8 hours ago
- For small datasets (e.g., ~1M documents), brute-force search using numpy is fast and often sufficient.
- Many teams do not need the complexity of a vector database, especially with low query traffic and precomputed embeddings.
- Performance benchmarks show decent QPS and latency even with 8.8M vectors using simple numpy operations.
- Further optimizations like threading and heap-based top-n collection can improve throughput.