TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · The Brief · AGE 17—19 · POSITIONAL · METHOD
Working a long document
Work one section at a time, because a page-true citation is one you opened.
When to use
Any document longer than a few pages: a development control regulation, a tender, a conservation report, a thesis you are reviewing.
The method
Start by asking the model for the document's structure — sections, headings, page ranges — and check it against the real contents page. Then work one section at a time, naming the section each turn. Demand page-true citations: every claim carries a page number, and "show me the sentence" is a question you ask often. Quote rather than paraphrase wherever a word matters — a bye-law's "shall" is not "should". Keep your own running note of what has been verified. When the conversation grows long, start a fresh one with the section you need, rather than trusting the model to hold the whole file.
Watch for this
A page number that looks exact and is not. Models produce citations the way they produce everything: plausibly. A page-true citation is one you opened. Watch the middle of a long file too — on some models, material there is recalled less reliably than the start and the end; test yours rather than assume it.
The Lab's note
Map first, one section at a time, page-true or nothing: this is the Lab's way of working a long document, arrived at by watching page numbers drift in student work. Other methods will do; this is the one we teach and can defend line by line.
Try it
Upload a long regulation. Ask for the rule on one specific matter, with page and quoted sentence. Open the page. Do this five times across the document — early, middle, late. Note where it was right and where it drifted.
Prove it
Show one set of claims extracted from a long document where every line carries a page reference you personally opened — and one line you caught that was wrong.
How it works
The document becomes tokens in the context window, and attention spreads unevenly across them. Measured work on 2023 models found recall weakest for material in the middle; newer models have narrowed that gap, and how much remains depends on the model and the task. Working section by section keeps each turn short enough that the answer does not depend on the curve at all. Some products add retrieval — they search the file and pass only the matching passages into the window. That helps, and it also means the model may never have "read" the page you assume it did.
What this idea builds on
- A starting idea.
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.