Response with endless spaces and Chinese characters

I’m using an assistant via API. Before I used it in the Playground.
In the playground I never had this issue:

  • The assistant has some instructions to extract information from a text and tell me the position of the found information. It shall return two markdown tables.

  • Via a user message I add the text and a list with each word and its position

  • In the playgoud I got 2 nice markdown tables

  • Via API I get also two markdown tables, but in between there are 40k-70k spaces and some of those ‘습습습습습습습습습습습습습습습습습습습습습’

My config:

  • model: gpt-4-turbo-preview
  • temperature: 0

It doesn’t happen for all my texts, but for one it seems to be reproducible.

Any ideas what could cause the issue?


We aren’t the only ones! We are seeing this 5-10 times a day on gpt-4-turbo-preview

we’ll get the markdown table header, then instead of row content, we get 10-70kB of whitespace, and sometimes a repeating chinese character (in our case, it is the same character you posted).

There’s other posts of similar problems others are seeing, we haven’t found a solution yet, beyond some code to detect that and rerun the prompt (which usually comes back correct, but sometimes it does the whitespace thing again)