NarrowToolsBrowse tools
Start here
First decide whether the PDF contains selectable text or scanned images. Extract a small sample, preserve SKU and unit columns, and compare the result with the source before converting the whole file. A spreadsheet that looks tidy can still contain commercially important extraction errors.
Classify the PDF before choosing an extractor
Try selecting a product description and a price. If they can be copied as text, a table extraction tool may work without OCR. If the page is an image, add OCR to the workflow. A mixed PDF can contain both kinds of pages, so inspect more than the cover or the first table.
Record whether tables continue across pages and whether prices are grouped under headings such as case, each, trade, or retail. These headings can define the meaning of many rows without appearing in every row. Losing one can turn an accurate number into an incorrect price.
Pick a route that matches the volume
For a few text-based files, Tabula lets you select table areas and export spreadsheet-friendly data. It is useful for a supervised task where a person can inspect the output. Its own documentation distinguishes text-based PDFs from scanned documents.
For a repeatable development pipeline, Docling provides document parsing components. It can be part of a system that keeps layout information and sends structured output to downstream checks. It still needs evaluation on your supplier documents. Neither approach removes the need to inspect difficult headers and multi-page tables.
Define the target columns before extraction
A useful target usually includes supplier, supplier SKU, product description, pack quantity, unit, currency, price, effective date, source filename, and source page. Keep the supplier’s original code as text so leading zeros survive. Preserve the original price string alongside the parsed numeric value during review.
If your internal SKU differs from the supplier code, keep a separate mapping table. Do not ask an AI model to recreate the mapping from product names every time a new file arrives. A stable reviewed mapping makes later price comparisons easier to explain.
Check the errors that matter to purchasing
Look for shifted columns, repeated page headings treated as products, decimal separators, currency symbols, and pack quantities accidentally merged into descriptions. Sample rows from the beginning, middle, and end of the file. Include unusually long descriptions and products near page breaks.
Compare the number of extracted products with the source where possible. Flag missing prices and duplicate supplier codes. A large price change should require confirmation against the page image before it reaches a purchasing system. A parser returning no error does not prove the content is correct.
Keep review separate from production updates
Save extracted rows into a staging sheet rather than overwriting the live product master. Add a reviewed column and a reason for corrections. Once approved, export the exact columns your purchasing system expects. Retain the source PDF and the reviewed extraction so a future discrepancy can be traced.
For recurring imports, measure how many rows need manual correction per supplier. If one supplier’s layout consumes most of the review time, create a supplier-specific template or ask for a CSV. The most effective automation may be changing the incoming format rather than adding another model call.
Sources
Updated October 30, 2026. Based on the linked documentation; the tools have not been benchmarked by NarrowTools. Editorial policy