Context Window
A context window is the amount of text, measured in tokens, that an AI language model can consider in one request. It limits how much of a construction spec book or proposal set the model can see at once.
Why it matters in construction
Construction documents are long, and that’s putting it nicely. A spec book for a mid-size commercial job can run 600 to 1,200 pages. A bid package with drawings, addenda, and sub proposals can be several times that. A CRM history for a repeat client stretches back years.
The context window limits how much material a model can receive for one question. That limit shows up when a tool finds a detail in Division 08 but misses a note in Division 26. It also determines whether a system can take a small bid package whole or needs to break a larger set into sections and retrieve the relevant pieces.
How it works
- Everything counts. The system prompt, the question, every attached document, the conversation so far, and the model’s reply all share one token budget.
- Hard cutoff. If the input exceeds the window, the request can fail or the tool can omit material to make room. That becomes dangerous when the tool does not clearly show what it left out.
- Attention is uneven. Even within the window, a model can have more trouble finding a note buried in a long spec book than answering from a short, relevant excerpt.
- Cost scales with length. Providers charge per token. A tool that sends 500 pages for every question adds cost with every click.
- Chunking and retrieval. Practical systems split documents into sections, store them, and retrieve only the relevant pieces for each question.
A larger context window does not make a full spec book the best input for every question. Fewer, more relevant tokens beat a stuffed prompt.
Example in practice
A precon team asks an AI assistant whether the owner’s contract for a $22M distribution center allows substitutions after bid. They upload the full 900-page project manual. The tool reports “no substitution restrictions found.”
An estimator who has read the document knows Section 01 25 00 prohibits post-bid substitutions except for discontinued products. The manual was about 650,000 tokens and the tool’s model handled 200,000. The tool sent what fit and dropped the rest without saying so. Section 01 25 00 was among the pages that never reached the model.
The team indexes the manual by section and retrieves the relevant pages for each question. For this query, the tool returns the substitution clause with a page reference and sends about 4,000 tokens instead of the full 650,000-token manual.
Frequently asked questions
How big is a typical context window?
Current frontier models can handle between 200,000 and 1,000,000 tokens, which is roughly 150,000 to 750,000 words. That can cover a large spec book, but not a whole project archive.
If the model can hold the whole spec, why do I need retrieval?
Sending 800 pages for every question is slow and expensive, and models can struggle to find one detail inside a very long prompt. Retrieval sends only the relevant pages.
Does the context window include the model's answer?
Yes. The prompt and the response share the same budget, so a long input leaves less room for a long output.
Go deeper
- From the blog GPT Models for Contractors: A Plain-English Guide A breakdown of GPT-5.6 for general contractors: what Sol, Terra, and Luna do, how effort levels work, and which model fits each precon task.
- From the blog AI for Estimators: 5 Practical Uses that Work Today (Not Someday) Construction's drowning in AI hype, but here are five use cases AI can actually do for estimators today, if it has access to your preconstruction data instead of just whatever's on the internet.