The News, Explained
On November 6, Kakao released Kanana-2-v-decide, an AI model that reads text and images and returns judgments in a defined format. Depending on the question, it gives yes-or-no probabilities, probabilities for several options, or a score based on a rating scale. An AI agent, software that carries out a task in steps, or an ordinary program can read those values to choose what to do next. Source Source
Suppose a customer sends a photo of a part and says it won’t fit. The service could give the model three options: compatibility question, installation help, or other. The model can return a probability for each after reading the message and photo. Where the ticket goes, whether to ask for another photo, and what the customer hears back are decisions for the service’s rules or its next step. Source Source
Kakao says the model turns questions and options into sets of numbers and compares how close those sets are. If the same options recur, the model can reuse the numbers it has already calculated. Kakao’s engineering team reports a response time of about 42 milliseconds on one H200 processor. That figure alone doesn’t say how long it takes to upload a photo and route a support ticket. The model card lists a maximum of 10 images and 20,480 tokens, the units a model splits its input into. Video is unsupported. Source Source
The model files are available on Hugging Face. The Kanana Open License allows developers to build and run their own services under its terms without a separate commercial license. Activities covered by section 4.1, such as giving third parties API access to the model or supplying it inside a customer’s system or device, require a separate commercial license from Kakao. Source Source Source
OYOPICK’s Take
For a support team, the consequence of a wrong classification matters more than which option got the highest probability. If a ticket is easy to move after it lands in the wrong queue, routing could be a reasonable place to start. If the decision is to deny a return, a rule that sends dark photos or cases with two close scores to a person matters more. The case for expanding automation should come from fewer misrouted tickets and fewer cases staff need to handle again in real work. The vendor’s response time can’t answer those questions. Source Source Source