Tips or flows for improving RAG Agent responses?

Questions
,
1

I am building a support chatbot. I have connected an AI Agent node with my OpenAI account and a Supabase containing vectors for markdown format documentation about my product. It works, but the answers it provides are often incomplete or lacking critical detail.

Is there a way to improve the responses by ‘correcting the bot’ and storing the feedback or should I simply add the new info into the knowledgebase it is drawing from and re-index that information?

i.e. if the user asks “how do I do X?” and the bot responds with instructions that are only half right, should I adjust the source content, or can I add some sort of human-in-the-loop that actually teaches the AI to return better responses?

(Running 1.95.3 on a Hostinger account.)

2

I recently saw these videos which is supposed to make your RAG responses more accurate. I also believe you have to give proper metadata.

  1. Use a reranker
  2. Improve your meta data
3

Interesting. I’ll have to update my n8n to get that option it seems! Worth a shot…

4

Let me know if it helps, and if so, please mark my answer as the solution. Maybe as a start you can use “npx n8n” to get the latest version on your local machine to test before you upgrade. But always a good idea to keep you n8n instance current.

5

Looks like my boss has asked me to drop the project for the moment, so I won’t have the opportunity to test this - especially since Reranker would cost some money to pull in, it seems. I do appreciate the input though. It looks like a good method from the videos anyway!

6

You should still be able to get decent results just using metadata without the reranker. But I understand priorities. If you want to share the workflow you had, we can still try and fix it up for you for future reference

7

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.