Skip to main content

Clarification Needed on RAG Feature and Continuous Learning Tab

Question

Hi Mindee team,

I’ve been testing the new RAG functionality in version 2, which, as I understand, should allow us to fine-tune the model by providing proper annotations. However, I’m having trouble understanding how the Continuous Learning tab is supposed to work in practice.

After uploading an invoice and letting the model make its predictions (fields are marked correctly), I don’t see any way to interact with or correct the predictions – there’s no clickable or editable interface that would let me provide feedback or annotate fields for further training.

Could you please clarify:

  • What exactly is the expected user interaction in the Continuous Learning tab?

  • How should we provide annotations or corrections to help improve the model?

  • Is there any documentation or a short demo video explaining how this flow works?

Thanks in advance – a walkthrough or clarification would be very helpful.

Posted by Laura Mitchell ·
Status: Completed4 comments

Log in to comment and vote

Comments4

  • Ianaré Sévi

    Team•

    Jul 24, 2025

    Hi and thanks for the feedback.

    To get started we do have some documentation for the RAG feature here:

    https://docs.mindee.com/models/improving-accuracy

    Let us know if that helps you.

    Thank you!

    • Laura Mitchell

      •

      Jul 24, 2025

      Hi Mindee team,

      I managed to figure out how to use the RAG feature, but I noticed some unexpected behavior during testing.

      I fine-tuned a model on an invoice where I annotated specific fields for a few labels. However, when I tested the model with a different invoice from the same company (used as a live test), I saw that for two of the annotated fields, the model returned values from the original training invoice – not from the test one.

      It seems as if those two values were hardcoded from the training invoice, because they don’t appear anywhere in the test document. That raises a concern that the model may be memorizing exact values rather than extracting them from the input.

      Let me know if there’s any insight into this, or if you need more details – happy to share samples if helpful.

      Thanks again for your support.

    • Felix Croizon

      •

      Dec 22, 2025

      Hi, it helps but I have a question : how to fix the result if it picked the correct answer but at the wrong place. For example I want to retrieve a car brand, and it detects the car dealership brand instead of the car brand.

      • Ianaré Sévi

        Team•

        Dec 29, 2025

        Hi, the model will select the best place to grab the info from. If the brand is wrong (not wrong place but wrong value), you can use RAG to “show” the model where to get the correct value.