RAG is working, but how to directly query an FAISS knowledge bank?

[Neil_B]

Hi All,

We're currently running version 12.3.2 and have had some good success setting up an LLM flow using the public preview recipes.  I've got a corpus of text going through the embeddings recipe and generating an FAISS knowledge bank.  I've been able to use retrieval augmented generation against the knowledge bank using a prompt recipe.  All of this is working great. 

What I'd like to understand is, how can I simply query against this knowledge bank for nearest neighbors?  I know that when i run the prompt recipe, the 5 nearest neighbors are being retrieved and passed to the LLM as context, and are also cited in the output.  If I have a new text string and simply want to retrieve and output the 5 nearest neighbors and their associated document chunks from the knowledge bank, without passing them back to the LLM as part of a prompt, how can I do this?  Can anyone provide a python example? 

Additionally, for the dataiku folks, will we eventually have another recipe allowing us to do this?

[AgatheG]

Hi Neil,

You might want to have a look at the API reference page for LLM Mesh, more precisely on the KnowledgeBank & associated methods: https://developer.dataiku.com/latest/api-reference/python/llm-mesh.html#dataiku.KnowledgeBank

Hope that helps!

Agathe

[Neil_B]

Thanks for the link.  I looked there first before posting, and the section of the documentation in question seems to have been greatly enhanced since I posted my question.  That definitely gives me what I need now, thank you.

-Neil