Large language models (Gemini, Claude, GPT and others) are the engine behind today’s AI assistants. Trained on large amounts of text, they understand freely worded questions in several languages and write clear answers. They are based on a type of neural network called a “transformer”, described in 2017 by Google researchers in the paper Attention Is All You Need.
What an LLM can do
- Understand a poorly worded question. Mistakes, abbreviations, incomplete sentences: it identifies the intent
- Write in the visitor’s language. It reads a document in French and answers in Dutch or English
- Summarise and rephrase. It condenses a conversation for the team or adapts text to the requested tone
- Follow instructions. Role, tone, prohibited topics: an LLM follows clear instructions well, although this is not an absolute guarantee
Its two limits
It only reliably knows what appeared in its training texts, not a company’s current prices or procedures: these are supplied with each question through RAG. It can also produce a false but convincing answer, a hallucination, when information is missing. These two limits explain why a business assistant is never just “an LLM connected to a website”.
How to choose a model
For a business assistant, the best model is not necessarily the most powerful. Useful criteria are answer quality in your languages and on your documents; cost per conversation; response speed; and data processing terms. The choice is based on tests with real questions and can change over time; the tests then need to be run again.
LLM, generative AI, assistant: the differences
Generative AI covers models that produce text, images or sound; an LLM is its text branch; an assistant is the product built around it, with documents, instructions, channels and the team that takes over.
During setup, KernChat chooses a model suited to each organization’s knowledge; conversations and documents are stored in the European Union, and messages are only sent to the model provider for the time needed to generate the answer (see how it works).