15.1.2 About AI Profiles

Data Studio offers AI Profiles for facilitating translation of natural language prompts to SQL statements and leveraging AI capabilities in Autonomous AI Database.

You can create guided AI Profiles in Data Studio for accessing and configuring LLMs using the following modes:

  • NL2SQL
  • Retrieval-Augmented Generation (RAG)

NL2SQL

It uses object lists of Table Metadata for processing plain-language questions as executable SQL queries.

Table Metadata

It is the structured description of a database schema that the system uses. It helps the AI Model to identify database objects such as tables, columns, data types, and so on and differentiate between them. For example, a date can be creation date in one table and expiration date in another. It also enforces security policies and aligns with IAM controls to prevent any unauthorized access or use of data.

Retrieval-Augmented Generation (RAG)

It configures RAG pipelines in Autonomous AI Database and helps the AI Model to load and query unstructured data from knowledge bases, such as documents or databases. It enforces data governance for indexed sources and allows access only to authorized data. For more details, see Select AI with Retrieval Augmented Generation (RAG).

RAG uses Vector Index for semantic search on unstructured data to leverage the vector search capabilities in Oracle AI Database.

Vector Index

It is a search mechanism designed to retrieve vector embeddings generated from RAG pipelines or table data. It enables rapid and efficient search to find similar vectors and returns more accurate responses for natural language queries.

You can create Vector Indexes from Data Studio Settings.

Note:

AI functionalities and LLMs are not supported in Oracle Database 21c, 19c, and earlier. AI features are introduced in Oracle AI Database 26ai.

Moreover, Data Studio also allows you to create an AI Profile with basic AI features using minimal attributes. This is not a specialized profile, therefore the AI searches through all the objects in the data store instead of running a selective query.