Tavus shows Raven-2, faster perception with visual memory for PALs

Tavus released a video on Raven-2, the perception model that lets a PAL see faster and remember what it saw earlier in a conversation. The model has been in Tavus's developer changelog since October 23.

Tavus says Raven-2 processes camera and screen content about 3x faster than the model before it. It also keeps a visual history of the call, so a PAL can refer back to objects, pages and errors that are no longer on screen. The docs give questions like "What was the previous error?" and "Which object did I hold up first?" as examples. That history is built automatically and covers a single conversation.

Raven-2 also adds dedicated screen perception, which starts once someone shares their screen, and lets a PAL speak up about what it sees on camera while you're quiet.

It's opt-in. Raven-1 is still the default, and you switch a PAL over by setting its perception_model to raven-2 and starting a new conversation.