Context Window Capacity and Use Record

Context Window Capacity and Use Record

Definition

A context window is the bounded input and generated-output space available to a model for one interaction or processing sequence, subject to the provider's model-specific rules.

More capacity can allow more documents, code, conversation, audio, images, or video-derived tokens to be considered without first reducing everything to a short prompt.

What capacity does not prove

Capacity does not establish that the model finds every relevant item, weighs evidence correctly, reasons across the full record, preserves chronology, follows every instruction, cites sources accurately, or produces a trustworthy decision.

Liu and colleagues' 2024 TACL paper found that tested models often performed better when relevant information appeared at the beginning or end of long context and worse when it appeared in the middle.

Google's current long-context documentation also states that performance can vary with context and that multiple-item retrieval does not match single-item retrieval accuracy. It notes cost and latency tradeoffs and recommends omitting unnecessary tokens.

Historical correction

The February 2024 Gemini 1.5 announcement described 128,000 tokens as the standard window and an experimental one-million-token window for a limited group. The E004 transcript says 1.5 million.

Current model limits must come from current model documentation and should not be projected backward.

Evaluation model

Test source coverage, retrieval, synthesis, contradiction handling, chronology, citation, abstention, cost, latency, privacy, security, and reviewer traceability on representative records.

Long context and retrieval systems are design choices, not opposing identities. The right architecture depends on the task, source change rate, access control, citation needs, cost, and evaluation evidence.