The issue is with Deepgram keyterms there’s a hard limit (about 500 tokens total, or 80–100 short terms). If you pass a huge list, the stream breaks after the first turn because the restart fails. Fix is simple trim/dedupe your keyterms, keep them short, or chunk them by topic. For niche phrases, put them in the LLM prompt instead of the keyterms list. Since I cut mine down under that limit, the assistant works fine again.