PDF data sheets turned into a structured product database
Product specifications are locked inside PDF data sheets and retyped by hand into the product database or the website. A checked extraction could prepare each product record, with someone reviewing the values before they are published.
Illustrative use case. It describes a problem we come across often and what we could build to solve it. It is not a client project: no results are claimed.
The starting point
The company sells thousands of technical products. For each one, the manufacturer provides a PDF data sheet: dimensions, weight, materials, performance, ingress protection rating, certifications, sometimes several variants in the same document. To fill the product database, the website or the catalogues, someone opens each PDF and copies the values one by one. A few manufacturers also send structured files; most don’t.
The actual problem
Data entry is slow, and data sheets pile up faster than they are processed. A digit copied wrong, or a unit mixed up (watts for kilowatts, millimetres for centimetres), puts a false specification in front of customers. Tables spread over several pages and multi-variant sheets are the hardest to copy without mistakes. Everyone names attributes their own way, so the database becomes inconsistent and the website filters stop working well.
What could be built
- A list of expected attributes per product family, with their units, aligned where possible with a recognised classification such as ETIM.
- Reading each PDF data sheet to extract those attributes, variant by variant, with the page each value came from.
- Automatic checks: expected unit, value within a plausible range for the family, missing required attribute, variants that contradict each other.
- A review screen showing the PDF next to the extracted values, with uncertain fields marked.
- Sending approved records to the product database or the tool that feeds the website and catalogues.
What would change for the team
- Product records are prepared instead of being typed from scratch.
- All records in a family follow the same structure and units, which improves search and filters.
- Every published value can be traced back to the original document.
- Reviewers focus on flagged fields and sensitive values.
What stays under human control
- Approving each record before publication.
- Values related to safety or regulation: certifications, protection ratings, compliance statements.
- Ambiguous or incomplete data sheets, and the questions to ask the manufacturer.
- Choosing the classification and the attributes kept for each family.
Going further
The article AI document extraction with human review explains how to organise the review of extracted values. For prices, which often arrive alongside data sheets, see Supplier price list updates: what to check before you import.