We built data extraction for n8n builders, ask us anything!

Hey Product Hunt 👋

I'm Felix from easybits. We've spent the last year building automation workflows for clients, and one problem kept showing up no matter the industry: getting clean, structured data out of messy inputs (PDFs, invoices, scanned docs) so it can actually flow into an automated workflow.

We tried the tools already out there, but kept hitting the same walls: some were a real pain to set up, and others promised "structured data" but didn't actually deliver it consistently. We'd get JSON back in a slightly different shape each time, so we ended up bolting on post-processing nodes just to force the output into a reliable format before anything downstream could use it.

So we built the easybits Extractor: point it at a document and it pulls out exactly the fields you need as structured data, ready to drop straight into n8n or any workflow tool.

A few things that make it click for us:

• You define the JSON structure on the platform. You set up the exact shape of the output yourself, so what comes back is consistent every time, nothing downstream breaks, and there's no post-processing just to clean up the format.

• Every data field gets a description. Think of it as a mini-prompt per field: it gives the model context for what that field actually means. This is where the real accuracy gains come from, when layouts change, the descriptions keep extraction on track, and refining them is an easy way to dial in precision without touching your workflow.

• Two extraction engines, pick one per pipeline:

  • General (based on Gemini) – great for everyday documents where you want fast, flexible extraction across varied layouts and content.

  • Specialized (based on Mistral OCR) – built for the hard stuff: dense scans, tricky layouts, and documents where OCR accuracy actually makes or breaks the result.

We built this first for our own client work, so it's been battle-tested on genuinely ugly, real-world documents rather than clean demo data. Being able to switch engines means you're not stuck with one model's blind spots.

Would love to hear from you:

• What's the messiest data source you've had to wrangle into a workflow?

• If you're already on n8n – where does extraction still break down for you?

Happy to dig into how it works under the hood.

Best,
Felix