Oracle Adds Bedrock to Select AI While Keeping More AI Work in Its Database

Oracle has added Amazon Bedrock support and nine other Select AI capabilities for Autonomous AI Database on Dedicated Infrastructure. The package brings local embedding, correction memory, synthetic data and agent tools under a shared configuration model, designed to keep more AI processing near the database.

The configuration layer

Select AI accepts natural-language prompts through SQL and PL/SQL. For SQL tasks, it builds an augmented prompt from database metadata and sends it to a customer-selected large language model, or LLM, to generate, run or explain a query. An AI profile holds the provider, model, credentials and database objects used by these functions.

The AWS option lets Bedrock models generate SQL against Oracle data. Users must specify a Bedrock model, inference profile or ARN, and configure network access for the chosen region. On Dedicated Infrastructure, a database permission to make an outbound call does not itself create the route needed to reach Bedrock.

Retrieval and correction inside the database

For retrieval-augmented generation, Select AI can run an imported ONNX embedding model through the database’s ONNX Runtime. Document chunking and vectorization stay inside the database perimeter, although retrieved context still goes to an LLM for the final answer. Oracle says embedding inference uses ECPUs, so a large initial vectorization job can compete with database workloads.

Select AI Feedback keeps prompts and corrections in a vector index, then retrieves similar examples as hints for future natural-language-to-SQL requests. That is retrieval-based prompt engineering, not model fine-tuning. Profiles can also supply table and column comments, annotations and foreign-key constraints as model context.

From questions to actions

• Schema-aware synthetic-data generation can use data types, constraints and referential integrity, with parallel processing for larger multi-table jobs.
• Conversations can be session-based or persistent, resumable and shareable across SQL, chat and narration operations.
• Property-graph metadata can guide the translation of natural-language questions into GRAPH_TABLE queries.

The agent capability extends the system beyond answers. It uses the ReAct reasoning-and-acting pattern with short- and long-term memory, and can call NL2SQL, retrieval, web-search and notification tools, along with custom PL/SQL procedures. Oracle warns that LLM-generated SQL can be wrong or create security risks; generated queries run in the customer’s database.