2. Tokens and the Context Window

Every AI model reads and writes in small chunks called tokens. A token is roughly a short word or part of a word — about 4 characters in English, slightly fewer in French. The word "Understanding" is one token; a long word like "anticonstitutionnellement" is split into several. This matters because everything the model does is counted in tokens: the cost of a request, how fast the answer arrives, and whether your content fits within the model's limits. A longer document means more tokens, which means higher cost and slower responses.

The context window is the total amount of text a model can consider at once for a single request. Think of it as the model's working memory. It includes everything: your message, the conversation history so far, any documents you attached, and the behind-the-scenes instructions that shape the model's behavior. When the window fills up, the oldest parts are dropped or summarized to make room. Crucially, the model cannot remember anything that is no longer inside its window — unless a memory system puts it back in.

Because the context window has a fixed size, there is a direct trade-off between what you put in and what the model can produce. If you attach a very large document, you leave less room for the model's answer and for conversation history. This is why cost, speed, and quality are all connected to token count. A focused, well-scoped request uses fewer tokens and gets a faster, cheaper, often better response than dumping everything in at once.

Remember: the model does not learn from your conversations. It is fixed once trained. What makes it feel like it "knows you" inside OUPI is that your memory, skills, and documents are re-injected into the context window each time. So the context window is not just a technical limit — it is the only channel through which the model receives any information at all for a given request.