Context window: complete definition in AI for SMEs

The context window is the maximum amount of text a language model can process in a single request — input (prompt + injected documents) and output (response) combined. It is measured in tokens: GPT-4o accepts 128,000 tokens, Claude 200,000, some open-source models 4,000 to 32,000. It is the model's "volatile memory": everything beyond the window is ignored.

What it changes for an SME

The context window determines what you can fit in a single request:

• a 500-word email (about 700 tokens) + 2,000 tokens of context → well within the window;
• a 20-page contract (about 8,000 tokens) + an instruction → possible on GPT-4o, tight on a light model;
• a 100-page document (about 40,000 tokens) → does not fit in a single call — RAG with chunking is needed.

Practical implications

The window drives architecture: for large volumes, RAG chunks documents to fit the window. For long conversations, the model's memory is limited to the window — old exchanges drop out if not re-injected. In fractional AI leadership, we size chunking and memory according to the chosen model.