There can be compute spikes when we scale up, which we noticed during our intake season while processing millions of records. Massive similarity searches on a lot of vectors can spike the database CPU and potentially slow down API requests, so we had to move to a higher plan in Supabase for handling this during our intake season. There is no native hybrid search yet, which can combine keyword search and vector search. Supabase supports both, but combining them requires writing a custom Postgres function, while dedicated tools on other platforms allow you to do that out of the box. On some level, we face indexing complexity with Supabase Vector because although vectors expedite searches, we need to use indexes such as HNSW or IVF Flat. Tuning these indexes in Postgres requires advanced knowledge, and we needed a dedicated Supabase expert or to hire someone capable of understanding these complex queries and set this up for us, making it not a plug-and-play solution for a massive scale project with tens of millions of vectors. Vectors are stored in Postgres, and we can perform a lot of similarity searches on millions of vectors, which can spike database CPU and potentially slow down the app, but apart from that, everything seems positive.
