So recently been seeing an issue where in the front end, text chunks miss out. The agent speaks out the entire llm response but the corresponding text chunks may cut off in between in the frontend.
Some notable events that happened before seeing this issue in the system is Avatar integration (Tavus), did an LLM update, updated Livekit version to 1.4.1 to support some extra plugins. The issue is completely random but comes up in the worst of times during demo calls. So wanted to do an RCA on it. I can reassure that LLM is completing its response. But the front end is not receiving the full text from backend during this scenario and I am unable to debug if its some wss issue under the hood.
If its a livekit version issue, would love to know the most stable version recommended by you guys for production usecases atm. Putting it out here to see if anyone else faced the same/any assumptions of why it is happening.