Slide context length and visualize the lost-in-the-middle effect. Everything outside the window is completely invisible to the model, and even content inside the window is not recalled equally well.
Sign in to save progress
The context window holds everything the model can read. Drag the slider to resize it.
Tokens available
4K
Approx. words
3,072
Approx. pages
~8
Even when content fits inside the window, recall quality drops for information buried in the middle of a long context.
Illustrative curve of the U-shaped "lost in the middle" pattern reported by Liu et al. (2023). It shows the qualitative shape, not measurements of any specific model.
Start (0–10%)
~86%
Best recalled
Middle (40–60%)
~57%
Often missed
End (90–100%)
~74%
Partially recalled
Context windows have grown dramatically across model generations. The table shows each model's documented limit at its release date; current models may differ. Page counts use the same conversion as Section 1 (0.75 words per token, 400 words per page).
Every message in the conversation accumulates tokens. Adjust the sliders to see how the window fills up.
Documents larger than the context window must be split into manageable pieces.
RAG (Retrieval-Augmented Generation) retrieves only the chunks relevant to the user's query, keeping the context window focused rather than stuffed.
Explore the RAG guide→Longer context does not equal better performance. Retrieval-augmented approaches often outperform "stuff everything in the context" because they focus attention on relevant content rather than burying the signal in noise.
Up next
Hallucinations: When Models Confabulate
Understand why LLMs produce confident but wrong answers.