Embedding

An embedding model converts text into a fixed-length sequence of numbers. The values mean nothing on their own; their value lies in the fact that similar meanings land close together in that space. "Annual leave request" and "how do I apply for holiday" share no words, yet sit as neighbours in the vector space.

This solves what keyword search cannot. Classic search looks for shared words; embedding-based search looks for shared meaning. In corporate documents the word a user asks with rarely matches the word in the document, which makes the difference decisive.

KobiGPT uses a multilingual embedding model. The practical consequence is that a question asked in Turkish can locate the relevant passage in an English-language document. In SME archives where Turkish and English content sit side by side, that is a common requirement.

The model choice is made once and binds the whole system. Embedding dimensionality is a fixed invariant on both the API and the embedding worker; changing the model requires re-indexing the entire archive. That makes the embedding model one of the hardest decisions to reverse in the architecture.

Embedding happens once per document and is metered at 0.1 Kobi in KobiGPT. At query time only the question needs embedding; the archive is not reprocessed. That is the source of RAG's cost advantage.

On scale, the limit is the plan tier: the Pro plan covers indexing up to 12500 documents. Because embedding jobs run asynchronously through a queue, uploading hundreds of documents at once does not stall the system — it queues them.