Long-Form Document Processing Strategies

Run the Long-Form Document Processing Strategies Diagram Fullscreen

About This Diagram

Sooner or later you will hand an AI something too big for it to read in one go. What you do next depends on two questions, in this order: does it fit, and what are you actually trying to get out of it?

Those two questions are the only forks in this flowchart. Everything downstream follows from them. Summarizing a long report, answering a specific question about it, and analysing its structure are three genuinely different jobs, and using the summarization approach for a question-answering task is one of the most common ways to waste tokens and get a worse answer.

How to Use

Start at the top with your document and follow the arrows.
Hover over any node to see a detailed explanation in the panel on the right, including the practical gotchas for that step.
Trace all four paths to the Final Output box. Notice that "Process directly" is a one-hop path and the others are not — fitting in the window is worth a lot.

Embed This Diagram

Place the following line in your website to include this diagram:

<iframe src="https://dmccreary.github.io/prompt-class/sims/long-form-document-processing-strategies/main.html"
        height="532px"
        width="100%"
        scrolling="no"></iframe>

Workflow Steps

Step	What happens
Long Document	Estimate the size in tokens before anything else
Fits in context window?	Compare against the window, leaving room for instructions and the response
Process directly	Send the whole thing with a clear instruction
Task type?	Summarization, question answering, and analysis diverge here
Chunk and summarize	Split, summarize each piece, then summarize the summaries
Extract relevant sections	Send only the sections that bear on the question
Hierarchical outline	Outline first, decide what matters, then drill down
Final Output	All paths converge; verify the result against the source

Lesson Plan

Learning Objective

Students will be able to select an appropriate long-document processing strategy given a document size and a task type, and justify the choice.

Bloom's Level: Apply (L3) — select, apply

Grade Level

High school through adult learners.

Duration

10-15 minutes

Prerequisites

Students should understand what a context window is and be able to roughly estimate token count from word count (Chapter 7).

Activities

Trace the paths (4 min): Students hover every node and identify which of the four routes to Final Output is shortest and why.
Route real scenarios (7 min): Give the class four scenarios — a 200-page contract with one question about a liability clause, a 12-page memo to summarize, a 60-page research paper to critique, a 3-page email thread. Students route each through the diagram and defend their answer.
Find the failure mode (4 min): Ask what goes wrong if you use chunk-and-summarize for a question-answering task. Students should reach the answer themselves: the chunk containing the answer gets summarized into oblivion before the question is ever asked.

Discussion Questions

The diagram treats "fits in the context window" as a yes or no question. In practice it is a spectrum. At what point should you treat a document that technically fits as though it does not?
Chunking loses information that spans a chunk boundary. What can you do about that?
Why does the analysis branch ask for an outline first instead of just chunking?

Assessment

Can the student route an unfamiliar scenario to the correct branch?
Can the student name a specific risk of the branch they chose?
Can the student explain why all four paths end at the same verification step?