Oracle AI Database 26ai is the layer everything else in this cluster sits on — and its headline change is that the database you already run for transactions can now do vector similarity search natively, next to the rows. A built-in VECTOR data type, AI Vector Search, and Select AI mean the RAG substrate can live where the business data lives, rather than in a second system you copy into and keep in sync. The honest read is about when that converged bet pays — and when a purpose-built vector store still wins.
Oracle AI Database 26ai is Oracle's current Long-Term Support release, generally available for Linux on-premises servers since January 2026 (the "23.x" release-update numbers you'll see — 23.6, 23.26.0 — run under the 26ai line). Its pitch is convergence: relational tables, JSON (with JSON-relational duality), and now vectors, queried through one SQL engine with one set of transactions, backups, and security.
For the rest of this cluster it is simply the ground: APEX runs inside it, ORDS exposes it, OCI decides where it sits. The 26ai-specific news is the vector work — and that is what makes an Oracle estate a plausible RAG substrate rather than just a source you extract from.
26ai introduces a native VECTOR data type (dense and sparse) and AI Vector Search: you store embeddings in a column alongside the business data and run similarity search in SQL, without moving anything to a separate vector database. Oracle documents the features as available in Release Update 23.6 or later with COMPATIBLE set to 23.6. Because the vectors live next to the rows, a single query can combine a similarity search with an ordinary relational filter — hybrid retrieval without a second datastore to join across in application code.
Oracle's framing is that you should not have to copy business data into a standalone vector store to make it searchable — keep it in place, add a vector column, and the transactional guarantees, access controls, and backups you already have extend to the embeddings. That's a real operational simplification. Whether it's the right performance and cost choice is the next section, not this one.
Select AI lets a user (or an agent) ask a question in natural language and have the database generate and run the SQL, with schema-metadata enrichment to make the schema legible to an LLM. Paired with ORDS 26.2's native MCP support, the picture for a legacy estate closes: the data stays put, gains a vector index and a natural-language surface, and becomes reachable by an agent through a governed REST/MCP door — no migration, no second system. That is the whole cluster's thesis, expressed at the data layer.
The case for 26ai's vector search is data gravity, not raw benchmark supremacy. If your records already live in Oracle, keeping the embeddings there avoids an ETL pipeline, a second system to secure, and the consistency problems of two stores drifting apart — that is a genuine and often decisive operational win. If you are building greenfield, a purpose-built vector database, or Postgres with pgvector at modest scale, may be cheaper and faster for pure retrieval, and you are not paying Oracle Database licensing to get it.
Two more caveats worth stating plainly. The vector features are newer than the relational core that has been hardened for decades — benchmark them on your own corpus and recall targets rather than assuming parity with a specialist. And the licensing is Oracle's: the convergence saves you a second system, not necessarily a second bill. The decision is "is my data already here?" — if yes, 26ai's vector search is a strong default; if no, it has to win on merits against tools built only for this.
The residency argument that runs through this whole cluster gets sharper here. If the embeddings for a RAG system live in a separate vector store, that store is another place personal information sits — another residency and POPIA question to answer, often a foreign SaaS. Keep the vectors in the Oracle database that already holds the records, on infrastructure you already located, and there is one residency answer, not two. For a POPIA-sensitive estate, collapsing the RAG substrate back into the system of record is a compliance simplification as much as an engineering one.