FinishReason(10) | Dagger | Page 1

brittle nimbus May 5, 2025, 9:56 PM

#

this one might go away if we switch fully to static tools, going on my theory that it's when it tries to call a tool before selecting it. instead the model would call the tool-calling tool.

(the error might come back if the LLM hallucinates a direct tool call instead of using the tool-calling tool, though 🫠)

jolly sinew May 5, 2025, 9:57 PM

#

Makes sense! This started happening a lot more with the new tool calling in 0.18.4 right?

brittle nimbus May 5, 2025, 9:59 PM

#

i think so - it happened before too, but seems to happen more often now. it's a bit mercurial - it would go away and come back with seemingly unrelated prompt changes, and some evals hit it more than others. it's possible we can system prompt our way out of it, but to be really sure we might need another eval that trips on it more predictably

jolly sinew May 5, 2025, 10:00 PM

#

awesome yeah it's pretty random right now too

jolly sinew May 6, 2025, 3:34 PM

#

https://developers.googleblog.com/en/gemini-2-5-pro-io-improved-coding-performance/

this new version will not only improve coding performance but will also address key developer feedback including reducing errors in function calling and improving function calling trigger rates

time to try it again maybe!

Gemini 2.5 Pro Preview: even better coding performance- Google Deve...

#

though I was having the issues with 2.0-flash so maybe not

brittle nimbus May 6, 2025, 3:37 PM

#

I have been wondering if this improves with 2.5+ yeah

jolly sinew May 6, 2025, 3:41 PM

#

I'll give it a shot. 2.5-pro is still a lot more expensive with lower rate limits compared to 2.5-flash

brittle nimbus May 6, 2025, 3:48 PM

#

ah have you tried 2.5-flash?

jolly sinew May 6, 2025, 4:00 PM

#

I'm using it now, hadn't tried it before now. Seems pretty strong?

#FinishReason(10)