General Question to RAG-System

[2paul]

I have some general questions to RAG-Systems in N8N.
I fill my qdrant-Vector-DB externally with docling (pdf-parser) in Python.
And want make that data accessable with N8N for Mail, Whatsapp, Elevenlabs, and so on.

Here is a simple example that is working:

But i also saw workflows with the “Answer question with vector store”-node, like this:

Can maybe someone explain me the differences and whats the benifits, to use one or the other ??

But the main question is, both workflows make 2 Requests to the Chat-Model in the execution-process.
But that shouldnt be needed.
In my Python-Chat were i directly communicating with the qdrant-Database, the workflow is like this:

• The question get vectorised by the embedded modell
• With this vector the DB is searched
• the question is sended to the LLM with the question and the results from the DB.
• Result in the chat.

So normally just one request to the LLM-Chat-Model should be needed.
Can I realize that in N8N ??
Or do i have to use a webhook to make make that external in this way ??

[mrrobot]

Hey @2paul ,

Can you try this
Gmail Trigger → Manual RAG Pipeline → Gmail Response without the Ai agent node

[2paul]

Thanks for your answer.
So there is no way to do a RAG-System with just one LLM-Call in n8n ??

[2paul]

For all that asking the same question, here is the solution:

RAG Movie Recommendations chatbot with Qdrant and n8n