -- 结构化输出（with_structured_output(...).ainvoke）从提示词里认出这一批有哪些表（「表 <表名>」）， - 返回 {"tables": [{"table": 表名, **replies[表名]}, …]}； +- 结构化输出（with_structured_output(...).ainvoke）从提示词里认出这一段请它写哪些表、每张表写哪部分（requested： + 整张表、只写表级的项、只写列出的几列、表级的项加几列），返回 {"tables": [{"table": 表名, …}, …]}，只给请它写的 + 那部分（整张表给 replies[表名] 的全部；只写几列时只给 replies 里这几列）； - 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 里的表所在的那一批抛异常，用来验证「一批失败只影响这一批」； +- 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」； - 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 +23,51 @@ from __future__ import annotations import json from typing import Any +from app.data.catalog import CATALOG_DRAFT_PROMPT_LABELS as _P + +#: 表名下一行写着「这次只写…」时，请模型写的是哪一部分（data/catalog_draft.py 按这几句写提示词） +_ONLY_TABLE = _P["only_table"] +_ONLY_COLUMNS = _P["only_columns"].split("{")[0] +_TABLE_AND = _P["table_and_columns"].split("{")[0] + + +def requested(text: str) -> dict[str, tuple[str, list[str] | None]]: + """提示词里请模型写哪些表、每张表写哪部分：{表名: (哪部分, 列)}。哪部分是 full（整张表）、table（只写表级的项）、 + columns（只写列出的列，列是表结构描述里列出的那几列）、table+columns（表级的项加「这次只写」那一行点名的列）。""" + out: dict[str, tuple[str, list[str] | None]] = {} + for chunk in text.split("\n\n"): + lines = chunk.split("\n") + head = lines[0] + if not head.startswith("表 "): + continue + name = head[2:] + if name.endswith("（视图）"): + name = name[:-len("（视图）")] + second = lines[1] if len(lines) > 1 else "" + listed = [line[2:].split(" ")[0] for line in lines if line.startswith("- ")] + if second.startswith(_ONLY_TABLE): + out[name] = ("table", None) + elif second.startswith(_ONLY_COLUMNS): + out[name] = ("columns", listed) + elif second.startswith(_TABLE_AND): + out[name] = ("table+columns", second.split("：", 1)[1].split("、")) + else: + out[name] = ("full", None) + return out + + +def answer(reply: dict, name: str, part: str, columns: list[str] | None) -> dict: + """按请它写的那部分，从 replies[表名] 里取。""" + if part == "full": + return {"table": name, **reply} + out: dict[str, Any] = {"table": name} + if part in ("table", "table+columns"): + out.update({k: v for k, v in reply.items() if k != "columns"}) + if part in ("columns", "table+columns"): + wanted = set(columns or []) + out["columns"] = [c for c in reply.get("columns") or [] if isinstance(c, dict) and c.get("name") in wanted] + return out + class FakeModel: def __init__(self, replies: dict[str, dict], fail: set[str] = frozenset(), *, @@ -44,17 +90,19 @@ class FakeModel: text = "\n".join(str(getattr(m, "content", m)) for m in messages) self.calls.append(text) self.modes.append(mode) - names = [t for t in self.replies if f"表 {t}\n" in text or f"表 {t}（" in text] - if any(t in self.fail for t in names): + asked = requested(text) + names = [t for t in asked if t in self.replies] + if any(t in self.fail for t in asked): raise RuntimeError("网关超时") shape = self.structured if mode == "structured" else self.text if shape == "error": raise NotImplementedError("这条路用不了") if shape == "hollow": return {"tables": [{"table": t} for t in names]} + answers = [answer(self.replies[t], t, *asked[t]) for t in names] if shape == "single": - return {"table": names[0], **self.replies[names[0]]} if names else {} - return {"tables": [{"table": t, **self.replies[t]} for t in names]} + return answers[0] if answers else {} + return {"tables": answers} class _Structured: diff --git a/backend/tests/test_catalog_draft.py b/backend/tests/test_catalog_draft.py index 96162564..85e98d74 100644 --- a/backend/tests/test_catalog_draft.py +++ b/backend/tests/test_catalog_draft.py @@ -274,10 +274,11 @@ async def test_model_single_table_object_without_wrapper_is_recognized(): async def test_hollow_structured_output_retries_with_text(): - """结构化输出只给了 {"tables":[{"table":"visits"}]} 这样的空壳：改走纯文本再问一次，用纯文本给的内容。""" + """结构化输出只给了 {"tables":[{"table":"visits"}]} 这样的空壳、这一段又拆不开（不到 10 列）：改走纯文本再问一次， + 用纯文本给的内容。纯文本只给了表级的项、没给列：再补问一次没给的列（只问缺的，data/catalog_draft.py）。""" model = FakeModel({"visits": {"label": "入园记录", "kind": "fact"}}, structured="hollow", text="tables") got = await catalog.draft_with_model(model, _source(_scenic_tables()), ["visits"]) - assert model.modes == ["structured", "text"] + assert model.modes == ["structured", "text", "text"] assert got["visits"].notes == {"label": make_item("入园记录", "llm"), "kind": make_item("fact", "llm")} diff --git a/backend/tests/test_catalog_draft_engine.py b/backend/tests/test_catalog_draft_engine.py new file mode 100644 index 00000000..a1ac2272 --- /dev/null +++ b/backend/tests/test_catalog_draft_engine.py @@ -0,0 +1,555 @@ +"""数据目录模型起草的引擎（data/catalog_draft.py）：分段、按失败原因处理、捡回被截断的部分、起草要求、用量。 + +用假模型复刻实测查明的几种失败（不调用真实模型）： +- 思考加输出用完额度：finish_reason=length 且正文为空（只有思考）、LengthFinishReasonError → 拆成两半再问； +- 只给空壳 → 拆开再问；拆不开的换纯文本问一次，写出来了这次起草后面都走纯文本； +- 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。 + +表名、列名、说明文字都是虚构的。 +""" +from __future__ import annotations + +import asyncio +import inspect +import json +from dataclasses import dataclass, field +from types import SimpleNamespace +from typing import Any, Callable + +import pytest +from langchain_core.messages import AIMessage + +from app.data import catalog +from app.data import catalog_draft as D +from tests.fixtures.catalog.fake_model import answer, requested + + +# -------------------------------------------------------------------------- +# 夹具 +# -------------------------------------------------------------------------- + + +def _col(name: str, type_: str = "INTEGER") -> dict: + return {"name": name, "type": type_, "nullable": True, "comment": None} + + +def _table(n: int, prefix: str = "c", comment: str | None = None) -> dict: + return {"qualified": "", "schema": None, "columns": [_col(f"{prefix}{i:03d}") for i in range(1, n + 1)], + "primary_key": [f"{prefix}001"], "comment": comment, "is_view": False, "foreign_keys": []} + + +def _source(tables: dict, **kw) -> SimpleNamespace: + for name, meta in tables.items(): + meta["qualified"] = name + return SimpleNamespace(id="src", name="shop", kind="sqlite", origin="manual", + description=kw.pop("description", ""), schema_cache={"tables": tables}) + + +def _reply(table: str, meta: dict) -> dict: + """一张表的完整草稿：表级三项 + 每列中文名和度量类型。""" + return {"label": f"{table}的中文名", "grain": "每条记录一行", "kind": "fact", + "columns": [{"name": c["name"], "label": f"{c['name']}的中文名", "measure": "attribute"} + for c in meta["columns"]]} + + +@dataclass +class Reply: + """一次回复。obj 是正常的 JSON 对象；text 给了就原样作正文（截断的 JSON 之类）。""" + + obj: Any = None + text: str | None = None + finish: str = "stop" + thinking: str = "" + usage: tuple[int, int] = (0, 0) + + def render(self, mode: str, include_raw: bool) -> Any: + if self.text is not None: + content = self.text + else: + content = json.dumps(self.obj, ensure_ascii=False) if self.obj is not None else "" + msg = AIMessage( + content=content, + additional_kwargs={"reasoning_content": self.thinking} if self.thinking else {}, + response_metadata={"finish_reason": self.finish}, + usage_metadata={"input_tokens": self.usage[0], "output_tokens": self.usage[1], + "total_tokens": sum(self.usage)}, + ) + if mode == "text": + return msg + parsed = self.obj if self.text is None else None + if not include_raw: + return parsed + return {"raw": msg, "parsed": parsed, "parsing_error": None if parsed is not None else ValueError("坏的")} + + +@dataclass +class Call: + mode: str + text: str + asked: dict + n: int + + +class Scripted: + """每次调用交给 handler(call) 决定怎么回：返回 Reply，或者抛异常；handler 可以是协程。""" + + def __init__(self, handler: Callable[[Call], Any], name: str = "scripted-model") -> None: + self.handler, self.model_name = handler, name + self.calls: list[Call] = [] + + def with_structured_output(self, schema: Any, include_raw: bool = False, **_: Any) -> Any: + owner = self + + class _S: + async def ainvoke(self, messages): + return await owner._run(messages, "structured", include_raw) + + return _S() + + async def ainvoke(self, messages: list[Any]) -> Any: + return await self._run(messages, "text", False) + + async def _run(self, messages: list[Any], mode: str, include_raw: bool) -> Any: + text = "\n".join(str(getattr(m, "content", m)) for m in messages) + call = Call(mode=mode, text=text, asked=requested(text), n=len(self.calls)) + self.calls.append(call) + out = self.handler(call) + if inspect.isawaitable(out): + out = await out + return out.render(mode, include_raw) + + +def answering(tables: dict, **kw) -> Callable[[Call], Reply]: + """照请求写：每张表按请它写的那部分给出 _reply 的内容。""" + def handle(call: Call) -> Reply: + return Reply(obj={"tables": [answer(_reply(t, tables[t]), t, *call.asked[t]) for t in call.asked]}, **kw) + return handle + + +class _Status(Exception): + def __init__(self, status: int, text: str) -> None: + super().__init__(text) + self.status_code = status + + +class LengthFinishReasonError(Exception): + """和 openai 的同名异常一样：解析结构化输出时发现 finish_reason=length。""" + + def __init__(self) -> None: + super().__init__("Could not parse response content as the length limit was reached") + self.completion = None + + +class ConnectError(Exception): + pass + + +@pytest.fixture(autouse=True) +def _no_wait(monkeypatch): + """退避不真等：记下要等多久。""" + waited: list[float] = [] + + async def fake_sleep(seconds: float) -> None: + waited.append(seconds) + + monkeypatch.setattr(D, "_sleep", fake_sleep) + return waited + + +def _status(outcomes: dict) -> dict: + return {t: o.status for t, o in outcomes.items()} + + +# -------------------------------------------------------------------------- +# 思考用完额度、空壳：拆开再问 +# -------------------------------------------------------------------------- + + +async def test_thinking_used_up_the_budget_splits_the_segment(): + tables = {"orders_x": _table(24)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.n == 0: # 只想不写：正文为空、只有思考，额度用完 + return Reply(text="", finish="length", thinking="先逐列想一遍……") + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["orders_x"]) + o = got["orders_x"] + assert o.status == "ok" and o.error is None and o.items == 3 + 24 * 2 + assert len(model.calls) == 3 + cols = [f"c{i:03d}" for i in range(1, 25)] + assert model.calls[1].asked == {"orders_x": ("table+columns", cols[:12])} + assert model.calls[2].asked == {"orders_x": ("columns", cols[12:])} + # 后半段在前半段之后问，带着刚起草的表级内容做背景 + assert catalog.CATALOG_DRAFT_PROMPT_LABELS["drafted"] in model.calls[2].text + assert "orders_x的中文名" in model.calls[2].text + + +async def test_length_finish_reason_error_also_splits(): + tables = {"orders_x": _table(24)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.n == 0: + raise LengthFinishReasonError() + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["orders_x"]) + assert got["orders_x"].status == "ok" and len(model.calls) == 3 + + +async def test_hollow_shell_splits_the_pack_by_table(): + tables = {"a_tab": _table(6), "b_tab": _table(6)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.n == 0: + return Reply(obj={"tables": [{"table": "a_tab"}, {"table": "b_tab"}]}) + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["a_tab", "b_tab"]) + assert _status(got) == {"a_tab": "ok", "b_tab": "ok"} + assert [set(c.asked) for c in model.calls] == [{"a_tab", "b_tab"}, {"a_tab"}, {"b_tab"}] + 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(): + tables = {"wide40": _table(40)} + model = Scripted(lambda call: Reply(text="", finish="length", thinking="……")) + 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" + assert min(len(c.asked["wide40"][1] or []) for c in model.calls[1:]) == 5 + assert o.status == "failed" and o.items == 0 + assert o.error == "模型的思考和输出用完了额度（32768 个 token），这一段没写出来；已拆小重试仍没写出" + + small = Scripted(lambda call: Reply(text="", finish="length")) + got = await D.run_model_draft(small, _source({"t9": _table(9)}), ["t9"]) + assert len(small.calls) == 1, "不到 10 列拆不出两段 5 列" + 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_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。 + + 用宽表看先后：表级那段（拆不开）写完之后才问列分段。""" + tables = {"wide45": _table(45)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.mode == "structured": + return Reply(obj={"tables": [{"table": t} for t in call.asked]}) + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["wide45"]) + assert got["wide45"].status == "ok" and len(got["wide45"].notes["columns"]) == 45 + assert [c.mode for c in model.calls] == ["structured", "text", "text", "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" + + +# -------------------------------------------------------------------------- +# 截断、缺表、缺列 +# -------------------------------------------------------------------------- + + +async def test_truncated_json_keeps_complete_columns_and_asks_only_the_rest(): + tables = {"t12": _table(12)} + full = json.dumps({"tables": [{"table": "t12", **_reply("t12", tables["t12"])}]}, ensure_ascii=False) + cut = full[:full.index('"c006"') + 8] # 第 6 列写到一半 + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.n == 0: + return Reply(text=cut, finish="length") + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["t12"]) + o = got["t12"] + assert o.status == "ok" and set(o.notes["columns"]) == {f"c{i:03d}" for i in range(1, 13)} + assert o.notes["label"]["value"] == "t12的中文名" + assert len(model.calls) == 2 + assert model.calls[1].asked == {"t12": ("columns", [f"c{i:03d}" for i in range(6, 13)])} + + +async def test_truncated_tool_call_arguments_are_salvaged_too(): + """结构化输出走工具调用时，截断的参数在 invalid_tool_calls 里（一段字符串）。""" + tables = {"t12": _table(12)} + full = json.dumps({"tables": [{"table": "t12", **_reply("t12", tables["t12"])}]}, ensure_ascii=False) + cut = full[:full.index('"c004"') + 3] # 第 4 列的列名写到一半 + ok = answering(tables) + + class ToolModel(Scripted): + async def _run(self, messages, mode, include_raw): + if not self.calls: + self.calls.append(Call(mode, "", {}, 0)) + raw = AIMessage(content="", invalid_tool_calls=[{"name": "catalog_draft", "args": cut, "id": "x", + "error": None, "type": "invalid_tool_call"}], + response_metadata={"finish_reason": "length"}) + return {"raw": raw, "parsed": None, "parsing_error": ValueError("截断")} + return await super()._run(messages, mode, include_raw) + + model = ToolModel(ok) + got = await D.run_model_draft(model, _source(tables), ["t12"]) + assert got["t12"].status == "ok" + assert model.calls[1].asked == {"t12": ("columns", [f"c{i:03d}" for i in range(4, 13)])} + + +async def test_missing_tables_are_asked_again_once(): + tables = {"a_tab": _table(4), "b_tab": _table(4), "c_tab": _table(4)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + reply = ok(call) + keep = {0: {"a_tab"}, 1: {"b_tab"}}.get(call.n, set()) + reply.obj["tables"] = [e for e in reply.obj["tables"] if e["table"] in keep] + return reply + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["a_tab", "b_tab", "c_tab"]) + assert len(model.calls) == 2 + assert model.calls[1].asked == {"b_tab": ("full", None), "c_tab": ("full", None)} + assert _status(got) == {"a_tab": "ok", "b_tab": "ok", "c_tab": "failed"} + assert got["c_tab"].error == "模型的回复里没有这张表" + + +async def test_columns_the_model_skips_are_asked_once_and_then_left_empty(): + 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 ("c007", "c008")] + return reply + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["t8"]) + assert len(model.calls) == 2 and model.calls[1].asked == {"t8": ("columns", ["c007", "c008"])} + o = got["t8"] + assert o.status == "ok" and o.error is None, "补问过仍然没给的列算拿不准、留空" + assert set(o.notes["columns"]) == {f"c{i:03d}" for i in range(1, 7)} + + +# -------------------------------------------------------------------------- +# 额度被拒、结构化输出不支持、网络错误、超时 +# -------------------------------------------------------------------------- + + +async def test_budget_rejection_falls_back_to_8192_for_the_rest(): + tables = {"wide45": _table(45)} + + def refuse(call: Call) -> Reply: + raise _Status(400, "max_tokens: 32768 > 8192, which is the maximum allowed number of output tokens") + + big, small = Scripted(refuse), Scripted(answering(tables)) + model = catalog.DraftModel(big, small, "m") + got = await D.run_model_draft(model, _source(tables), ["wide45"]) + assert got["wide45"].status == "ok" + assert len(big.calls) == 1 and len(small.calls) == 3, "退回之后，表级那段重来、两个列分段都用小额度" + assert model.budget == 8192 + + +async def test_exhausted_message_names_the_budget_actually_used(): + def refuse(call: Call) -> Reply: + raise _Status(400, "max_tokens is too large") + + model = catalog.DraftModel(Scripted(refuse), Scripted(lambda c: Reply(text="", finish="length")), "m") + got = await D.run_model_draft(model, _source({"t4": _table(4)}), ["t4"]) + assert got["t4"].error == "模型的思考和输出用完了额度（8192 个 token），这一段没写出来" + + +async def test_structured_output_unsupported_switches_to_text_for_the_rest(): + tables = {"wide45": _table(45)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.mode == "structured": + raise _Status(400, "This response_format type is unavailable now") + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["wide45"]) + assert got["wide45"].status == "ok" + assert [c.mode for c in model.calls] == ["structured", "text", "text", "text"] + + +async def test_a_model_without_structured_output_goes_to_text(): + class TextOnly(Scripted): + def with_structured_output(self, schema, **_): + raise NotImplementedError + + tables = {"t4": _table(4)} + model = TextOnly(answering(tables)) + got = await D.run_model_draft(model, _source(tables), ["t4"]) + assert got["t4"].status == "ok" and [c.mode for c in model.calls] == ["text"] + + +async def test_network_errors_are_retried_with_backoff(_no_wait): + tables = {"t4": _table(4)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + if call.n < 2: + raise ConnectError("All connection attempts failed") + return ok(call) + + model = Scripted(handle) + got = await D.run_model_draft(model, _source(tables), ["t4"]) + assert got["t4"].status == "ok" and len(model.calls) == 3 + assert _no_wait == list(catalog.MODEL_RETRY_DELAYS_S) + + +async def test_server_errors_give_up_after_two_retries(): + def handle(call: Call) -> Reply: + raise _Status(503, "Service Unavailable") + + model = Scripted(handle) + got = await D.run_model_draft(model, _source({"t4": _table(4)}), ["t4"]) + assert len(model.calls) == 3 + assert got["t4"].status == "failed" + assert got["t4"].error == "调用模型出错：Service Unavailable（已重试 2 次）" + + +async def test_other_errors_are_not_retried(): + def handle(call: Call) -> Reply: + raise RuntimeError("网关拒绝了这次请求") + + model = Scripted(handle) + got = await D.run_model_draft(model, _source({"t4": _table(4)}), ["t4"]) + 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): + monkeypatch.setattr(catalog, "MODEL_TIMEOUT_S", 0.05) + + async def slow(call: Call) -> Reply: + await asyncio.sleep(1) + return Reply(obj={"tables": []}) + + 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）" + + +# -------------------------------------------------------------------------- +# 分段：宽表拆成表级一段 + 列分段，窄表拼段 +# -------------------------------------------------------------------------- + + +async def test_a_wide_table_is_split_into_a_table_part_and_column_parts(): + tables = {"wide95": _table(95, comment="虚构的宽表")} + model = Scripted(answering(tables)) + done: list[str] = [] + + async def part_done(table: str) -> None: + done.append(table) + + got = await D.run_model_draft(model, _source(tables), ["wide95"], hooks=D.Hooks(part_done=part_done)) + o = got["wide95"] + assert o.status == "ok" and o.parts_total == 4 and o.parts_done == 4 and done == ["wide95"] * 4 + assert len(o.notes["columns"]) == 95 and o.notes["label"]["value"] == "wide95的中文名" + 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] + 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)} + + +async def test_a_wide_table_with_one_failed_part_is_partial_and_says_which(): + tables = {"wide95": _table(95)} + ok = answering(tables) + + def handle(call: Call) -> Reply: + part, cols = call.asked["wide95"] + if part == "columns" and "c041" in cols: + raise RuntimeError("网关拒绝") + return ok(call) + + 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" + 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_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()} + model = Scripted(answering(tables)) + got = await D.run_model_draft(model, _source(tables), list(widths)) + assert set(_status(got).values()) == {"ok"} + asked = sorted((tuple(sorted(c.asked)), c.asked[next(iter(c.asked))][0]) for c in model.calls) + assert asked == sorted([ + (("ta", "tb"), "full"), (("tc", "td", "te", "tf"), "full"), (("tg",), "full"), + (("th",), "table"), (("th",), "columns"), (("th",), "columns"), + ]) + + +# -------------------------------------------------------------------------- +# 起草要求、用量 +# -------------------------------------------------------------------------- + + +async def test_guidance_goes_into_the_prompt_after_the_rules(): + tables = {"t4": _table(4)} + model = Scripted(answering(tables)) + src = _source(tables, description="虚构的景区票务库") + await D.run_model_draft(model, src, ["t4"], + background=D.Background(guidance="金额一律按元写单位", note="这次重点看会员相关的列")) + text = model.calls[0].text + block = ("使用者补充的背景和要求：\n数据源说明：虚构的景区票务库\n起草要求：金额一律按元写单位\n" + "本次说明：这次重点看会员相关的列") + assert block in text + rule = "它和这里的规矩冲突时，以这里的规矩为准" + assert rule in catalog.CATALOG_DRAFT_SYSTEM + assert text.index(rule) < text.index(block) < text.index("请为下面") < text.index("表 t4") + + plain = Scripted(answering(tables)) + await D.run_model_draft(plain, _source(tables), ["t4"]) + assert "使用者补充的背景和要求：" not in plain.calls[0].text, "都没有就不出现这一段（系统提示词里那条规矩照旧在）" + only_note = Scripted(answering(tables)) + await D.run_model_draft(only_note, _source(tables), ["t4"], background=D.Background(note="只看金额")) + assert "使用者补充的背景和要求：\n本次说明：只看金额\n\n请为下面" in only_note.calls[0].text + + +async def test_the_prompt_asks_not_to_draft_every_column_while_thinking(): + assert "思考时不必把要输出的内容逐列草拟一遍" in catalog.CATALOG_DRAFT_SYSTEM + + +async def test_usage_is_added_up_through_the_hooks(): + tables = {"orders_x": _table(24)} + ok = answering(tables, usage=(120, 60)) + + def handle(call: Call) -> Reply: + if call.n == 0: + 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)] + + +def test_budget_words_match_the_other_drafts(): + from app.data import recipe_ai + + assert D.BUDGET_WORDS == recipe_ai._BUDGET_WORDS From 90a560d07e6e07b8b5ebec82518cf7052f527ccf Mon Sep 17 00:00:00 2001 From: ThearchyHelios Date: Tue, 13 Oct 2026 15:24:57 +0800 Subject: [PATCH 03/29] =?UTF-8?q?=E6=95=B0=E6=8D=AE=E7=9B=AE=E5=BD=95?= =?UTF-8?q?=E8=B5=B7=E8=8D=89=E6=94=B9=E6=88=90=E5=90=8E=E5=8F=B0=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=EF=BC=88=E5=89=8D=E7=AB=AF=EF=BC=89=EF=BC=9A=E5=AF=B9?= =?UTF-8?q?=E8=AF=9D=E6=A1=86=E9=80=89=E8=8C=83=E5=9B=B4=E3=80=81=E5=86=99?= =?UTF-8?q?=E8=B5=B7=E8=8D=89=E8=A6=81=E6=B1=82=E5=92=8C=E6=9C=AC=E6=AC=A1?= =?UTF-8?q?=E8=A1=A5=E5=85=85=E8=AF=B4=E6=98=8E=E5=90=8E=E5=BB=BA=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=EF=BC=8C=E4=B9=8B=E5=90=8E=E6=98=BE=E7=A4=BA=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=E7=9A=84=E8=BF=9B=E5=BA=A6=E5=92=8C=E7=BB=93=E6=9E=9C?= =?UTF-8?q?=E2=80=94=E2=80=94=E6=80=BB=E8=BF=9B=E5=BA=A6=E3=80=81=E6=96=B0?= =?UTF-8?q?=E5=A2=9E=20/=20=E6=9B=B4=E6=96=B0=20/=20=E5=88=A0=E9=99=A4?= =?UTF-8?q?=E5=90=88=E8=AE=A1=E3=80=81=E9=80=90=E8=A1=A8=E7=8A=B6=E6=80=81?= =?UTF-8?q?=E3=80=81=E5=AE=BD=E8=A1=A8=E7=9A=84=E6=AE=B5=E8=BF=9B=E5=BA=A6?= =?UTF-8?q?=E3=80=81=E6=B2=A1=E5=86=99=E5=85=A8=E7=9A=84=E5=8E=9F=E5=9B=A0?= =?UTF-8?q?=EF=BC=9B=E5=8F=AF=E4=BB=A5=E5=81=9C=E6=AD=A2=E3=80=81=E5=9C=A8?= =?UTF-8?q?=E5=90=8E=E5=8F=B0=E7=BB=A7=E7=BB=AD=EF=BC=8C=E5=85=B3=E6=8E=89?= =?UTF-8?q?=E5=AF=B9=E8=AF=9D=E6=A1=86=E4=BB=BB=E5=8A=A1=E7=85=A7=E5=B8=B8?= =?UTF-8?q?=E8=BF=9B=E8=A1=8C=EF=BC=8C=E7=9B=AE=E5=BD=95=E9=A1=B5=E4=B8=8A?= =?UTF-8?q?=E6=96=B9=E6=98=BE=E7=A4=BA=E4=B8=80=E6=9D=A1=E8=BF=9B=E5=BA=A6?= =?UTF-8?q?=EF=BC=8C=E7=82=B9=E5=BC=80=E5=9B=9E=E5=88=B0=E5=AF=B9=E8=AF=9D?= =?UTF-8?q?=E6=A1=86=EF=BC=8C=E7=BB=93=E6=9D=9F=E6=97=B6=E6=8F=90=E7=A4=BA?= =?UTF-8?q?=E5=B9=B6=E5=8F=AF=E6=9F=A5=E7=9C=8B=E7=BB=93=E6=9E=9C=EF=BC=9B?= =?UTF-8?q?=E7=BB=93=E6=9E=9C=E9=87=8C=E5=8F=AA=E5=88=97=E6=B2=A1=E5=86=99?= =?UTF-8?q?=E5=85=A8=E7=9A=84=E8=A1=A8=EF=BC=8C=E3=80=8C=E5=8F=AA=E9=87=8D?= =?UTF-8?q?=E8=AF=95=E5=A4=B1=E8=B4=A5=E7=9A=84=E8=A1=A8=E3=80=8D=E6=96=B0?= =?UTF-8?q?=E5=BB=BA=E5=8F=AA=E5=90=AB=E5=A4=B1=E8=B4=A5=E3=80=81=E9=83=A8?= =?UTF-8?q?=E5=88=86=E5=AE=8C=E6=88=90=E3=80=81=E5=B7=B2=E5=81=9C=E6=AD=A2?= =?UTF-8?q?=E7=9A=84=E8=A1=A8=E7=9A=84=E4=BB=BB=E5=8A=A1=EF=BC=9B=E6=9C=8D?= =?UTF-8?q?=E5=8A=A1=E9=87=8D=E5=90=AF=E4=B8=AD=E6=96=AD=E7=9A=84=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=E5=86=99=E6=98=8E=E5=8E=9F=E5=9B=A0=E3=80=81=E5=90=8C?= =?UTF-8?q?=E6=A0=B7=E8=83=BD=E9=87=8D=E8=AF=95=EF=BC=9B=E5=B7=B2=E6=9C=89?= =?UTF-8?q?=E4=BB=BB=E5=8A=A1=EF=BC=88409=EF=BC=89=E8=BD=AC=E5=8E=BB?= =?UTF-8?q?=E6=98=BE=E7=A4=BA=E8=BF=9B=E8=A1=8C=E4=B8=AD=E7=9A=84=E9=82=A3?= =?UTF-8?q?=E4=B8=80=E4=B8=AA=EF=BC=8C=E8=B5=B7=E8=8D=89=E6=A8=A1=E5=9E=8B?= =?UTF-8?q?=E7=94=A8=E4=B8=8D=E4=BA=86=EF=BC=88409=EF=BC=89=E7=95=99?= =?UTF-8?q?=E5=9C=A8=E8=AE=BE=E7=BD=AE=E9=98=B6=E6=AE=B5=E5=86=99=E6=98=8E?= =?UTF-8?q?=E5=8E=9F=E5=9B=A0=EF=BC=9B=E8=B5=B7=E8=8D=89=E8=A6=81=E6=B1=82?= =?UTF-8?q?=E6=94=B9=E8=BF=87=E5=85=88=E5=AD=98=E5=86=8D=E5=BB=BA=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=EF=BC=8C=E5=86=99=E6=98=8E=E7=94=A8=E7=9A=84=E6=98=AF?= =?UTF-8?q?=E5=93=AA=E4=B8=AA=E8=B5=B7=E8=8D=89=E6=A8=A1=E5=9E=8B=E3=80=81?= =?UTF-8?q?=E6=95=B0=E6=8D=AE=E6=BA=90=E8=AF=B4=E6=98=8E=E4=B9=9F=E4=B8=80?= =?UTF-8?q?=E5=B9=B6=E4=BA=A4=E7=BB=99=E6=A8=A1=E5=9E=8B=EF=BC=9B=E6=B8=85?= =?UTF-8?q?=E5=8D=95=E4=B8=8A=E6=A0=87=E5=87=BA=E4=B8=8A=E6=AC=A1=E8=B5=B7?= =?UTF-8?q?=E8=8D=89=E5=A4=B1=E8=B4=A5=E3=80=81=E9=83=A8=E5=88=86=E5=AE=8C?= =?UTF-8?q?=E6=88=90=E7=9A=84=E8=A1=A8=E5=B9=B6=E5=86=99=E5=8E=9F=E5=9B=A0?= =?UTF-8?q?=EF=BC=8C=E8=A1=A8=E8=AF=A6=E6=83=85=E9=87=8C=E7=BB=99=E3=80=8C?= =?UTF-8?q?=E9=87=8D=E6=96=B0=E8=B5=B7=E8=8D=89=E8=BF=99=E5=BC=A0=E8=A1=A8?= =?UTF-8?q?=E3=80=8D=EF=BC=9B=E8=BF=9B=E8=A1=8C=E4=B8=AD=E6=AF=8F=201.5=20?= =?UTF-8?q?=E7=A7=92=E5=8F=96=E4=B8=80=E6=AC=A1=EF=BC=8C=E5=8F=96=E5=9B=9E?= =?UTF-8?q?=E7=9A=84=E7=8A=B6=E6=80=81=E5=8F=AA=E5=BE=80=E5=89=8D=E8=B5=B0?= =?UTF-8?q?=E3=80=82check-catalog=20=E7=94=A8=E5=81=87=E7=9A=84=E4=BB=BB?= =?UTF-8?q?=E5=8A=A1=E6=8E=A5=E5=8F=A3=E8=A6=86=E7=9B=96=E8=BF=99=E4=BA=9B?= =?UTF-8?q?=EF=BC=88=E4=BB=BB=E5=8A=A1=E6=80=8E=E4=B9=88=E5=BE=80=E5=89=8D?= =?UTF-8?q?=E8=B5=B0=E7=94=B1=E6=A3=80=E6=9F=A5=E8=87=AA=E5=B7=B1=E6=8E=A8?= =?UTF-8?q?=EF=BC=89=EF=BC=8C=E6=97=A7=E7=9A=84=E5=90=8C=E6=AD=A5=E8=B5=B7?= =?UTF-8?q?=E8=8D=89=E6=A3=80=E6=9F=A5=E6=8D=A2=E6=8E=89?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- frontend/src/api/client.ts | 40 +- frontend/src/lib/terms.ts | 144 ++++- frontend/src/pages/catalog/CatalogPage.tsx | 119 +++- frontend/src/pages/catalog/DraftDialog.tsx | 508 +++++++++++----- frontend/src/pages/catalog/TableDetail.tsx | 35 +- frontend/src/pages/catalog/TableIndex.tsx | 21 +- frontend/src/pages/catalog/draftJob.ts | 192 ++++++ frontend/src/types.ts | 99 +++- scripts/check-catalog.mjs | 657 +++++++++++++++++---- 9 files changed, 1479 insertions(+), 336 deletions(-) create mode 100644 frontend/src/pages/catalog/draftJob.ts diff --git a/frontend/src/api/client.ts b/frontend/src/api/client.ts index 00edd8da..e4e99d80 100644 --- a/frontend/src/api/client.ts +++ b/frontend/src/api/client.ts @@ -1,5 +1,5 @@ import type { - ActivateSnapshotBody, ActivateSnapshotOut, AiPreview, CatalogDetail, CatalogDraftOut, CatalogImpact, CatalogList, CatalogNotes, + ActivateSnapshotBody, ActivateSnapshotOut, AiPreview, CatalogDetail, CatalogDraftJob, CatalogImpact, CatalogList, CatalogNotes, CatalogPatchPreview, CatalogPatchSubmit, CatalogProfileOut, CatalogReviewAction, CommitOut, CurrentRecipe, EditPreview, EditRequest, ImportRecord, ManifestOut, PurgeRawBody, PurgeRawOut, QuestionAnswer, Recipe, RedraftRulesOut, RemovePeriodBody, RemovePeriodOut, @@ -31,6 +31,9 @@ export type { ApiErrorKind, RequestOptions, UploadOptions, UploadProgress } from const isObj = (v: unknown): v is Record => !!v && typeof v === 'object' && !Array.isArray(v) +/** 数据目录的接口前缀：/datasources/{id}/catalog */ +const catalogBase = (sourceId: string) => `/datasources/${encodeURIComponent(sourceId)}/catalog` + /** * 上传表格被退回、要用户先选怎么处理（422 带 decision）时，取出那份决定；别的错误返回 null。 * @@ -409,12 +412,37 @@ export const api = { post(`/datasources/${encodeURIComponent(sourceId)}/catalog/${encodeURIComponent(table)}/patch`, { changes, if_version: ifVersion }), /** - * 同步起草（注释、外键、命名推断；use_model 时再请助手的模型起草）。不给 tables 时按使用次数取前 20 张。 - * 人工确认、驳回过的项不动。没有表结构时回 409 + * 建一个起草任务（后台进行：注释、外键、命名推断；use_model 时再请起草模型起草，带上数据源的起草要求和这次的 note）。 + * 不给 tables 时按使用次数取前 20 张；表结构里没有的表照样建进任务、记 skipped。人工确认、驳回过的项不动。 + * 409：同一个数据源已有排队、进行中的任务（code 为 catalog_draft_busy，改去取 activeDraftJob）；起草模型用不了 + * （catalog_draft_model_unavailable，不建任务）；没有表结构 */ - draft: (sourceId: string, body: { tables?: string[]; use_model: boolean }, opts?: RequestOptions) => - request(`/datasources/${encodeURIComponent(sourceId)}/catalog/draft`, - { method: 'POST', body: JSON.stringify(body), ...opts }), + createDraftJob: (sourceId: string, body: { tables?: string[]; use_model: boolean; note?: string }) => + post<{ job: CatalogDraftJob }>(`${catalogBase(sourceId)}/draft-jobs`, body), + /** 排队或进行中的那个起草任务；没有为 null */ + activeDraftJob: (sourceId: string, opts?: RequestOptions) => + get<{ job: CatalogDraftJob | null }>(`${catalogBase(sourceId)}/draft-jobs/active`, opts), + /** 最近建的一个起草任务（任何状态）；没有为 null */ + latestDraftJob: (sourceId: string, opts?: RequestOptions) => + get<{ job: CatalogDraftJob | null }>(`${catalogBase(sourceId)}/draft-jobs/latest`, opts), + /** 一个起草任务的当前状态；不存在或不属于这个数据源回 404 */ + draftJob: (sourceId: string, jobId: string, opts?: RequestOptions) => + get<{ job: CatalogDraftJob }>(`${catalogBase(sourceId)}/draft-jobs/${encodeURIComponent(jobId)}`, opts), + /** + * 停止起草：排队的直接停；进行中的停掉正在进行的模型调用，没开始和没写完的表记已停止，已经写进目录的不回滚。 + * 已经结束的回 409（catalog_draft_finished） + */ + cancelDraftJob: (sourceId: string, jobId: string) => + post<{ job: CatalogDraftJob }>(`${catalogBase(sourceId)}/draft-jobs/${encodeURIComponent(jobId)}/cancel`), + /** + * 只重试失败的表：新任务只含原任务里失败、部分完成、已停止的表，用模型和补充说明沿用，起草要求取数据源现在的。 + * 没有可重试的表 422；有进行中的任务 409（catalog_draft_busy） + */ + retryDraftJob: (sourceId: string, jobId: string) => + post<{ job: CatalogDraftJob }>(`${catalogBase(sourceId)}/draft-jobs/${encodeURIComponent(jobId)}/retry`), + /** 存数据源的起草要求（去首尾空白，最多 2000 字，空串清除）。上传表格建的源存在它自己的数据源上 */ + putGuidance: (sourceId: string, guidance: string) => + put<{ guidance: string | null }>(`${catalogBase(sourceId)}/guidance`, { guidance }), /** * 同步剖析：对业务库发少量只读查询，核对关系、取码值候选，结论写进目录。不给 tables 时按使用次数取前 10 张 * 有待核实关系的表。没开启、数据源停用、没有表结构、同一个数据源正在剖析、快照被改动时回 409（detail 是一句中文） diff --git a/frontend/src/lib/terms.ts b/frontend/src/lib/terms.ts index 61073e35..bae8db47 100644 --- a/frontend/src/lib/terms.ts +++ b/frontend/src/lib/terms.ts @@ -43,10 +43,11 @@ */ import type { - CatalogCardinality, CatalogMeasure, CatalogProfileNumberKey, CatalogSource, CatalogStatus, CatalogTableKind, NodeType, SnapshotReasonCode, + CatalogCardinality, CatalogDraftJobStatus, CatalogDraftTaskStatus, CatalogMeasure, CatalogProfileNumberKey, CatalogSource, CatalogStatus, + CatalogTableKind, NodeType, SnapshotReasonCode, SqlCheckCode, SqlCheckLevel, ToolTrust, } from '../types' -import { formatCost, formatNumber } from './format' +import { formatCost, formatNumber, formatTokens } from './format' export const TERMS = { workflow: '工作流', @@ -2198,7 +2199,7 @@ export const CATALOG_TEXT = { noSchemaBody: (note: string) => `${note.replace(/[。.]$/, '')}。数据目录按表结构逐表记录，请先回到数据源卡片点击「探查结构」。`, noSchemaUploadBody: '这个表格还没有可用的表结构，请重新上传后再编写数据目录。', emptyTitle: '还没有数据目录', - emptyBody: '先起草，再逐表确认。起草按数据库注释、外键约束和列名推断，也可以请助手的模型起草中文名和含义。', + emptyBody: '先起草，再逐表确认。起草按数据库注释、外键约束和列名推断，也可以用模型起草中文名和含义。', // ---- 导入表格的说明 systemTitle: '导入时生成的说明', systemBody: '这个数据源由导入表格生成，表和列的说明由系统按核对结果生成，只读，不能在数据目录中改写。', @@ -2260,29 +2261,17 @@ export const CATALOG_TEXT = { scopeAll: (n: number) => `全部 ${formatNumber(n)} 张表`, draftBase: '按数据库注释、外键约束和列名推断起草。已确认、已驳回的项不会被改动；起草出的项都是推断，需要逐项确认。', useModel: '用模型起草', - useModelHint: '请助手的模型按表结构起草中文名、说明、粒度和列的含义。只发送表结构（表名、列名、类型、约束、注释），不发送数据；耗时较长，并产生模型费用。', + useModelHint: '请起草模型按表结构起草中文名、说明、粒度和列的含义。只发送表结构（表名、列名、类型、约束、注释）和起草要求，不发送数据；在后台进行，耗时较长，并产生模型费用。', draftStart: (n: number) => `起草 ${formatNumber(n)} 张表`, - drafting: '正在起草…', - draftProgress: (done: number, total: number) => `已完成 ${formatNumber(done)} / ${formatNumber(total)} 张表`, - draftStop: '停止', - draftStopping: '当前这一批完成后停止…', - draftDone: '起草完成', - draftStopped: (done: number, total: number) => `已停止：完成 ${formatNumber(done)} / ${formatNumber(total)} 张表，其余未起草`, - draftFailed: '起草中断', draftSummary: (added: number, updated: number, removed: number) => `新增 ${formatNumber(added)} 项、更新 ${formatNumber(updated)} 项${removed ? `、删除 ${formatNumber(removed)} 项` : ''}`, - // 只在起草确实给出了内容、只是和现有目录一致时用：有表的模型起草失败或没给出可用内容（model_error）时 - // 不能说「没有变化」，改用 draftNothingWritten + // 只在起草确实给出了内容、只是和现有目录一致时用：模型没给出可用内容时说 draftNothingWritten，有表没写全时说 + // draftNothingWrittenFailed，不能说「没有变化」 draftNoChange: '没有变化：起草出的内容与现有目录一致', draftNothingWritten: '目录没有更新：模型没有给出可用的内容，原因见下方', - draftModelSkipped: (reason: string) => `模型未参与起草：${reason.replace(/[。.]$/, '')}。已按注释、外键和命名起草。`, - draftModelUsed: (model: string) => `模型：${model}`, - draftTableErrors: (n: number) => `${formatNumber(n)} 张表未能起草`, - draftModelErrors: (n: number) => `${formatNumber(n)} 张表的模型起草失败，已按注释、外键和命名起草`, + draftNothingWrittenFailed: '目录没有更新：部分表未写全，原因见下方', draftClose: '完成', - moreTables: (n: number) => `…另有 ${formatNumber(n)} 张`, cancel: '取消', - draftBackground: '起草仍在进行，当前这一批完成后停止', // ---- 表详情 backToList: '返回表清单', prev: '上一张', @@ -2389,6 +2378,123 @@ export const CATALOG_TEXT = { reloadDiscardAction: '放弃修改并重新载入', } +// =========================================================================== +// 数据目录的起草任务（/api/datasources/{id}/catalog/draft-jobs）：起草在服务端后台进行，关掉对话框不停；逐表起草， +// 每张表一结束就写进目录。 +// +// 术语： +// - 「起草任务」的状态：排队中、正在起草、起草完成、部分表未写全、起草失败、已停止、服务重启中断。 +// 一张表：等待、起草中、已完成、部分完成、失败、已跳过（表结构里没有）、已停止。 +// - 「起草要求」：使用者补充给模型的业务背景和用词约定，存在数据源上，以后每次用模型起草都带上； +// 「本次补充说明」只用于这一次（重试时沿用）。两者都只交给模型，不改起草规则：冲突时以规则为准。 +// - 「起草模型」：设置里单独的一项（偏好设置 → 起草模型），留空跟随助手的模型。 +// - 宽表的「段」：超过 40 列的表拆成表级一段和若干列分段，逐段起草；写的是「已完成 3 / 9 段」。 +// - 「只重试失败的表」：重新起草失败、部分完成和已停止的表；已经写进目录的内容不回滚、不受影响。 +// =========================================================================== + +/** 起草要求、本次补充说明的字数上限（服务端同一个数：超长 422） */ +export const DRAFT_GUIDANCE_MAX = 2000 +export const DRAFT_NOTE_MAX = 1000 + +export const CATALOG_DRAFT_TEXT = { + // ---- 设置阶段：起草要求、本次补充说明、起草模型 + guidance: '起草要求', + guidanceHint: '写给模型的业务背景和用词约定，保存在这个数据源上，以后每次用模型起草都会带上。例如：「核销」指检票入园；金额单位是元', + guidanceRule: '起草要求只帮助模型理解业务、统一用词；与起草规则冲突时以规则为准，起草出的项仍需逐项确认', + guidancePlaceholder: '业务背景、常用说法、单位约定……', + note: '本次补充说明', + noteHint: '只用于这一次起草，例如「这一批是会员相关的表」；重试时沿用', + notePlaceholder: '可不填', + count: (n: number, max: number) => `${formatNumber(n)} / ${formatNumber(max)} 字`, + tooLong: (max: number) => `不超过 ${formatNumber(max)} 字`, + sourceDescription: (text: string) => `数据源说明也会一并交给模型：${text}`, + model: '起草模型', + /** 设置里没配起草模型：跟随助手的模型，写出当前是哪个（取不到就不写） */ + modelFollow: (current: string | null) => (current ? `跟随助手的模型（当前：${current}）` : '跟随助手的模型'), + /** 设置里配了：模型（没填模型就是那个接入的默认模型）和接入 */ + modelNamed: (model: string | null, provider: string | null) => (model + ? (provider ? `${model}（模型接入「${provider}」）` : model) + : `模型接入「${provider}」的默认模型`), + modelWhere: '可在「设置 → 偏好设置 → 起草模型」中更换', + tooMany: (max: number) => `一次最多起草 ${formatNumber(max)} 张表，请缩小范围`, + saveGuidanceFailed: (reason: string) => `起草要求未能保存，未开始起草：${reason}`, + startFailed: (reason: string) => `未能开始起草：${reason}`, + // ---- 上一次起草（设置阶段，没写全的才提） + last: (state: string, at: string) => `上一次起草（${at}）：${state}`, + lastOpen: '查看', + // ---- 进行中 + queued: '排队中，等待开始…', + running: '正在起草…', + stopping: '正在停止：当前的模型调用结束后停止…', + progress: (done: number, total: number) => `已完成 ${formatNumber(done)} / ${formatNumber(total)} 张表`, + /** 结束以后：全部写出来的表有几张（停止、失败、部分完成的不算） */ + progressDone: (ok: number, total: number) => `写完 ${formatNumber(ok)} / ${formatNumber(total)} 张表`, + parts: (done: number, total: number) => `${formatNumber(done)} / ${formatNumber(total)} 段`, + partsHint: '超过 40 列的表拆成表级一段和若干列分段，逐段起草', + hide: '在后台继续', + hideHint: '关闭对话框不会停止起草，进度显示在目录页上方；每张表起草完就写进目录', + stop: '停止', + busy: '这个数据源已有进行中的起草任务，下面显示的是它的进度', + modelUsed: (model: string) => `模型：${model}`, + usage: (calls: number, input: number, output: number) => + `调用模型 ${formatNumber(calls)} 次，输入 ${formatTokens(input, { compact: true })}、输出 ${formatTokens(output, { compact: true })}`, + // ---- 结束 + title: { + queued: '排队中，等待开始…', + running: '正在起草…', + succeeded: '起草完成', + partial: '起草完成，部分表未写全', + failed: '起草失败', + cancelled: '已停止', + interrupted: '服务重启，起草中断', + } as Record, + interrupted: '服务重启时这次起草还没有完成，没写完的表未起草。已写进目录的内容保留；可以只重试这些表，接着起草', + cancelled: '已写进目录的内容保留，没写完的表未起草', + jobFailed: (reason: string) => `起草任务出错：${reason}`, + /** 没写全的表：失败、部分完成、已停止各几张；跳过的（表结构里没有）另说 */ + problems: (c: { failed: number; partial: number; cancelled: number; skipped: number }) => { + const parts = [ + c.failed ? `失败 ${formatNumber(c.failed)} 张` : '', + c.partial ? `部分完成 ${formatNumber(c.partial)} 张` : '', + c.cancelled ? `已停止 ${formatNumber(c.cancelled)} 张` : '', + ].filter(Boolean) + const bad = parts.length ? `${formatNumber(c.failed + c.partial + c.cancelled)} 张表未写全：${parts.join('、')}` : '' + const skipped = c.skipped ? `${formatNumber(c.skipped)} 张表已跳过（表结构里没有）` : '' + return [bad, skipped].filter(Boolean).join('；') + }, + retry: (n: number) => `只重试失败的表（${formatNumber(n)} 张）`, + retryHint: '重新起草失败、部分完成和已停止的表；已写进目录的内容不受影响', + showAll: (n: number) => `显示全部 ${formatNumber(n)} 张表`, + showProblems: '只看未写全的表', + tasks: '逐表进度', + taskStatus: { + pending: '等待', running: '起草中', ok: '已完成', partial: '部分完成', failed: '失败', skipped: '已跳过', cancelled: '已停止', + } as Record, + taskAdded: (added: number, updated: number) => (added || updated + ? `+${formatNumber(added)}${updated ? ` · 改 ${formatNumber(updated)}` : ''}` : ''), + taskAddedHint: (added: number, updated: number, removed: number) => + `新增 ${formatNumber(added)} 项、更新 ${formatNumber(updated)} 项${removed ? `、删除 ${formatNumber(removed)} 项` : ''}`, + // ---- 目录页上方的进度、结束提示 + strip: (done: number, total: number) => `正在起草：已完成 ${formatNumber(done)} / ${formatNumber(total)} 张表`, + stripQueued: (total: number) => `起草任务排队中，共 ${formatNumber(total)} 张表`, + stripOpen: '查看进度', + toast: { + succeeded: (summary: string) => `数据目录起草完成：${summary}`, + partial: (n: number) => `数据目录起草完成，${formatNumber(n)} 张表未写全`, + failed: (reason: string) => `数据目录起草失败：${reason}`, + cancelled: '数据目录起草已停止', + interrupted: '服务重启，数据目录起草中断', + }, + viewResult: '查看结果', + // ---- 清单上的标记、表详情里的说明 + lastFailed: '上次起草失败', + lastPartial: '上次起草部分完成', + lastWith: (label: string, reason: string) => `${label}：${reason}`, + redraft: '重新起草这张表', + redraftHint: '用模型重新起草这张表（不带补充说明，带上数据源的起草要求）；人工确认、驳回过的项不会被改动', + redraftBusy: '已有起草任务在进行，结束后再重新起草这张表', +} + // =========================================================================== // 目录修改提案：用户在助手（画布）或问数据页说出一条数据事实时，助手提出的数据目录修改。 // diff --git a/frontend/src/pages/catalog/CatalogPage.tsx b/frontend/src/pages/catalog/CatalogPage.tsx index 48442f3a..15f8e24e 100644 --- a/frontend/src/pages/catalog/CatalogPage.tsx +++ b/frontend/src/pages/catalog/CatalogPage.tsx @@ -7,9 +7,13 @@ import { ApiError, api } from '../../api/client' import type { CatalogDetail, CatalogList, CatalogTableRow } from '../../types' import { EmptyState, ErrorState, PageHeader, Skeleton, Spinner, toast, useTicker } from '../../components/ui' import { useCatalog, useDatasources, useOnReconnect } from '../../store/catalog' -import { CATALOG_TEXT as CT, PROFILE_STOP_LABEL, PROFILE_TEXT as PT, SCHEMA_PARTIAL_TEXT as SPT } from '../../lib/terms' +import { + CATALOG_DRAFT_TEXT as DT, CATALOG_TEXT as CT, PROFILE_STOP_LABEL, PROFILE_TEXT as PT, SCHEMA_PARTIAL_TEXT as SPT, +} from '../../lib/terms' import { StatusLegend, SystemNotesNotice } from './parts' import { DraftDialog } from './DraftDialog' +import { isJobActive, retryableCount, useDraftJob } from './draftJob' +import type { DraftStartBody } from './draftJob' import { ProfileDialog } from './ProfileDialog' import type { ProfileJob } from './ProfileDialog' import { ProfileSettingsDialog } from './ProfileSettings' @@ -99,7 +103,11 @@ export function CatalogPage() { const search = writeListParams(params, list).toString() searchRef.current = search ? `?${search}` : '' const [selected, setSelected] = useState>(() => new Set()) - const [drafting, setDrafting] = useState(false) + /** + * 起草对话框开着没有、打开时显示什么：setup 选范围开始；job 显示页面跟着的任务（进行中的进度、刚结束的结果）。 + * busy：是撞上已有任务（409）转过去显示的 + */ + const [draftOpen, setDraftOpen] = useState<{ view: 'setup' | 'job'; busy?: boolean } | null>(null) /** 起草写进了这些表：详情据此重新载入（编辑中的不动，保存时服务端会用版本号拦下） */ const [drafted, setDrafted] = useState<{ seq: number; tables: Set }>({ seq: 0, tables: new Set() }) const [profiling, setProfiling] = useState(false) @@ -142,11 +150,59 @@ export function CatalogPage() { })) }, []) - const onDrafted = useCallback((tables: string[]) => { - setDrafted((cur) => ({ seq: cur.seq + 1, tables: new Set(tables) })) - setSelected(new Set()) - void load() - }, [load]) + /** + * 起草任务（后台进行，状态在页面上）：每张表一结束就重取清单、正在看的表在里面就重新载入（编辑中的不动，保存时服务端 + * 会用版本号拦下）；结束时重取一次清单。对话框关着时结束的，弹提示并给「查看结果」 + */ + const draftOpenRef = useRef(draftOpen) + draftOpenRef.current = draftOpen + const draft = useDraftJob(sourceId, data?.active_job, { + onTablesEnded: (tables) => { + setDrafted((cur) => ({ seq: cur.seq + 1, tables: new Set(tables) })) + void load() + }, + onFinished: (job) => { + void load() + if (draftOpenRef.current) return + const src = job.source_id + const opts = { + key: `catalog-draft:${src}`, duration: 10000, + action: { label: DT.viewResult, onClick: () => { if (sourceRef.current === src) setDraftOpen({ view: 'job' }) } }, + } + if (job.status === 'succeeded') toast.ok(DT.toast.succeeded(CT.draftSummary(job.added, job.updated, job.removed)), opts) + else if (job.status === 'partial') toast.warn(DT.toast.partial(retryableCount(job)), opts) + else if (job.status === 'failed') toast.error(DT.toast.failed(job.error ?? ''), opts) + else if (job.status === 'interrupted') toast.warn(DT.toast.interrupted, opts) + else toast.warn(DT.toast.cancelled, opts) + }, + }) + const draftJob = draft.job + const draftActive = isJobActive(draftJob) + /** 打开起草对话框：有进行中的任务就直接显示它；上一次的结果看过了，再点就是重新开始 */ + const openDraft = useCallback(() => setDraftOpen({ view: draftActive ? 'job' : 'setup' }), [draftActive]) + const startDraft = useCallback(async (body: DraftStartBody) => { + const r = await draft.start(body) + // 选中的表已经交给起草任务了：清空，免得接着点「起草所选」又起草一遍 + if (!r.busy) setSelected(new Set()) + return r + }, [draft.start]) + /** 详情里「重新起草这张表」：只含这张表、用模型、不带补充说明。模型用不了等开始不了的，直接提示原因 */ + const redraftTable = useCallback(async (t: string) => { + try { + const r = await draft.start({ tables: [t], use_model: true }) + setDraftOpen({ view: 'job', busy: r.busy }) + } catch (e) { + toast.error(e) + } + }, [draft.start]) + const onGuidanceSaved = useCallback((g: string | null) => { + setData((cur) => cur && ({ ...cur, guidance: g })) + }, []) + // 上方的进度：跟着的任务在进行就按它；跟着的就是清单说的那一个、已经结束了，就不信清单（清单还没重取回来） + const listActive = isJobActive(data?.active_job) ? data?.active_job ?? null : null + const stripJob: { done: number; total: number; status: string } | null = draftJob && (draftActive || draftJob.id === listActive?.id) + ? (draftActive ? draftJob : null) + : listActive const options = savedOptions?.sourceId === sourceId ? savedOptions.options : source?.options const profileSettings = useMemo(() => profileSettingsOf(options), [options]) @@ -262,7 +318,7 @@ export function CatalogPage() { const empty = rows.length > 0 && rows.every((r) => progressOf(r.counts) === 'none') const hasTables = rows.length > 0 const draftButton = (primary = false) => ( - ) @@ -281,7 +337,7 @@ export function CatalogPage() { ) // 页头的「起草」窄屏只留图标，标题才放得下 const headerDraft = ( - ) @@ -335,7 +391,7 @@ export function CatalogPage() { onSelected={setSelected} active={table} onOpen={openTable} - onDraft={() => setDrafting(true)} + onDraft={openDraft} onProfile={openProfile} /> @@ -343,7 +399,8 @@ export function CatalogPage() { {table ? ( + onFillDone={onFillDone} onOpen={openTable} onBack={backToList} onDetail={onDetail} + draftBusy={!!stripJob} onRedraft={redraftTable} /> ) : } @@ -373,16 +430,27 @@ export function CatalogPage() { onGoSource={() => navigate(back)} /> )} {hasTables && usage.used > 0 && !empty && } + {hasTables && stripJob && !draftOpen && } {body} - {drafting && data && ( + {draftOpen && data && ( setDrafting(false)} - onDrafted={onDrafted} + job={draftJob} + initial={draftOpen.view} + initialBusy={!!draftOpen.busy} + stopping={draft.stopping} + onStart={startDraft} + onCancel={draft.cancel} + onRetry={draft.retry} + onTrack={draft.track} + onGuidanceSaved={onGuidanceSaved} + onClose={() => setDraftOpen(null)} /> )} {profiling && data && ( @@ -412,6 +480,29 @@ export function CatalogPage() { ) } +/** + * 起草任务在进行：目录页上方一条进度（关掉对话框任务照常进行，每张表起草完就写进目录）。点「查看进度」回到对话框 + */ +function DraftStrip({ job, onOpen }: { job: { done: number; total: number; status: string }; onOpen: () => void }) { + const text = job.status === 'queued' ? DT.stripQueued(job.total) : DT.strip(job.done, job.total) + return ( +
+ +
+

{text}

+
+
+
+
+ +
+ ) +} + /** 媒体查询的当前结果，跟着窗口大小变 */ function useMedia(query: string): boolean { const [on, setOn] = useState(() => typeof matchMedia !== 'function' || matchMedia(query).matches) diff --git a/frontend/src/pages/catalog/DraftDialog.tsx b/frontend/src/pages/catalog/DraftDialog.tsx index c9fba7fc..47e11c65 100644 --- a/frontend/src/pages/catalog/DraftDialog.tsx +++ b/frontend/src/pages/catalog/DraftDialog.tsx @@ -1,66 +1,41 @@ -import { useMemo, useRef, useState } from 'react' -import { Sparkles } from 'lucide-react' +import { useEffect, useMemo, useRef, useState } from 'react' +import { CircleCheck, CircleDashed, CircleMinus, CircleSlash, CircleX, RotateCcw, Sparkles, TriangleAlert } from 'lucide-react' +import type { LucideIcon } from 'lucide-react' import clsx from 'clsx' -import { api } from '../../api/client' -import type { CatalogDraftRow, CatalogTableRow } from '../../types' -import { Modal, Notice, Spinner } from '../../components/ui' +import { ApiError, api } from '../../api/client' +import type { CatalogDraftJob, CatalogDraftTask, CatalogDraftTaskStatus, CatalogTableRow } from '../../types' +import { Field, Modal, Notice, Spinner } from '../../components/ui' import { errorMessage } from '../../lib/errors' -import { CATALOG_TEXT as CT } from '../../lib/terms' +import { formatTime } from '../../lib/format' +import { CATALOG_DRAFT_TEXT as DT, CATALOG_TEXT as CT, DRAFT_GUIDANCE_MAX, DRAFT_NOTE_MAX } from '../../lib/terms' +import { + DRAFT_MAX_TABLES, DRAFT_MODEL_UNAVAILABLE, DRAFT_TOP_N, RETRYABLE, isJobActive, nothingWritten, retryableCount, taskReason, +} from './draftJob' +import type { DraftStartBody } from './draftJob' // =========================================================================== -// 起草数据目录。接口是同步的，一次请求起草完才返回；表多时分批调用，每批回来就更新进度和累计的新增、更新、删除， -// 中途可以停止（当前这一批做完后不再发下一批）。模型整体用不了时第一批就会知道：之后的批次不再请模型，结果里 -// 写明原因。已经写进去的批次不回滚——起草只增改推断项，人工确认、驳回过的项服务端不会动。 +// 起草数据目录。起草是服务端的后台任务：这里选范围、写起草要求和本次补充说明，建任务；之后显示页面跟着的那个任务 +// （进行中的进度、结束后的结果），任务本身的状态在页面上（draftJob.ts），关掉对话框照常进行。 +// +// 三段：设置（范围、用不用模型；用模型时写起草要求、补充说明，写明用哪个起草模型）→ 进行中（总进度、合计、逐表的 +// 状态和宽表的段进度、没写全的原因；可以停止、在后台继续）→ 结束（汇总、没写全的表和原因、只重试失败的表）。 +// 已经写进目录的内容不回滚——起草只增改推断项，人工确认、驳回过的项服务端不会动。 // =========================================================================== -/** 每批几张表。只按注释、外键和命名起草时只看探查缓存，很快；请模型时服务端每 4 张一批、最多 3 批并发 */ -const BATCH_PLAIN = 40 -const BATCH_MODEL = 12 -/** 服务端不给 tables 时取前 20 张，这里给出同样的范围，进度才算得出来 */ -const TOP_N = 20 -/** 结果里逐张列出的表最多几张，其余合成一句 */ -const LIST_MAX = 8 - type Scope = 'selected' | 'filtered' | 'top' | 'all' -type Phase = 'setup' | 'running' | 'done' -interface Outcome { - done: number - total: number - added: number - updated: number - removed: number - tableErrors: CatalogDraftRow[] - modelErrors: CatalogDraftRow[] - modelSkipped: string | null - model: string | null - failure: string | null - stopped: boolean - /** 这一次勾了「用模型起草」 */ - modelAsked: boolean - /** 模型给了可用内容（model_items > 0）的表有几张；服务端没给 model_items（老服务端）时 modelItemsKnown 为假 */ - modelItems: number - modelItemsKnown: boolean -} +/** 字数按服务端的算法：去掉首尾空白后的字符数（按码位，不按 UTF-16 的长度） */ +const charCount = (s: string) => [...s.trim()].length -const fresh = (total: number, modelAsked = false): Outcome => ({ - done: 0, total, added: 0, updated: 0, removed: 0, tableErrors: [], modelErrors: [], modelSkipped: null, model: null, - failure: null, stopped: false, modelAsked, modelItems: 0, modelItemsKnown: false, -}) - -/** - * 一项都没写进去时怎么说。「没有变化」只在确实是「起草出了内容、和现有目录一致」时说：有表没起草、模型在某些表上 - * 没给出内容（model_error），或者勾了用模型、却没有一张表拿到模型的可用内容（模型整体用不了、model_items 都是 0）， - * 都说「目录没有更新」，原因写在下面 - */ -function nothingWrittenText(out: Outcome): string { - const failed = out.modelErrors.length > 0 || out.tableErrors.length > 0 - const modelEmpty = out.modelAsked && (!!out.modelSkipped || (out.modelItemsKnown && out.modelItems === 0)) - return failed || modelEmpty ? CT.draftNothingWritten : CT.draftNoChange -} - -export function DraftDialog({ sourceId, rows, visible, selected, filtered, onClose, onDrafted }: { +export function DraftDialog({ + sourceId, sourceDescription, guidance, rows, visible, selected, filtered, job, initial, initialBusy = false, stopping, + onStart, onCancel, onRetry, onTrack, onGuidanceSaved, onClose, +}: { sourceId: string + /** 数据源说明：用模型起草时连同起草要求一起交给模型 */ + sourceDescription: string + /** 数据源存着的起草要求（没有为空串） */ + guidance: string /** 全部表（服务端的顺序：使用次数多的在前） */ rows: CatalogTableRow[] /** 当前筛选结果 */ @@ -68,11 +43,25 @@ export function DraftDialog({ sourceId, rows, visible, selected, filtered, onClo selected: string[] /** 清单上有没有生效的搜索或筛选：有时才给「当前筛选结果」这一项 */ filtered: boolean + /** 页面跟着的起草任务（进行中的、刚结束的） */ + job: CatalogDraftJob | null + /** 打开时显示什么：setup 选范围、开始；job 显示页面跟着的那个任务 */ + initial: 'setup' | 'job' + /** 打开时显示的任务是撞上已有任务（409）转过来的：写明 */ + initialBusy?: boolean + /** 点了停止、服务端还没停下来 */ + stopping: boolean + /** 建任务；409 已有任务时页面改取进行中的那一个（busy 为真） */ + onStart: (body: DraftStartBody) => Promise<{ job: CatalogDraftJob; busy: boolean }> + onCancel: () => Promise + onRetry: (from: CatalogDraftJob) => Promise<{ job: CatalogDraftJob; busy: boolean }> + /** 设置阶段点开了上一次起草的结果（或者发现别处建的进行中的任务）：页面改跟这个任务 */ + onTrack: (job: CatalogDraftJob) => void + /** 起草要求存好了：页面记下新的这一份 */ + onGuidanceSaved: (guidance: string | null) => void onClose: () => void - /** 有批次写进去了：重取清单、刷新正在看的表 */ - onDrafted: (tables: string[]) => void }) { - // 表结构里已经没有的表起草不了（服务端会回「表结构里没有」），不算进范围 + // 表结构里已经没有的表起草不了（服务端会记「已跳过」），不算进范围 const live = useMemo(() => rows.filter((r) => r.in_schema).map((r) => r.table_name), [rows]) const liveSet = useMemo(() => new Set(live), [live]) const scopes = useMemo(() => { @@ -81,80 +70,156 @@ export function DraftDialog({ sourceId, rows, visible, selected, filtered, onClo if (sel.length) out.push({ key: 'selected', label: CT.scopeSelected(sel.length), tables: sel }) const vis = visible.filter((r) => r.in_schema).map((r) => r.table_name) if (filtered && vis.length && vis.length !== live.length) out.push({ key: 'filtered', label: CT.scopeFiltered(vis.length), tables: vis }) - if (live.length > TOP_N) out.push({ key: 'top', label: CT.scopeTop(TOP_N), tables: live.slice(0, TOP_N) }) + if (live.length > DRAFT_TOP_N) out.push({ key: 'top', label: CT.scopeTop(DRAFT_TOP_N), tables: live.slice(0, DRAFT_TOP_N) }) out.push({ key: 'all', label: CT.scopeAll(live.length), tables: live }) return out }, [selected, visible, filtered, live, liveSet]) const [scope, setScope] = useState(() => scopes[0].key) - const [useModel, setUseModel] = useState(false) - const [phase, setPhase] = useState('setup') - const [out, setOut] = useState(() => fresh(0)) - const [stopping, setStopping] = useState(false) - const stopRef = useRef(false) const tables = scopes.find((s) => s.key === scope)?.tables ?? scopes[0].tables + const tooMany = tables.length > DRAFT_MAX_TABLES + const [useModel, setUseModel] = useState(false) + const [guidanceText, setGuidanceText] = useState(guidance) + const [note, setNote] = useState('') + const guidanceLen = charCount(guidanceText) + const noteLen = charCount(note) + const guidanceOver = guidanceLen > DRAFT_GUIDANCE_MAX + const noteOver = noteLen > DRAFT_NOTE_MAX + const guidanceChanged = guidanceText.trim() !== guidance.trim() - const run = async () => { - stopRef.current = false - setStopping(false) - setPhase('running') - const acc = fresh(tables.length, useModel) - setOut({ ...acc }) - const touched: string[] = [] - let withModel = useModel - let i = 0 - while (i < tables.length) { - if (stopRef.current) { acc.stopped = true; break } - const batch = tables.slice(i, i + (withModel ? BATCH_MODEL : BATCH_PLAIN)) - try { - const r = await api.dataCatalog.draft(sourceId, { tables: batch, use_model: withModel }) - acc.added += r.total.added - acc.updated += r.total.updated - acc.removed += r.total.removed - acc.tableErrors.push(...r.tables.filter((t) => t.error)) - acc.modelErrors.push(...r.tables.filter((t) => !t.error && t.model_error)) - acc.modelItems += r.tables.filter((t) => !t.error && (t.model_items ?? 0) > 0).length - if (r.tables.some((t) => t.model_items !== undefined)) acc.modelItemsKnown = true - if (r.model) acc.model = r.model - // 模型整体用不了（没配置、已停用……）：后面的批次不再请模型，免得每批都撞一次同样的错 - if (withModel && r.model_error && !r.model_used) { acc.modelSkipped = r.model_error; withModel = false } - touched.push(...r.tables.filter((t) => !t.error).map((t) => t.table_name)) - } catch (e) { - acc.failure = errorMessage(e) - break + const [mode, setMode] = useState<'setup' | 'job'>(initial === 'job' && job ? 'job' : 'setup') + const [starting, setStarting] = useState(false) + const [startError, setStartError] = useState(null) + /** 是 409（已有任务）转过来显示的那一个 */ + const [busy, setBusy] = useState(initialBusy) + const [acting, setActing] = useState(false) + const [actionError, setActionError] = useState(null) + /** 上一次起草（没写全的才提），设置阶段写一行、可以点开 */ + const [last, setLast] = useState(null) + /** 起草模型的说法：设置里的起草模型，没配就是跟随助手的模型 */ + const [modelLine, setModelLine] = useState<{ text: string; warn: string | null } | null>(null) + const onTrackRef = useRef(onTrack) + onTrackRef.current = onTrack + + const inSetup = mode === 'setup' || !job + // 设置阶段：取上一次起草。别处（另一个标签页）建的任务还在进行：直接显示它 + useEffect(() => { + if (!inSetup) return + let alive = true + api.dataCatalog.latestDraftJob(sourceId).then(({ job: latest }) => { + if (!alive || !latest) return + if (isJobActive(latest)) { + onTrackRef.current(latest) + setBusy(true) + setMode('job') + } else if (latest.status !== 'succeeded') { + setLast(latest) } - i += batch.length - acc.done = i - setOut({ ...acc }) + }, () => {}) + return () => { alive = false } + }, [inSetup, sourceId]) + + // 起草模型：设置里的 catalog_draft 填了就是它；没填跟随助手的模型（写出当前是哪个，用不了时写原因） + useEffect(() => { + let alive = true + void Promise.allSettled([api.settings.get(), api.copilot.getModel()]).then(([s, c]) => { + if (!alive) return + const cfg = s.status === 'fulfilled' ? s.value?.catalog_draft : null + const provider = typeof cfg?.provider === 'string' && cfg.provider.trim() ? cfg.provider.trim() : null + const model = typeof cfg?.model === 'string' && cfg.model.trim() ? cfg.model.trim() : null + if (provider || model) { + setModelLine({ text: DT.modelNamed(model, provider), warn: null }) + return + } + const assistant = c.status === 'fulfilled' ? c.value : null + setModelLine({ text: DT.modelFollow(assistant?.effective_model ?? null), warn: assistant?.error ?? null }) + }) + return () => { alive = false } + }, []) + + const begin = async () => { + if (starting || !tables.length || tooMany || (useModel && (guidanceOver || noteOver))) return + setStarting(true) + setStartError(null) + try { + // 起草要求改过：先存（存不上就不开始，免得模型拿着旧的那份起草） + if (useModel && guidanceChanged) { + try { + const saved = await api.dataCatalog.putGuidance(sourceId, guidanceText.trim()) + onGuidanceSaved(saved.guidance) + } catch (e) { + setStartError(DT.saveGuidanceFailed(errorMessage(e))) + return + } + } + const extra = useModel && note.trim() ? { note: note.trim() } : {} + const r = await onStart({ tables, use_model: useModel, ...extra }) + setBusy(r.busy) + setMode('job') + } catch (e) { + // 起草模型用不了（没有接入、已停用、只有演示接入）：服务端那句话已经写了原因和去处 + setStartError(e instanceof ApiError && e.code === DRAFT_MODEL_UNAVAILABLE ? e.message : DT.startFailed(errorMessage(e))) + } finally { + setStarting(false) } - setOut({ ...acc }) - setPhase('done') - if (touched.length) onDrafted(touched) } - const stop = () => { - stopRef.current = true - setStopping(true) + const act = async (fn: () => Promise) => { + setActing(true) + setActionError(null) + try { + await fn() + } catch (e) { + setActionError(errorMessage(e)) + } finally { + setActing(false) + } + } + const stop = () => act(onCancel) + const retry = (from: CatalogDraftJob) => act(async () => { + const r = await onRetry(from) + setBusy(r.busy) + }) + const openLast = (j: CatalogDraftJob) => { + onTrack(j) + setBusy(false) + setMode('job') } - // 进行中关窗 = 停止：当前这一批做完再关，结果照常显示 - const requestClose = () => (phase === 'running' ? stop() : onClose()) + const phase: 'setup' | 'running' | 'done' = inSetup ? 'setup' : isJobActive(job) ? 'running' : 'done' + const canStart = !!tables.length && !tooMany && !(useModel && (guidanceOver || noteOver)) && !starting + const retryable = job && phase === 'done' ? retryableCount(job) : 0 const footer = phase === 'setup' ? ( <> - ) : phase === 'running' - ? - : + ? ( + <> + + + + ) + : ( + <> + {retryable > 0 && job && ( + + )} + + + ) return ( - -
+ +
{phase === 'setup' && ( <>
@@ -168,71 +233,192 @@ export function DraftDialog({ sourceId, rows, visible, selected, filtered, onClo ))}
+ {tooMany &&

{DT.tooMany(DRAFT_MAX_TABLES)}

}

{CT.draftBase}

- +
+ + {useModel && ( +
+
+ {DT.model}： + {modelLine ? modelLine.text : '…'} + {DT.modelWhere} + {modelLine?.warn && {modelLine.warn}} +
+
+ + {(p) => ( +