Bug: Audio/Video Files Over 25MB Not Rejected — Size Limit Not Enforced

[dev-bitrupt]

What happened?

30MB MP4 uploaded successfully with no error. Size limit not enforced.

Steps to reproduce

1. Open AI Chat attachment
2. Upload an MP4 file of 30MB

Expected behavior

Upload rejected with clear error: audio/video files are limited to 25MB.

Affected service

Platform

Version or commit

0.2.58

Browser and OS

Chrome

Relevant log output

Screenshots or recordings

No response

Before submitting

• I searched existing issues and this hasn't been reported yet.

[larryro]

@dev-bitrupt We don't impose a size limit on audio/video uploads ourselves. The 25MB cap is OpenAI's model-side limit, not a constraint we apply to user uploads — our backend automatically compresses and splits the file before sending it to the model.

[yannickmonney]

Triaged against current main.

The 25 MB figure is OpenAI's Whisper API limit, but Tale doesn't send raw uploads to that API. Audio/video attachments are re-encoded server-side before transcription — video is stripped (-vn), audio is silence-trimmed and re-encoded to 32 kbps Opus — so a 30 MB MP4 comfortably lands under the downstream cap. Because of that, the raw-upload size limit for audio/video is intentionally large (2 GB), and the real gate is duration, not bytes:

• file-types.ts documents this directly: "the raw upload can be far larger than OpenAI's 25 MB cap. Duration is the real gate — see CHAT_AUDIO_MAX_DURATION_SEC."
• Audio/video over the 4-hour limit is rejected, with a clear message (audioDurationExceeded).

So a 30 MB MP4 uploading successfully is expected behavior, not a missing size check. There is no UI or docs text that promises a 25 MB ceiling, so nothing is misleading the user toward that expectation either.

Enforcing a hard 25 MB cap would regress a deliberate feature (transcribing large media via re-encoding). Closing as working-as-intended. If a specific large file fails to transcribe (as opposed to upload), that's a transcription-pipeline concern tracked separately (e.g. #1457 / #1619).