Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion CLAUDE.md
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,7 @@ Load unpacked at `chrome://extensions/` → select `BillNote_extension/extension
- `app/routers/` — API routes: `note.py` (generation), `note_group.py` (note groups), `provider.py`, `model.py`, `config.py`
- `app/services/` — Business logic:
- `note.py` — `NoteGenerator` orchestrates the full pipeline (download → transcribe → LLM → notes)
- `task_serial_executor.py` — task queue
- 并发调度没有独立队列模块:`routers/note.py` 的 `_ParallelGate` 是唯一并发上限(设置里的并行数,热更新),任务体直接执行;批量入口用固定 8 worker 的分发池接住请求
- `cookie_manager.py` — per-platform cookie storage; injected into yt-dlp by downloaders (e.g. Bilibili)
- `transcriber_config_manager.py` — persisted transcriber settings
- `app/downloaders/` — Platform adapters with a shared `base.py` interface; only the Bilibili adapter is registered for generation (`SUPPORT_PLATFORM_MAP` in `app/services/constant.py`), while source intake (`app/services/source_intake.py`) stays platform-neutral so more adapters can be added without touching the intake contract.
Expand Down
24 changes: 23 additions & 1 deletion backend/app/db/init_db.py
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
import logging
import os

from sqlalchemy import inspect, text

Expand Down Expand Up @@ -32,6 +33,26 @@ def init_db():
return
engine = get_engine()

# spec 022 D4:启动时把统一数据根打进日志,并提示旧版 data/data 残留。
# 不自动迁移历史文件——迁移会改写用户数据,必须由人决定。
from app.utils.path_helper import get_data_dir

data_root = get_data_dir()
logger.info("媒体数据根目录(DATA_DIR):%s", data_root)
legacy_root = os.path.join(data_root, "data")
try:
legacy_entries = os.listdir(legacy_root) if os.path.isdir(legacy_root) else []
except OSError:
legacy_entries = []
if legacy_entries:
logger.warning(
"检测到旧版数据目录 %s 非空(%d 项);spec 022 不自动迁移,"
"需要时请手动把文件移入 %s",
legacy_root,
len(legacy_entries),
data_root,
)

Base.metadata.create_all(bind=engine)
if engine.dialect.name == "sqlite":
# Idempotent additive migration for older SQLite installations.
Expand Down Expand Up @@ -207,4 +228,5 @@ def init_db():
connection.execute(text("DROP TABLE note_groups"))
NoteGroup.__table__.create(bind=connection)
engine_module._INIT_DB_DONE = True
backfill_legacy_resources()
inserted = backfill_legacy_resources()
logger.info("媒体资源历史回填完成:新增 %d 条", inserted)
5 changes: 3 additions & 2 deletions backend/app/downloaders/base.py
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@
from app.enmus.note_enums import DownloadQuality
from app.models.notes_model import AudioDownloadResult
from app.models.transcriber_model import TranscriptResult
from os import getenv
from app.utils.path_helper import get_data_dir

QUALITY_MAP = {"fast": "32", "medium": "64", "slow": "128"}

Expand Down Expand Up @@ -66,7 +66,8 @@ class Downloader(ABC):
def __init__(self):
# TODO 需要修改为可配置
self.quality = QUALITY_MAP["fast"]
self.cache_data = getenv("DATA_DIR") or "data"
# spec 022 D4:与下载落盘、媒体登记共用同一个数据根。
self.cache_data = get_data_dir()

@abstractmethod
def download(
Expand Down
26 changes: 18 additions & 8 deletions backend/app/gpt/universal_gpt.py
Original file line number Diff line number Diff line change
Expand Up @@ -242,12 +242,14 @@ def judge_note_frames(
- `plan_frame_timestamps` 决定「抽哪些帧」,服务于 hybrid 采样的配额;
- 本方法决定「抽出来的帧里哪些真的进笔记」,是插图逻辑的唯一准入闸门。

解析不出结构化结果时抛 WorthJudgementError,**绝不**返回空清单——否则一次
网络抖动就会让整篇笔记静默地一张图都没有(spec 008 §5.6b)。模型明确判定
「都不值得」时返回空 approved,那是合法结果。
网络/鉴权失败向上抛。spec 022 D6 起,解析不出结构化结果**不再**抛到任务级:
任务级重试会重做下载/转写/总结,代价极高而结果大概率一样。显式空清单记
degraded=empty_frames,全部批次解析失败记 degraded=parse_failed,两者都
返回「全部候选」(受 note_max_images 截断),由上层写进任务 detail。
"""
from app.services.ocr_assist import (
WorthJudgementError,
_degraded_outcome,
build_judge_system_prompt,
parse_worth_judgement,
)
Expand Down Expand Up @@ -282,7 +284,11 @@ def _build_messages(batch: list) -> list:
if not image_urls:
response = self._chat_completion_create(_build_messages([]))
text = response.choices[0].message.content or ""
return parse_worth_judgement(text, candidates, max_images=limit)
try:
return parse_worth_judgement(text, candidates, max_images=limit)
except WorthJudgementError as exc:
logger.warning("判定层单次调用解析失败,降级为保留全部候选:%s", exc)
return _degraded_outcome(candidates, "parse_failed", limit)

# base 必须带一张图:_build_messages([]) 在无图时把 content 装配成字符串,
# 往里追加图片会因为没有 list 可 append 而静默丢图(体积估算恒为空)。
Expand All @@ -305,11 +311,15 @@ def _build_messages(batch: list) -> list:
)

if not items:
# 每一块都解析不出条目时按「判定失败」处理:静默返回空清单会让整篇
# 笔记一张图都没有,而且完全看不出出过错(spec 008 §5.6b)。
raise WorthJudgementError(
"判定结果里找不到帧列表(既不是数组,也没有 frames 字段)"
# spec 022 D6:全部批次都解析不出条目时降级,而不是让整任务失败。
# 保留全部候选至少能保证笔记有插图,degraded 让这件事在任务 detail
# 里可见——比「静默无图」和「白白重跑三次」都好。
logger.warning(
"判定层全部批次解析失败(图片=%d 批次=%d),降级为保留全部候选",
len(image_urls),
len(batches),
)
return _degraded_outcome(candidates, "parse_failed", limit)

# 在**完整候选清单**上只解析一次:这样 max_images 上限、去重、排序和
# 统计字段都是全局口径,而不是「每块各算一遍再拼起来」。
Expand Down
Loading
Loading