Fenêtres contextuelles

Text models process tokenized input, where a token may represent a word, subword, character, or byte-based unit depending on the tokenizer. A fixed token-to-word conversion should not be assumed across languages and models. Context capacity describes what a configured model can consider in one request; it is not the size of its training dataset or a permanent memory guarantee. In a decoder-style example with one shared sequence budget, supplied instructions, conversation turns, document tokens, and generated continuation all need to fit. A model or interface can also impose a separate maximum on new output tokens. Distinguish these limits rather than assume that the entire remaining context can always be generated as an answer. For an illustrative shared budget of 8,000 tokens and input using 6,500 tokens, 1,500 remain before considering any other limit or additional overhead. Oversized inputs do not have one universal behavior. Depending on the interface and settings, a request can be rejected, truncated, or shortened by an application before submission. The model does not automatically retain excluded text. Longer accepted input also does not guarantee stronger retrieval or reasoning. Lost in the Middle research reports position-sensitive use of information in its evaluated tasks, including weaker performance when relevant material appeared in the middle. These findings motivate testing the actual model and task rather than assuming identical behavior for every long-context system. Retrieval can select passages for the input, but the complete retrieval-and-generation workflow still needs its own quality and resource evaluation.