How to Parse Scanned PDFs into Clean Data with Reducto
Reducto is a document parsing API. You send a PDF, scan, or image, and it returns text with reading order preserved, tables as structured data, and optionally a JSON object matching a schema you define. Old rule-based extractors choke on skewed scans and merged table cells; Reducto reads the page as an image the way a person would.
When is a vision parser worth it over a text extractor?
When the document is scanned or badly typeset. A born-digital PDF with clean text extracts perfectly with a free library in a few lines of code. Sending it to a paid vision model adds cost for no gain.
The line is usually about scan quality and layout. Multi-column invoices, rotated pages, forms with fields in boxes, and anything with tables spanning pages are where the extra cost pays back.
Run one test: if your existing extractor returns empty strings or jumbled columns on 20% of your documents, a vision parser fixes that immediately.
How do you run a schema extraction?
- Upload the file to the parse endpoint and get a parse ID.
- Define a schema: field names, types, and a short description for each.
- Call the extract endpoint with the parse ID and schema.
- Review the output as JSON — numbers as numbers, not strings.
- Add validation rules downstream: totals that should add up, dates that should be in range.
Field descriptions do the heavy lifting. “Line item quantity as an integer” gets a different result from “quantity”. Write them like you are briefing a clerk.
How well does it handle tables and charts?
Tables come back with structure preserved, including merged header cells flattened sensibly. That is the main reason people switch — cell alignment is where cheap parsers fail hardest.
Charts are harder. It extracts the surrounding text and can describe a chart, but reading values off a bar chart is approximate. If chart values matter, treat extraction as a hint and verify against the underlying data.
What does it cost per page?
Pricing is per page processed, with a free tier for evaluation and volume discounts on paid plans. Schema extraction usually costs more than plain parsing because it involves a second model pass. A 50-page annual report with schema extraction is a reasonably sized job, not a free one.
Budget before you batch. Twenty thousand pages at even a few cents each is a real line item, so sample 200 pages first and measure accuracy — the cost of a bad run is the re-run.
Where does accuracy drop?
Handwriting. Printed text is reliable; a form filled in by hand is not. Treat handwritten fields as reviewable.
Stamps and overlays. A signature or a rubber stamp across a table cell confuses alignment. Some vendors offer to detect and ignore them.
Very low-resolution scans — under about 200 DPI — lose small print. Rescan at 300 DPI if you control the source.
How do you fit it into a RAG pipeline?
Use layout-aware chunking rather than fixed character counts. Reducto returns chunks that follow document structure, so a table stays with its heading instead of being split mid-row.
Store the page number with each chunk. When an answer is wrong, you need to know which page produced it, and that is the difference between a five-minute fix and an afternoon.
For high volume, the async pipeline API lets you submit batches and collect results rather than blocking per document.
Who is this for?
- Finance teams ingesting invoices and statements at volume.
- Legal and compliance teams turning archives into searchable, structured data.
- Engineering teams building document Q&A who are tired of writing regex for eighteen different layouts.
Not for: born-digital PDFs you can parse for free, or documents where a single field error carries legal weight — those need a human in the loop regardless of vendor accuracy claims.
Quick Summary
- Reducto uses a vision model to parse scanned and complex PDFs, returning text, tables and schema-shaped JSON.
- Only worth the per-page cost on scans and messy layouts — clean text PDFs parse free.
- Field descriptions in the schema matter more than field names.
- Tables extract well; values inside charts are approximate.
- Weak spots: handwriting, stamps over table cells, scans below about 200 DPI.
Endpoint details and per-page pricing are on the Reducto tool page.
