need vector search for millions of vectors

For millions of vectors, vector search is absolutely feasible — the key is choosing the right index, hardware, and query pattern.

What to use

1) Approximate Nearest Neighbor (ANN)

For millions of vectors, you usually do not want exact brute-force search. Use ANN methods like:

• HNSW — very popular, high recall, fast queries
• IVF / IVF-PQ — good for larger scale and memory efficiency
• PQ / OPQ — for compressing vectors to reduce RAM
• DiskANN — strong for very large datasets, optimized for SSD

Good options by scale

If you have up to tens of millions

• FAISS (great library)
• Milvus
• Qdrant
• Weaviate
• Elasticsearch/OpenSearch kNN
• pgvector if scale and latency requirements are moderate

If you need high performance at very large scale

• FAISS with IVF/HNSW/PQ
• DiskANN
• Managed/vector DBs depending on ops preference

Practical recommendation

If you’re starting from scratch:

• Need simple, production-ready DB: Qdrant or Milvus
• Need maximum control/performance: FAISS
• Need billion-scale / SSD-based search: DiskANN

Important considerations

• Vector dimension: higher dims cost more memory and latency
• Recall vs latency: ANN lets you trade accuracy for speed
• Filtering: metadata filters can change index choice
• Updates/deletes: some engines handle dynamic data better than others
• RAM vs disk: millions of vectors can get expensive if stored uncompressed

Rule of thumb

If you have, say:

• 1M vectors: HNSW is often enough
• 10M+ vectors: consider IVF-PQ or DiskANN if memory is tight
• Very low latency + high recall: HNSW or tuned FAISS

If you want, I can recommend a specific setup based on:

1. number of vectors
2. vector dimension
3. latency target
4. update rate
5. whether you need metadata filtering