Just brute force your embeddings

For datasets with around one million documents, brute-force vector search using numpy on standard hardware can be efficient enough, eliminating the need for complex vector databases. This approach leverages simple matrix operations to achieve high query throughput and low latency.