AI and agents
Context window
A context window is the maximum amount of text, counted in tokens, that a language model can take in and produce in a single request.
How it works in practice
Everything the model works with in one call has to fit: the instructions, the conversation so far, any retrieved documents, tool results and the answer it writes. Current large models accept hundreds of thousands of tokens, which is enough for long contracts, but more context costs more per call and can still make the model miss details in the middle.
In practice we send what the task needs and no more. Retrieval picks the relevant passages, long threads are summarised, and tool results are trimmed to the fields that matter. That keeps cost and response time down and usually improves accuracy.
A worked example
A legal team asks an assistant about a 300-page lease bundle. Rather than sending the whole bundle on every question, the system retrieves the ten clauses that match and sends those, so each answer costs a fraction as much and cites the clause it used.
Examples in this glossary describe typical builds for companies of 1 to 1,000 people. They are not named client results.















