Sure? It answers. Unsure? It asks.
decidekit answers what it's sure of in milliseconds, on your device. When it isn't sure, it asks a slow LLM and learns from the answer: every eight answers it retrains. It starts out knowing only the names of twelve card-banking intents.
answered on device
answered on device, disagrees with BANKING77
sent to the LLM
Messages: BANKING77 test set (PolyAI, CC BY 4.0). The LLM is Llama 3.1 8B on Cloudflare Workers AI, given the same one-sentence label descriptions Decide uses; its answers to these messages were recorded once, so the replay is the same for everyone and costs nothing. It agrees with BANKING77's labels on about 89% of them, and Decide learns its mistakes too. How to do this in your app