š Hey dev.to community,
One of our users reached out with a problem I think a lot of people hit: he was extracting a data table from a multi-page PDF, and the cells kept bleeding into each other. About 95% of the data came out right, but 5% got mixed up with the wrong rows, so he could never fully trust the result.
This week we shipped something to fix exactly that: an extraction engine dropdown you can set per pipeline. In the video I run the same messy multi-page table through both engines, with a small n8n workflow that checks every extracted cell against a reference so you can actually see what slipped.
What the two engines are:
The General engine runs on Gemini and covers about 90% of everyday extraction (image description, classification, normal documents). The Specialized engine runs on Mistral and is OCR-optimized for document-heavy work like dense or multi-page tables.
What the test showed:
The General engine slipped on a couple of rows and came back with pass = false. Switching the pipeline to Specialized took the same document to 100%, every cell correct. The bonus I did not expect: Specialized also ran faster on the multi-page PDF.
A couple of takeaways even if you skip the video:
For dense or multi-page tables, reach for the Specialized extraction engine. For most other jobs, General is the right default.
Do not eyeball table extraction. A tiny workflow that cross-checks each cell against a known-good reference tells you exactly which rows are wrong, instead of you scanning 20 rows by hand.
If rows still bleed after switching engines, it is almost always the response structure. Model the table as one records field set as an array of objects with each column nested inside, not one separate list per column.
Want to try the new engine on your own tables? The easybits Extractor is a verified community node with 50 free monthly API requests included. On n8n Cloud, just search 'easybits Extractor' in the node panel, no install needed. Self-hosted, install '@easybits/n8n-nodes-extractor' from Settings, Community Nodes.
I put a full step-by-step guide (PDF) for setting up your extractor for data tables here: https://github.com/felix-sattler-easybits/n8n-workflows/tree/ee1ed5fe0a3e898843422a619922cedb7cf618c4/easybits-data-table-extraction (the validation workflow from the video is in that same folder too, so you can import it and try it on your own tables)
What is the most stubborn multi-page document you have tried to pull a table out of?
Best,
Felix
Top comments (0)