Case Tested

Turning Supplier Documents Into Structured Product Data

Supplier invoices, price lists and spec sheets arrived in every format and were re-keyed by hand. We built rule-based extraction and matching that turns them into structured product data, flags every ambiguous match for a person, and never fills a value it cannot source. In a matching trial, 44 of 45 models matched against about 6,500 records.

Status: Tested on real data — results reviewed and accepted by the business; not yet a production pipeline.

The problem

Supplier invoices, promotional price lists and spec sheets arrived as PDFs and spreadsheets, each supplier in its own format. The same model could appear with a prefix, extra spaces or a variant suffix. Staff re-keyed invoice lines, looked up specifications one by one, and cross-checked everything against internal records.

What we built

  • Invoice PDF → structured sheet — one fixed format per invoice: model, net unit cost, quantity, supplier, invoice number, status and notes, with a final total row that reconciles cost × quantity. Tax, freight and backorder-only lines are left out; credits keep a negative quantity.
  • Supplier and invoice identification — supplier name variants mapped to one canonical name; invoice numbers extracted.
  • Model matching — exact match first, then normalized (prefixes, spacing, case), then a controlled fuzzy fallback. Anything that matches more than one record is flagged, never picked silently.
  • Specification enrichment — dimensions, unit conversion and technical specifications filled from official sources in a fixed priority order, with the source recorded. If no reliable source exists, the field stays blank and is listed as an exception.
  • Price and product sheets — filterable product lists from supplier PDFs, price write-back by model, and combined product, price and stock sheets with gaps marked.
  • Image mapping — official product images mapped to SKUs, up to three per SKU, on a tested sample.

Results Tested

  • Model matching trial: 44 of 45 models matched against ~6,500 records; 10 ambiguous matches sent to a person; 1 reported as unmatched
  • Invoice rules applied to real supplier invoices, with the total reconciled on every sheet
  • Product specification sheets produced with dimensions, conversions, source notes and an explicit exception list
  • Image mapping tested on a 10-SKU sample; large-scale runs are still being made reliable

These workflows have run on real business data, and the business reviewed and accepted the results. They are not yet a production pipeline: every output is checked by a person, and we do not publish accuracy rates we have not measured.

Where people stay in control

  • Any match with more than one candidate
  • Whether two variant model numbers are the same product
  • Whether a source is reliable enough to fill a missing value
  • The final check of every sheet before it is used

Key principle Leave a field blank rather than guess it. A visible gap costs less than a silent error.

Which task takes too much time?

Tell us how your team does it today and which software and files are involved. We can work through where a tool or a different process would help.

Discuss Your Project