What you're looking for is called "Dissimilarity Search". Similarity search is computationally less expensive than dissimilarity search. The domain is constrained per se, which is harder to guarantee for dissimilarity (bottom-k) search.

What this does remind me of though, is Maximal Marginal Relevance (MMR). A technique used to prioritize diversity in similarity searches. After retrieving the some n (chunk size) * k similar documents, we find the most dissimilar document to K1 (most similar result) from the rest, This is repeated till K results are retrieved, all of which are dissimilar to each other but similar to the query document.

The latter works because of a constrained domain, maybe it could be useful to you.

https://www.cs.bilkent.edu.tr/~canf/CS533/hwSpring14/eightMinPresentations/handoutMMR.pdf

Dissimilarity search is pretty tough, because you have to calculate distances to many more samples. Think of this, we're embedding in 3 Dimensions. Embed and plot the query, draw a sphere around it till you have roughly K samples inside the sphere. Use your similarity function and select the top-k results, your domain is constrained by the sphere. This is incredibly contrived to illustrate the point over text succinctly, but such optimizations deliberately reduce the number of samples to check, so you aren't comparing all the results in your vector database.

You could also set the top_k to the number of samples in your database and select the last N results. This won't scale, though.