Engineering · 6 min read · September 2026
Converged databases for AI: Oracle's talk at Startup Summit 2026 and what it changes in practice
Vectors, JSON, graph, relational and text search in the same engine. That was Oracle's topic at Startup Summit 2026. The question is when it's worth it and when it's dead weight.
What a converged database is
In Oracle's talk at Startup Summit 2026, the core idea was this: instead of a relational database + a vector database + a search index + a JSON store, all of that lives in a single engine. You store the document, the embedding, the metadata and the relationship in the same place, and query it with a single language. Oracle has been selling this for years; Postgres with pgvector, SQL Server and others are heading the same way.
Why this matters for applied AI
A typical RAG system has an annoying pain point: the data feeding semantic search and the data driving the business rule live in different databases, and keeping them in sync is work. In a converged database, vector search and the relational query happen in the same transaction — less copying, fewer "the vector is stale" problems, less infrastructure to run.
The honest counterpoint
- Lock-in: real convergence ties you to the vendor. pgvector on Postgres ties you to no one.
- Cost: a converged Oracle license doesn't compete with a managed Postgres for most cases.
- Maturity: native vector search in open databases still trails dedicated solutions at scale — but the gap is closing fast.
What Tensoor does today
For most projects, Postgres with pgvector does the job: one database, no lock-in, and the client's team can maintain it. When the client is already an Oracle shop and the volume justifies it, converged stops being overkill. The choice belongs to the problem, not the showcase — as always.
More on how we choose our stack is in Software by company size.