Why Generic AI Translation Can't Resume After Failure
Every translation of a long document will hit a transient failure at some point. A provider rate limit, a network blip, a model that times out on one specific chunk. The question is what happens next. In chat-based AI translation, what happens next is usually that you start over.
What goes wrong
The session ends, the tab closes, the browser crashes, the laptop reboots. The translation is gone. You have either to paste the source back in and try again, or to manually figure out which chunks completed and which did not, and stitch the result together by hand.
Why generic AI translation fails here
State lives in the tab. The conversation is the project. Close the conversation and the project is gone. Even when the chat persists, the partial translation is not addressable as a unit; you cannot point to "chunk forty-three" and say "redo this one and only this one".
How TranslationAI solves it
Project state lives in the database, chunk by chunk. The browser tab is a view, not the work. A failure marks the project failed and stops the worker; a manual retry resets the failure record and resumes from the last good chunk. The project record persists until you explicitly act on it, even when the work failed mid-way. Refreshing the page is safe. Closing the tab is safe. Coming back tomorrow is safe.
Further reading: why failures must be surfaced honestly first, why long documents need server-side orchestration, how context survives the resume.