Find answers from the community

Updated 9 months ago

How do implement something like Auto

How do implement something like Auto Merging Retriever but using two Supabase VectorStoreIndex .

I have defined two separate vector store indices one for nodes and the other for leaf nodes. How do I retrieve from these nodes such that the output is something like Auto Merging Retriever?

14 comments

LLogan M

you cant rely purely on a vector store for this. You need the docstore (or some way to provide a key-val lookup)

bbeaverTango

I have an impletentation that sort of works but gives better output when coupled with a re-ranker

bbeaverTango

But I would want to shift the docstore to supabase at some point

bbeaverTango

It dosen't seem to supported as of now

bbeaverTango

https://github.com/run-llama/llama_index/issues/10696

Just saw your comments on the feature request. What do you think?

bbeaverTango

The primary problem with my implementation is that I'm not getting the same outputs as AutoMergingRetriever when used with a SimpleDocumentStore and VectorStoreIndex. I don't know where I'm messsing this up

LLogan M

I know supabase will likely retrieve different nodes than the default vector index (supabase likely uses HNSW retrieval)

LLogan M

So the results won't be exactly the same, but they should still be good.

Like you mentioned, high top k + reranking is a good choice

bbeaverTango

Oh, okay. This makes sense why the retrieved nodes are different,

bbeaverTango

Could you specify what algorithm VectorStoreIndex uses for retrieval ?

LLogan M

The default vector store just does pairwise comparison between the query eng every node lol very simple

supabase is (likely) using HNSW, an approximate method that scales much better

bbeaverTango

The pairwise comparison might be in efficient but still seems to produce a better results for queries as compared to hnsw then. Is there someway I can specify any other retrieval method?

LLogan M

not sure, I haven't use supabase at all 😅

bbeaverTango

No worries 😄

Add a reply