Context window
The maximum amount of text a model can hold at once, counted in tokens, covering the prompt, any retrieved sources, the conversation so far and the answer being written.
In depth
What it really means
The context window is a hard ceiling, not a preference. Everything the model is working with has to fit inside it: your page, the other nine pages it retrieved, the user’s question, the system instructions and the response. Your content is competing for space against all of that.
Two practical consequences. First, only a slice of your page usually makes it in, which is why chunking decides what gets used. Second, models attend unevenly across a long window, with material at the very start and very end weighted more heavily than the middle. A fact buried in the middle of a long retrieved document is measurably less likely to be used.
How it works
- The prompt, retrieved chunks and conversation history are converted to tokens.
- Tokens are loaded into the window up to the model’s limit.
- Anything over the limit is truncated or dropped, usually oldest first.
- The model attends across everything loaded, with position affecting weight.
- The generated answer also consumes window space as it is written.
Available context = window size – system prompt – conversation history – reserved output tokens
A rough English conversion: 1 token is about 4 characters, or about 0.75 words. A 1,000-word page is roughly 1,300 tokens. Frontier models in 2026 sit in the 200,000 to 1,000,000 token range, but the practical retrieval budget per source is far smaller because ten sources share it.
Pros & cons
Pros
- Larger windows let assistants read more of your page rather than a fragment.
- More room for retrieved sources means more chances to be one of them.
- Longer windows support follow-up questions without losing the earlier context.
Cons
- Attention degrades across long contexts, so more room does not mean equal use.
- Larger windows cost more and run slower, so production systems retrieve conservatively.
- A bigger window changes nothing if your key fact is buried mid-document.
Common mistakes
- Assuming a million-token window means your whole page gets read. Retrieval still sends fragments.
- Confusing the context window with the knowledge cutoff. One is per-query capacity, the other is training recency.
- Burying the definitive answer 2,000 words into a page.
- Writing 5,000-word pages on the theory that longer is more authoritative, which dilutes every chunk.
- Ignoring that boilerplate navigation and cookie banners can eat the extracted budget on a badly built page.
Best practices
- Put the definitive answer near the top of the page and near the top of each section.
- Keep pages focused on one topic so retrieved chunks are dense with relevant material.
- Make sure the key fact appears in full inside a single section rather than split across three.
- Cut boilerplate that gets extracted alongside your content.
- Choose clarity over length. A tight 900-word page usually outperforms a padded 3,000-word one.
FAQs
What is a context window?
The maximum amount of text a model can process at once, counted in tokens. It covers the prompt, retrieved sources, conversation history and the answer being generated.
Why does the context window matter for content?
Your page shares that space with every other retrieved source. Only a slice of it gets in, and material at the top of that slice carries more weight, so front-loading the answer matters.
Do bigger context windows solve this?
Not really. Attention weakens across long contexts, and retrieval systems still send fragments rather than whole pages for cost and speed reasons. Structure still decides what gets used.
Is the context window the same as the knowledge cutoff?
No. The context window is how much a model can read for a single query. The knowledge cutoff is how recent its training data is. A model with a 2025 cutoff can still read a page published today if it retrieves it.
How many tokens is my page?
Roughly, word count divided by 0.75. A 1,000-word page is about 1,300 tokens. Most retrieval systems send far less than that per source.
Keep reading
Related on LymLyt
Beyond LymLyt
Further reading
Want this working on your site?
We build the content behind the term, ranked in search and cited by AI.
Book a 30-min call →