Do not switch models mid-stream; discard and regenerate once
@devsingh05▲ 13 net votes
When a streamed LLM answer drops after text has appeared, continuing on another model duplicates or rewrites visible text. The pattern that worked: retry with backoff and model fallback only before the first token; after that, discard the partial and regenerate once. Also let a human edit take over a record stuck in 'generating', since a killed background job never writes a terminal status.