Oracle Select AI : Why Metadata is Important

Metadata is the key enabler of Select AI because general-purpose Large Language Models (LLMs) have no inherent knowledge of your specific business, enterprise, or private information. Without metadata, an LLM cannot answer questions about your data in a meaningful way and is likely to hallucinate, inventing table and column names that do not exist in your database

👉 What is Oracle Select AI

Metadata provides the necessary context that transforms a general AI into a specialized developer assistant for your specific database.

•  Preventing Hallucinations: While an LLM might produce a syntactically correct SQL query on its own, it will fail if it "hallucinates" the schema. Metadata ensures the generated SQL uses your actual table and column names.

• Overcoming Temporal and Knowledge Limits: LLMs are limited by their training data and have "temporal limitations," meaning they only know information up to a certain date. Metadata bridges this gap by providing real-time information about your database structure.

• 
  Handling Generic Schemas: In "pathological cases" where tables are named generically (e.g., if we have "Table 1" ,"Table 2" and Table3 and columns are equally generic as shown below, if we ask LLM a question against this schema , it will have no idea about how to formulate a query. But if we define comments for our table and columns then this information can be added to table's metadata that provided to LLM.That will help LLM to formulate runnable queries). Metadata—specifically comments—is the only way to help the LLM understand how to formulate a query.

 How Select AI Uses Metadata

 Select AI utilizes a process called Prompt Augmentation to integrate your database structure into the AI's reasoning process.

 

Step	How Metadata is Utilized
Augmentation	Select AI combines your natural language prompt with table names, column names, data types, and optional comments.
Contextual Delivery	This "augmented prompt" is sent to the LLM (e.g., OCI Gen AI, OpenAI) to provide the context needed for SQL generation.
AI Profiles	Users specify an Object List in an AI Profile to limit the metadata sent, ensuring the prompt does not exceed token limits while maintaining accuracy.
Tuning	By adding comments to columns, you "tune" the metadata, making it a significant asset for the LLM to produce better queries.

 The Future of Metadata in Select AI

Oracle is currently developing features to make metadata management even more efficient:

• Automated Metadata Selection: For large schemas with hundreds or thousands of tables, Select AI will use AI Vector Search to automatically determine which specific metadata is most relevant to a user's prompt.

• Enhanced Constraints: Future updates will include foreign key constraints and annotations in the metadata to help the LLM better understand the relationships between different tables and views.

• Metadata Clones: This allows for the creation of a new database containing the same metadata as the source but without the actual data, specifically for synthetic data generation.