Neon has released a new version of pg_embedding, its Postgres extension for vector similarity search. The key update moves HNSW index construction from in-memory to on-disk, enabling compatibility with Neon's serverless architecture including autoscaling and scale-to-zero. The release also adds support for cosine similarity, Manhattan distance, and Euclidean distance metrics. While on-disk indexing is roughly 2x slower than in-memory, the trade-off enables AI/LLM applications to scale to millions of users without managing a separate vector store. The post covers HNSW tuning parameters (dims, m, efConstruction, efSearch) and how Neon's read replicas can offload vector search workloads.
