Large language model (LLM)

A language model probabilistically continues the text it is given. It carries broad world knowledge from its training data and produces fluent prose. But that knowledge is frozen at a cutoff date and stored inside the model's weights — its source cannot be shown.

In enterprise use, those two properties fall short. Your company's leave procedure is in no model's training data, and if you cannot show where an answer came from you cannot rely on it in an audit. A language model alone is therefore not an enterprise knowledge solution.

KobiGPT uses a chat model from the Gemini family by default. The model's job is not to carry knowledge but to read the document passages placed in front of it and summarise them in a way that answers the question. The knowledge comes from the archive, not the model's memory.

That design makes the model a replaceable component. It is set by configuration, and changing it does not require re-indexing the archive. The embedding model is the opposite case — there, the choice is a system-wide invariant.

Language model usage is billed per token, and KobiGPT converts that into Kobi Kredi. An AI message is metered at a minimum of 0.1 Kobi; actual consumption is computed from the input and output token counts returned by the model, with output tokens weighted more heavily.

The monthly quota is defined per plan tier: 120 on Ücretsiz plan, 1000 on Starter, and 12500 Kobi on Pro. This is the mechanism that keeps external service cost predictable.