What is a language model?
A language model produces text from patterns it learned during training and the information supplied with a request. It can explain, summarise and draft; it can also be wrong.
By Exemplary AI · Updated · Illustrated explainer
The key idea
The model writes the answer. The surrounding system decides what information it can use.
The idea, illustrated.
How a question becomes an answer
Figure 01The employee’s request.
“Late-shift taxi claims require manager approval.”
Builds the response from small pieces of text.
Check the answer against the supplied policy before relying on it.
- 01
Supply the context
The request contains the question, instructions and any relevant information, such as a company policy.
- 02
Generate a continuation
The model processes small pieces of text called tokens and produces a response a piece at a time.
- 03
Check the result
A convincing sentence still needs to agree with the source. Fluency and correctness are different things.
What does a model actually learn?
Training adjusts a model’s internal numerical settings so it can represent patterns in its training material. Those settings are called parameters. Many modern language models use a Transformer architecture, which lets information in one part of a sequence influence how another part is processed.
For a business reader, the useful distinction is between training and a request. Training creates the model. A request gives an already trained model something to work on. Adding a policy to a question does not, by itself, retrain that model or make the policy available to every future conversation.
Primary source: Vaswani et al.: Attention Is All You Need
How does it know about your company?
A general model does not automatically know your latest contract, price list or approval rules. An application can retrieve the relevant information and include it with the request. The amount the model can work with at once is its context window. More context is not always better: unrelated pages can distract from the clause that matters.
This is why the surrounding application matters. It connects the right documents, checks access and makes the source available to the reader. Changing the model alone does not repair a missing policy or an incorrect permission.
Primary source: Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
An everyday example
A policy question with a visible basis
An employee asks whether a taxi home after a late shift is covered. The application supplies the current travel policy. The model can explain the relevant conditions in everyday language.
If the policy does not cover the situation, a useful response says what is missing and points the employee to the policy owner. It should not invent a reimbursement rule.
A common misconception
“If it sounds certain, it must know.”
The wording can be confident even when the evidence is weak. Source checks, current information and a clear “I don’t have enough information” path matter.
How Exemplary helps
We connect the model to the work.
Exemplary’s Model Hub brings model choices and deployments into one place. We help configure the model and connect it to the information the task needs, so your team can focus on the answer.
Explore Model HubSources & further reading
Primary publications and documentation. Accessed 22 September 2026.
- 01Attention Is All You Need
Vaswani et al. · NeurIPS · 2017
- 02Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis et al. · NeurIPS · 2020