Single-index filtering vs. Categorical HNSW

I am trying to decide on the right platform for my application. In my application, a user will need to be able to search his documents. The solution needs to scale to thousands of users and millions of documents per user. It should have memory requirements proportional to the number of documents for one user, not the overall number of documents of all users.

Here is my question. Is single-index filtering in Pinecone similar to Categorical HNSW in Qdrant (i.e. adding edges that connect vertices corresponding to the same meta-value)?

Pinecone doesnt even use HNSW (which can have these performance issues) it appears they use FreshDiskANN (Great Algorithms Are Not Enough | Pinecone) with some modifications Iā€™m sure.

ARXIV: [2105.09613] FreshDiskANN: A Fast and Accurate Graph-Based ANN Index for Streaming Similarity Search

1 Like

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.