Files and research with sources
A model only knows what was in its training – unless you give it documents or let it search. How to have files analysed and research done without believing references you never opened.
Boaz Lichtenstein
A language model knows neither your contract nor this morning’s price list. There are exactly two ways to give it that kind of knowledge: you upload documents, or you let it search the web. Both cut the error rate considerably – if you demand references and check them.
Files: structure first, then the question
PDFs, spreadsheets, presentations and images can be uploaded to every major assistant. Three habits turn that into reliable work:
Have it show you the structure first. “Which sections does the document have, and what is in each?” That takes ten seconds and shows whether the file was read completely – scanned pages or tables as images tend to drop out.
Demand the reference. Every statement with page or section.
Answer the question from the attached contract only.
Give page and section for every statement.
If the contract does not cover something, write “not covered”.The third sentence is the most important. Without it the model fills gaps with whatever is usual in contracts.
Have it calculate, not read off. With tables: totals and averages should be computed by the assistant with its analysis tool, showing the working. A number it “reads from the table” is an estimate – why is explained in Where AI reliably fails.
Web search: quick or thorough
Simple search fetches a few pages and answers in seconds. It is good for facts with one clear answer: a date, a price, who is responsible.
Thorough research – called “deep research” or something similar, depending on the provider – works for minutes, reads dozens of sources and writes a report with evidence. It pays off for questions with several sides: a market overview, a comparison, the state of a debate.
The same four-question check applies to both:
- Does the source exist, and does the link open?
- Does it really say that? Find the passage on the page.
- When is it from? A price from two years ago is not a price.
- Who is speaking? The manufacturer, an authority – or a blog copying from the manufacturer?
The fourth question separates primary source from retelling. For prices, deadlines and features, only the page of whoever sets them counts.
The rule that stays
A reference you have not opened is not a reference. That sounds strict and in daily work is a matter of two minutes: you do not check all twelve sources, only the two or three your decision rests on. Why models invent evidence in the first place is explained in the next station.
And before you upload anything: customer data, staff records and other parties’ contract details have rules of their own – ChatGPT and privacy says which.