AI Tools for Research and Documents: Which Shape of Tool Fits

Two jobs look alike and are not: questioning a fixed set of sources, and producing something from material you supply. The first needs a tool grounded in your documents that cites them; the second needs one that returns a real file. Using one tool for both is the source of the frustration.

The distinction

A question that separates them: once you have the result, what will you do with it? Read it and decide something - that is questioning. Send it to someone - that is production.

Grounded questioning: what matters

The problem being solved: you have thirty documents and the answer is in one of them - but you do not know which.

The requirement not to compromise on: citation

This is the single point that decides whether this is research or a well-phrased guess.

An answer without a source reference is a summary you have to trust. An answer with a quote and a source passage is a finding you can verify in thirty seconds.

So: ask explicitly for a quote with the document and page, not a rephrasing. And when the finding matters - open the source and look. Not on every answer; on the one that will drive a decision.

Where this is done

A dedicated research notebook is built exactly for this - a fixed set of sources you question repeatedly, grounded in them.

A chat with document upload does this well for a small number of files and a one-off question - and there too, ask for a quote rather than a rephrasing.

Connecting to the file store suits the case where you do not know in advance which documents are relevant - which is precisely when the searching is the work.

Production: what matters

The problem being solved: you have the material and the structure, and it needs to come out as a document.

The requirement: a real file

A spreadsheet, document or deck you download and open - not text you paste and restyle. The copying and reformatting is what eats most of the time saved, and a tool that returns only text returns the problem.

The rule: assembly, not invention

A document built from material you supplied is reliable; one written from nothing is a guess dressed as a report.

Hence the central rule: every figure in the document must be traceable to a source you supplied. If it cannot be traced, it should not be there.

Where this is done

If the material lives in Google - work in place, because location saves both copying steps.

If the material is in files on your computer - upload and ask for a file back.

The shared failure: a messy source folder

This breaks both jobs the same way, and it is not a problem with the tool.

If the same folder holds three versions of one document - a draft, the version that was sent, the signed one - the answer can be built on the obsolete one, and it will look exactly like a correct answer.

Three rules:

1. Ask it to state which document it relied on, and from what date.
2. Remove superseded versions from the collection before starting.
3. Remember: the tool exposes an organisation problem, it does not solve one.

What fails in both

• A scanned PDF. A document that is an image behaves differently from one with text, and in Hebrew that stands out. Check in two seconds: try to select a line.
• A table in a PDF. It has no structure - it is lines of text whose positions create the impression of columns.
• Dirty Hebrew. Double spaces and invisible characters break search and matching. A file not found is not necessarily a file that does not exist.
• A single source. An answer resting on one document is a quotation, not research.

The decision

1. What will you do with the result? Read and decide - questioning. Send it - production.
2. If questioning: is there a citation to the source? If not, that is not the tool.
3. If production: does a real file come back? If not, half the work remains.
4. Where does the material live? A tool in place beats a better tool somewhere else.

And the question that precedes all of them: if this has to happen repeatedly without you asking - that is automation, not tool choice.