What is a context window? Per-request capacity and the limits of recall
Understand context windows through multi-contract questions, distinguishing input length, output limits, and persistent memory, with budgeting and long-document checks.
Understand the concept
A context window limits the tokens a model can handle for a request. Instructions and conversation history can consume capacity, and models account for input, output, and reasoning tokens differently. A large window is neither permanent memory nor a guarantee of complete retrieval from long documents. Exceeding limits may cause errors or truncation.
Inspect the content actually included in the request
Comparing three contracts may involve document text and previous discussion. Visible chat history does not establish that all of it was sent; the system may summarize or remove older turns. Inspect actual inputs before attributing omissions to model behavior.
Capacity is measured in tokens, not directly in Chinese characters or pages. Multimodal inputs have their own accounting. Use the target model’s counting tool and reserve output capacity rather than relying on a universal conversion ratio.
Separate output limits from total context capacity
Some models share a total input-output budget while others also impose separate limits. A large context figure does not imply equally long output. Check the interface and version, including reasoning budgets and finish reasons where relevant.
Long contracts can leave insufficient capacity for a full comparison. Extract relevant clauses or process sections, retaining source identifiers to prevent mixing conditions across contracts. Reduce irrelevant content without discarding critical exceptions.
Test long-context behavior with multiple details
Test dispersed exceptions rather than only one easy sentence, checking each against the source. Google’s long-context guidance distinguishes multi-detail retrieval from single-needle lookup. Supported length alone is not an accuracy promise.
This is a checking method, not a contract upload or model test. Cross-session persistence needs storage and retrieval beyond the context window. Link summaries to originals so compressed details can be recovered from evidence.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.