Hey folks, was wondering if anyone has experience ...
# agents
w
Hey folks, was wondering if anyone has experience with using Gemini
gemini-live-2.5-flash-preview-native-audio-09-2025
and issues with tool calls / interruptions? Seems like whenever a user interrupts an agent, it goes silent, then a follow up prompt makes the agent speak again but then it thinks a tools call was made and starts hallucinating. Works fine if there are no interrupts tho...
Have been tweaking the instruction prompt.. which i swear it may work one day, but breaks the next day! 🥲 If i switch out Gemin i for something like OpenAI then everything works flawlessly..
b
Are you using preemptive generation
w
🤔 i'm using Gemini's native live model (through Vertex AI), is this config still applicable? edit: seems like option also only available in Python.. we're using the JS SDK
Oh I see the config, but doesn't seem to help...
With the latest 1.0.18 update, it just stops working altogether
Copy code
ERROR (60645): Text message received from Realtime API with audio modality. This usually happens when text chat context is synced to the API. Try to add a TTS model as fallback or use text modality with TTS instead.
c
Hey @wonderful-plastic-41083. Did you find any solution? I'm also facing the same along with latency issues.
w
Yes, I've identified a bug in the livekit google js plugin, i'm still doing a bit of testing before I raise a PR.
n
hey @calm-book-72252 did you find a solution for the latency, i am using the same gemini model that @wonderful-plastic-41083 use and i have like a 3s > latency