Hacker News Reader

Not sure where else to post this, but when attempting to use any of the Gemini 2.5 models via API, I receive an "empty content" response about 50% of the time. To be clear, the API responds successfully, but the `content` returned by the LLM is just an empty string.

Has anyone here had any luck working around this problem?

View on HN

What finish reason are you getting? Perhaps your code sets a low max_tokens, so the generation stops while the model is still thinking, without giving any actual output.

The finish reason is `length`. I have tried setting minimal token budgets, really small prompts, and max lengths of various sizes from 100-4000 and nothing seems to make a consistent dent in the behavioral pattern.

This can happen if the prompt or response is blocked by a safety filter. Check some of the other fields in the response.

Reply on news.ycombinator.com