At the scale of hundreds of millions of embeddings (the most I have experience with) pgvector has some sharp edges. Vacuuming is a huge one; it’s difficult to v

At the scale of hundreds of millions of embeddings (the most I have experience with) pgvector has some sharp edges. Vacuuming is a huge one; it’s difficult to vacuum indices that are hundreds of GBs in size. And reasoning about recall is therefore challenging. Pgvector’s HNSW implementation also lacks support for certain features (e.g. filtering), which is fine when silently falling back to brute force is acceptable but otherwise not so much. To be clear I think pgvector is an awesome project but it certainly has limitations.