Tokens are pieces of input
A token may represent a word, part of a word, punctuation or another unit. Tokenization depends on the model’s tokenizer; the same text may have different counts across models. Character or word counts are only rough proxies. Use the appropriate tokenizer when exact limits or cost estimates matter. Non-text inputs may be represented differently.
The context window is a working limit
The context window limits the amount of information a model can process in a request. Instructions, conversation, retrieved passages and generated output may share that budget, depending on the API. A long chat interface may summarize or omit older messages. Even when text fits, important details may be missed. Provide the relevant excerpt and identify the question clearly instead of assuming unlimited memory.
Attention relates positions in a sequence
In a transformer, attention computes weighted combinations of representations. This lets a token’s representation use information from other positions allowed by the attention mask. In a causal language model, that typically means earlier positions rather than future tokens. Multiple layers and heads create different transformations. Attention weights are mathematical quantities, not proof of human-style understanding or a reliable explanation of why a specific answer was produced.
Make the context easier to use
Separate instructions from quoted material. Give documents stable source IDs and page labels. Prefer complete, relevant passages to isolated fragments. If an answer refers to a deadline, include the date and surrounding conditions. If sources disagree, keep both rather than quietly merging them. A useful context package preserves enough detail for the reader to check the eventual answer.
Try it yourself
Compare “Summarize this” with a prompt specifying audience, length, source IDs and missing-information behavior. Use the same short document for both.
What to look for
The second prompt gives clearer success criteria. It can make evaluation easier, but it does not guarantee factual accuracy or that every constraint will be followed.
Sources & further reading
Independent educational material. Examples are simplified and do not replace checking the documentation for the model or service you use.