Search is the second database most systems add: something that answers "biryani near me, open now" or "messages from Asha last March" in milliseconds, which the main database can't do well (A separate search index). Since 2023 it has also become where vector embeddings live, for semantic search and retrieval-augmented generation (RAG).
The clouds offer three families: a managed Elasticsearch-style engine, a search platform with built-in AI ranking, and vector search bolted onto a database or object store you already have. This page compares them, covers the licence history that split Elasticsearch in two, and looks at two companies that ran search at scale and what they changed. The Elasticsearch lesson and Hybrid search cover the concepts.
Side by side
| AWS | Google Cloud | Azure | Self-hosted | |
|---|---|---|---|---|
| Search engine | OpenSearch Service (domains) and OpenSearch Serverless (collections) | Elastic Cloud from the marketplace, or Agent Search (formerly Vertex AI Search) as a managed search product | Azure AI Search | OpenSearch, Elasticsearch, Solr, or Lucene directly |
| Vectors in the engine | k-NN indexes in OpenSearch; vector search collections in Serverless | Vector Search (formerly Vertex AI Vector Search) | Vector fields in the same index, hybrid queries with RRF | k-NN in OpenSearch or Elasticsearch |
| Vectors next to your data | pgvector on RDS and Aurora; S3 vector buckets | pgvector and the ScaNN index on AlloyDB | pgvector on Azure Database for PostgreSQL | pgvector on your own Postgres |
| You pay for | Domains: instance-hours and storage. Serverless: OpenSearch Compute Units for indexing and search | Per query or per node, depending on product | A pricing tier, scaled by adding replicas and partitions | Machines and people |