alldayrec docs

English

简体中文

English

简体中文

Appearance

How to troubleshoot common problems ​

Start with the log: with Docker, run docker compose logs -f. The web page 处理进度 (Progress, /progress) also shows how far transcription and summaries have got, and the system log.

The server exits right after it starts ​

  • The log says not set: ...: these variables are missing from .env. See Configuration.
  • The log says ENGINE, DRAFT_ENGINE, ENGINE_FALLBACK must be one of ...: one of these three has a wrong value, and the log lists the allowed values. If your value is in Change the ASR engine but not in the log, the image on your machine is older than the docs: run docker compose pull first.
  • The log says ffmpeg is not on PATH: install ffmpeg (macOS: brew install ffmpeg; Debian, Ubuntu: sudo apt install ffmpeg). The Docker image has it.
  • The log says ENGINE=local needs sherpa-onnx: install it with the local extra: pip install 'alldayrec[local]', uvx with --from 'alldayrec[local]', from source uv sync --extra local. The Docker image has it.
  • The log says others can read it, and it holds every token: chmod 600: run chmod 600 on the tenants.json in DATA_DIR (with Docker, chmod 600 data/tenants.json).
  • The log says a model file failed to download, is missing or has another sha256: delete that file and restart, or download it from the same address the log gives and put it in MODELS_DIR (data/models with Docker). A ReDimNet2 file you export yourself with export_onnx.py differs byte for byte from the released one, and the server refuses it.

The web page shows only token required ​

This browser isn't signed in yet. Open http(s)://<your-server>/?token=<your token> once; after that a cookie keeps you signed in.

Summaries fail, or "ask AI" reports "模型调用失败" (model call failed) ​

  • When a Bailian key is restricted to specific models in the console, calls to the chat models return 403. Use a key without a model restriction.
  • Another LLM service must support tool calls and JSON output (response_format of json_object).

Some conversations have no summary ​

A conversation with under 60 seconds of speech doesn't go to the LLM: its title is its first sentence, and it has no summary or to-dos (MIN_SPEECH_S in server/alldayrec/summary.py).

The first transcript is slow with ENGINE=local ​

When the first audio arrives, the server first downloads a model tarball of about 136 MB, so the first transcript takes about a minute. It doesn't download it again after that.

The last part of imported recordings is never transcribed ​

After you upload a batch of files, call POST /api/chunks/end once. See Import recordings you already have. Until then, the batch's last conversation may also still show as 「进行中」 (in progress).

The phone can't connect after scanning the QR code ​

  • The server address in the QR code is the one picked under 「手机用哪个地址」 (Address for the phone) on the 连接手机 page. If it is 127.0.0.1 or another address the phone can't reach, pick one the phone can reach, or type it under 「自己填」 (Enter it yourself), then scan again.
  • With an http:// address, APKs 1.5.0 and older can't connect: upgrade to 1.5.1 or newer. See Reach the server from a phone.

The app on the iPhone no longer opens ​

An app installed with a free Apple ID stops working after 7 days. Install it again from Xcode. A paid developer account doesn't have this limit.

Importing a file returns 413 behind an nginx reverse proxy ​

By default nginx accepts request bodies of up to 1 MB only. Raise client_max_body_size.

Can older Safari play the recordings? ​

Yes. The server decodes the audio for playback and live listening on the web page to WAV before the browser gets it (/api/audio, /api/listen), because Safari before 18.4 can't decode the recordings' Ogg Opus format.

The card's OMI drive never appears on macOS ​

The app's 录音设备 › U 盘模式 › 存储卡当 U 盘 (Recorder › USB drive mode › Card as USB drive) gives the recorder's card to the computer as a USB drive. Normally the OMI drive appears after about 20 seconds.

What the maintainer measured on macOS 26: the system allows about 20 seconds to mount a FAT32 volume, so a large FAT32 card that reads slowly (such as 32 GB) may never mount; the same card formatted as exFAT mounts in about 2 seconds. The firmware reads both FAT32 and exFAT. Also, macOS mounts no disks while the Mac is locked.