← All posts

Five steps from a PDF catalog to specs an AI can compare

Manufacturers and distributors still keep most product specs in PDF catalogs and spreadsheets. Here is the order of work for turning them into data an AI assistant can filter and compare, shown on a fictional example.

3 min read

By ZAG

日本語版 →

Japan IT Week Kansai opens at INTEX Osaka on November 18 and runs to the 20th. Plenty of the people walking the floor will be the ones who look after product data at manufacturers and distributors.

Ask them where the specs live and the answer is usually some mix of "the latest PDF catalog", "a spreadsheet we keep internally" and "whatever format each supplier sends". That works while people do the reading. Once buyers start asking an AI assistant "which products meet these requirements?", it stops working: specs in that shape don't make the shortlist.

Here is the order of work we follow to turn PDF specs into data an AI can compare. Every example is fictional.

1. Decide the attributes first

Start by deciding what you are going to extract. If attributes get added as you read, each catalog ends up with its own "output", "rated output" and "max output", and none of them can be compared.

For each of your own categories, write down the attribute names, their units and the range of values you expect. It doesn't have to be perfect. What matters is the order: decide, then read, then refine.

2. Separate value, unit and condition

Spec tables pack value, unit and condition into one cell. Suppose the catalog for a fictional pump, EXAMPLE ZC-4820S, says:

Flow 50/60 Hz: 120/140 L/min Operating temperature 0–40 °C

In a form an AI can compare, that becomes:

Attribute (condition) Min Max Unit
Flow (50 Hz) 120 120 L/min
Flow (60 Hz) 140 140 L/min
Operating temp. 0 40 °C

Product, model number and values are all fictional.

Ranges are split into min and max, and conditions like frequency are pulled out of the value and kept with the attribute. Now "at least 130 L/min at 60 Hz" is a filter that works.

3. Keep the source

Attach to every extracted value the catalog, page and table it came from.

With a source, a reviewer only opens the original page for the values that look doubtful. When a catalog is revised, you know exactly which values to recheck.

4. When in doubt, leave it blank

Merged cells, margin notes and caveats like "except some models" trip up human readers too. When the reading isn't certain, leave the field blank and flag it for review rather than guessing.

Everyone can see that a blank is still open. A plausible wrong value goes straight into quotes and comparison sheets. If you only track how full the table is, you lose that distinction.

5. People review the differences, not everything

People still check the result, but not every row. They look at:

  • values flagged for review
  • values that changed since the previous version
  • values that are an order of magnitude off from similar products in the same category

With review narrowed to those three, the effort depends on how many values have problems, not on how many catalogs you have.

For distributors with many suppliers

A distributor gets the same category of product from dozens of manufacturers, each in its own layout. This is where step 1 pays off. With your own attribute definitions in hand, each supplier's way of writing things gets absorbed once, at intake.

In short

Decide the attributes, separate value, unit and condition, keep the source, leave doubtful fields blank, and have people review only the differences. None of it is exotic. But skip the order and you end up with data that can't be compared.