Calculation and long documents
You know why a model makes calculation errors and misses details in a long document, and how to avoid it.
ObjectiveBy the end of this lesson, you will be able to have a calculation set out in a spreadsheet and query a long document section by section, with quotes and pages.
Two other limits often come as a surprise: calculation and very long documents. They have the same origin. The model processes text token by token and does not naturally perform exact operations or exhaustive reading.
Calculation
A model can write correct reasoning and get the final sum wrong. It predicts a result instead of calculating it. Errors mainly affect large numbers, chained percentages, conversions and the totals of long lists.
Calculate the total cost of the training for the 47 participants with the discounts.
Set out the total cost calculation in a spreadsheet: one row per participant category, the discounts in a separate column, the total as a formula. Show me the table.
The calculation goes through a tool that actually calculates, and you can check each formula.
Long documents
A model has a context window: the amount of text it can consider at once. Even when a long document fits in it, its attention is not uniform. A detail in the middle of a 200-page report can be missed, and a summary can give too much weight to the beginning and the end.
- Ask for a summary chapter by chapter rather than an overall summary in one go.
- Ask precise questions: “What does section 4 say about deadlines?”
- Require the reference page or section for each point.
- For an exhaustive search, ask for a list of every occurrence, then check a few of them.
Learnya’s knowledge library helps here. The agent fetches the relevant passages from the documents you have placed in it, instead of relying on an overall reading.