DEV Community

Cover image for K-shot training with LLMs
Anton Loss
Anton Loss

Posted on

K-shot training with LLMs

I built a tool for teaching LLMs how to extract structured data from documents by annotating, not prompt engineering. I’d love your feedback.

How it works:

  • Upload a document (DOCX, PDF, image, etc.)
  • Select and tag parts of it (supports nesting, arrays, custom tag structures)
  • Upload another document → click "predict" → see editable annotations
  • Amend them and save as a new example - Call the API with a third document → get JSON back

Use cases:

  • Identify "important clauses" in contracts
  • Extract "total value" from invoices
  • Anything subjective, like "healthy ingredients" on a label
  • Anything objective, like "postcode" or "phone number"
  • You could even tag things like "good rhymes" in a poem — basically anything an LLM can understand and extrapolate

The key idea: instead of iterating endlessly on prompts (and sometimes regressing), you just iterate on examples. Each example improves accuracy in a concrete way, and you need far fewer than traditional ML approaches.

We’re launching on Product Hunt today (currently #5)
https://www.producthunt.com/products/deeptagger

Top comments (0)