Oracle Database 23ai puts vector search and JSON documents inside the relational engine

September 19, 2026 Benjamin

Oracle Database 23ai expands the company’s long-running relational database with features aimed at AI application development and modern document-oriented software.

Two of the most important are AI Vector Search and JSON Relational Duality. The first adds semantic vector search inside the database, while the second lets applications work with JSON document views over normalised relational data.

TechnologyBlog.co.za has not benchmarked Oracle Database 23ai. This article focuses on the architecture and developer implications documented by Oracle.

AI Vector Search keeps embeddings next to business data

Vector embeddings encode semantic information from text, images and other content as numerical arrays. Similarity search compares those vectors to find items that are close in meaning rather than merely matching keywords.

Database 23ai adds a native VECTOR data type, vector indexes and SQL operations for similarity search.

That means an application can combine a semantic search with ordinary filters, joins and security controls without automatically moving data into a separate vector database.

RAG is a natural use case

Retrieval-augmented generation, or RAG, retrieves relevant enterprise information before asking a language model to produce an answer.

Native vector search can support that retrieval step while keeping governed business data inside the database.

RAG can improve relevance, but it does not guarantee factual answers. Application designers still need evaluation, access control and source attribution.

JSON Relational Duality separates storage format from application format

Developers often like JSON because an application object can be represented as a self-contained document. Database administrators value relational normalisation because it reduces duplication and maintains consistency.

JSON Relational Duality allows data to remain in relational tables while exposing updatable JSON documents through Duality Views.

Oracle’s goal is to let developers work with a document model without duplicating the underlying source of truth.

The converged-database argument is architectural

Oracle has long positioned its database as capable of handling relational, JSON, spatial, graph and other data types. Vector support extends that strategy to AI.

The advantage is fewer data movements and fewer specialised systems to operate. The trade-off is deeper dependence on one database platform and its operational model.

Architecture teams should therefore evaluate workload fit, licensing, skills and portability rather than assuming one database is automatically preferable to a specialised service.

Who is Database 23ai for?

The strongest fit is an organisation already using Oracle Database that wants to add semantic search, RAG or document-style application development without immediately introducing another database technology.

New applications can also use the features, but teams should compare them against cloud-native and open-source alternatives based on the wider architecture, not on AI branding alone.

Oracle Database 23ai feature overview

Specification	Details
Database model	Relational converged database
AI Vector Search	Native VECTOR type, indexing and similarity queries
RAG support	Vector retrieval can be combined with enterprise data
JSON Relational Duality	Updatable JSON views over relational tables
Query language	SQL plus supported APIs
Other data models	JSON, graph, spatial, text and relational
Primary users	Enterprise developers, DBAs and data teams
Key architectural idea	Bring AI search to governed data instead of moving data elsewhere

South African deployment questions

South African Oracle estates may value continuity with existing skills and data governance, but current support matrices and local partner capability should be checked before an AI-search project. Procurement teams should also evaluate local partner capability, support escalation, data-location requirements and the skills needed to operate the platform after consultants leave. Those factors can turn a technically sound deployment into a durable system rather than a permanent implementation project.

The metric that matters is operational change

The final test for Oracle Database 23ai is whether it changes an outcome: shorter resolution times, fewer manual steps, faster analysis, better data quality, lower infrastructure overhead or a new service that could not be delivered before. Feature adoption is not the same as value. A smaller deployment with clear measurable impact can be more successful than a broad rollout that employees rarely use.

Architecture comes before the feature list

Oracle’s database release that introduced vector search and modernised JSON and graph development before being superseded in name by AI Database 26ai. The enduring architectural point is that semantic vector search can live beside relational data, transactions and security controls instead of requiring a separate vector store. Enterprise software is rarely valuable because one screen contains more buttons. The important question is how data, identity, workflows, integrations and operational responsibility fit together. As of 2026 the 23ai name is no longer Oracle’s current long-term-support branding; Oracle says AI Database 26ai replaces it through a release update. A product can be technically capable and still fail if the organisation has not defined ownership or cleaned up the process it intends to automate.

AI makes governance more important, not less

Many current enterprise platforms now add assistants, agents or model connectivity. That can reduce manual work, but it also introduces new paths through which data is retrieved and actions are triggered. Permissions, audit logs, evaluation and human escalation therefore need to be designed alongside the AI feature. The responsible approach is to treat a model as another production component that must be monitored rather than as an infallible expert.

Integration determines how quickly value appears

Most large organisations already have databases, identity systems, document stores and line-of-business applications. The implementation challenge is therefore not greenfield installation but connecting the platform without duplicating uncontrolled data or creating brittle point-to-point integrations. Open APIs, connectors and migration tooling matter because the cost of integration can exceed the licence cost over the life of a programme.

How to evaluate the platform

Teams should evaluate the current 26ai documentation, compatibility, licensing, vector-index behaviour, application APIs and migration path rather than planning a new deployment around an outdated product name. A proof of concept should use a real business workflow, real permission boundaries and measurable success criteria. Demonstrations built on clean sample data can hide the messy conditions that determine whether enterprise software works in production.

Five questions worth asking before committing

Before adopting Oracle Database 23ai, write down the problem it is meant to solve, the metric that will show improvement, the systems or people it depends on, the failure mode that would hurt most, and the support path when something goes wrong. Teams should evaluate the current 26ai documentation, compatibility, licensing, vector-index behaviour, application APIs and migration path rather than planning a new deployment around an outdated product name. That exercise prevents a technically impressive product from becoming a solution in search of a problem. It also creates a baseline for later review: if the expected outcome does not improve, the organisation can change configuration, training or even the product choice instead of defending the original purchase.