Is there a way to get the least relevant results?

Feature Requests
1

Hey!

I‘ve been using Pinecone for a bit and it works great!

However, I have a use case for which I didn‘t find anything in the documentation: Essentially I have a spectrum of statements and want to find contradictions in it. Inspired by Four-valued logic - Wikipedia and Catuṣkoṭi - Wikipedia

Now, in this prototype of an experiment I‘m envisioning, I would not only need the best matches (i.e. what we can already get through top_k), but also the worst ones. So that I input an embeddings vector and get the existing embedding IDs from Pinecone where the cosine similarity score is as low as possible. Is that possible somehow?

2

I’m not sure how to do that with Pinecone, but one potential solution would be to compute the inverse (negative) of the query vector and use that for retrieval. The results you retrieve using this method would likely be the least similar, as the cosine similarity score would be at its lowest when using the inverse vector.