How to pull specific data out of a document with ChatGPT

Name the fields, the format, and the source

Compare a typical request with a specification.

Typical:

Pull out the key details from this contract.

Specified:

Extract the following from this contract. Return a markdown table with one row per item and these columns: field, value, section reference.

Fields: parties, effective date, term length, notice period for termination, total contract value, payment schedule, liability cap, governing law, auto renewal (yes or no).

If a field is not present in the document, write "not stated" in the value column. Do not infer anything.

The second gets you something you can use and check. The first gets you a paragraph.

Ask for the format you actually want

A table for reading: ask for a markdown table with named columns.

A spreadsheet: ask for CSV with a header row, and say what to use if a value contains a comma.

Structured data: ask for JSON and give the exact shape you want, with an example object. Naming the keys avoids getting a different structure on every run.

One value: ask for the value alone with no commentary. "Return only the total contract value, as a number, with the section it came from."

Require the source, always

For each value, include the section number or page it came from.

Two benefits and they compound. You can verify without rereading the document, and paraphrase drift mostly stops, because a citation either exists or it does not.

It also makes the failure visible. A value with no source is either invented or genuinely unsourced, and both are worth knowing before you use the number.

Extraction over many documents

Where this becomes genuinely valuable rather than merely convenient.

Use the identical prompt on each document, one at a time, and collect the rows. Identical prompts produce comparable output, which is what turns twenty contracts into one table you can sort.

1. Write the extraction prompt once

   Fields, format, source requirement, "not stated" instruction.

2. Run it against each document separately

   Separately matters. Several documents at once and the values start blending between them.

3. Collect the rows into one spreadsheet

   Now you can sort, filter and find the outliers, which is usually the actual question.

4. Spot check the outliers against the source

   The unusual values are both the most interesting and the most likely to be extraction errors, so check those first.

Where extraction goes wrong

Numbers in charts. Images inside documents are discarded on every plan except Enterprise, so a figure that only appears in a graph is not available. It may be extracted from a caption instead, which is not the same number.

Truncated documents. If the document was past 2 million tokens, fields that appear late come back as "not stated" when they are actually present. Check what the final section is called before trusting any absence.

Ambiguous fields. "Total value" in a contract with fees, expenses and options has several defensible answers. Say which one you mean.

Scans. Nothing to extract until OCR, and OCR errors become extraction errors that look exactly like real values.

Why this is the task to trust most

Of the things people ask of documents, extraction is the most grounded. It is a search problem with a verifiable answer, rather than a judgement call.

Summarising involves deciding what matters. Analysis involves an opinion. Extraction asks what the document says in a specific place, and a source reference makes the answer checkable in seconds. That is why it is worth structuring properly.