Hey, I’m trying to “multithread” images to a VLM. What i’d liketo do: grab a VideoFrame triggered by user starting speech, and immediately send it to a VLM, with a dummy prompt if necessary. Then in on_user_turn_completed, I send the same exact image “again” with the actual user prompt appended this time, but now the VLM has cached the image. Is this achievable?

Yes, that’s achievable. You can pre-send the VideoFrame with a placeholder prompt to warm the VLM cache, then reuse the same image reference (or hash) in on_user_turn_completed with the real prompt so the model leverages the cached embedding instead of reprocessing the image. I’ve implemented similar vision+speech pipelines and can help design the caching flow and latency optimization end-to-end. @acceptable-leather-47181