Context window

noun

A context window is the portion of prior text or conversation an AI language model can reference to generate its next output. It acts as the model’s short-term memory, determining how much prior information it can use to maintain coherence, answer follow-up questions, or follow complex instructions.

Most consumer AI models, like those in chatbots or image generators, have a fixed context window (e.g., 4,096 tokens, roughly 3,000 words). This means if a conversation exceeds that limit, the model may forget earlier details, leading to errors or irrelevant responses. Some advanced models dynamically adjust the window, but this increases computational cost.

For users, this means keeping interactions concise or summarizing key points to stay within the window. For developers, it requires optimizing prompts and managing memory usage. For example, a user asking an AI to summarize a 10,000-word document would need to split it into chunks, as the model might lose track of earlier sections if the full text exceeds its window.

If you ask an AI to analyze a 5,000-word report, it may only process the last 3,000 words if its context window is 4,096 tokens.

See also: Token · Attention mechanism

Découvrez nos autres services

ElpisIA, c’est tout un univers — explorez nos autres services.