Your AI Agent Wants Your Data. Here's How to Lock It Down - Jeff Smith, Oracle

하이라이트

• Oracle native support for vectors allows databases to participate directly in RAG pipelines alongside relational and graph data workloads.

• Natural language to SQL conversion frequently generates flawed queries that compile successfully but return incorrect business data.

• Database-level security dynamically strips unauthorized data from query results based on user single sign-on roles before information reaches AI agents.

• Schema annotation tools consume unstructured documents from network drives and store metadata directly inside the database to improve query generation accuracy.

• Oracle database instances integrate with the Model Context Protocol via specialized servers to manage tool execution and context windows efficiently.

타임라인

1. 00:00 - 05:07

   Oracle AI Database Capabilities and RAG Integration

   Database management tools have evolved over decades to handle various data types, now incorporating native vector support for financial, healthcare, and enterprise data. This native capability allows organizations to query structured information and vectorized content through a single interface, eliminating the need to duplicate data across multiple external platforms.

   • Oracle databases integrate vector and graph data workloads to support modern RAG pipelines within a single instance.
   • A single Model Context Protocol server connects AI agents directly to enterprise databases without data replication.
   • Database infrastructure teams typically initiate the deployment of new AI and Model Context Protocol technologies.
2. 05:07 - 11:18

   Risks of Natural Language to SQL and Report-Based Alternatives

   Business users frequently ask ad-hoc operational questions without possessing deep SQL expertise. While large language models can translate natural language into database queries, they often produce syntactically valid queries that retrieve incorrect business insights. Routing agent requests through pre-defined, parameterized reports prevents query errors and reduces unnecessary token consumption.

   • Unchecked natural language to SQL generation creates high token costs and unreliable business analytics.
   • Pre-written parameterized reports limit LLM errors by replacing dynamic SQL generation with controlled query templates.
   • Model Context Protocol servers support both custom report execution tools and raw SQL query execution modes.
3. 11:18 - 19:27

   DevOps Troubleshooting and Schema Documentation

   Database administrators can utilize read-only standby instances, such as DataGuard, to let AI agents safely diagnose performance bottlenecks and collect statistics. Furthermore, documenting legacy database schemas often relies on consuming unstructured text files and storing annotations natively, which drastically improves the precision of subsequent AI-driven queries.

   • Read-only standby database instances provide safe environments for AI agents to investigate production performance issues.
   • Schema annotation tools ingest documentation from network shares to maintain contextual metadata inside the database.
   • Database-level security controls prevent unauthorized data exposure when external tools query production environments.
4. 19:27 - 21:28

   Enterprise Security and Roadmap

   Enterprise deployments require strict access controls to ensure AI agents only retrieve information matching the user's authorization level. By tying agent requests to single sign-on identity providers like Azure Entra ID, the database dynamically applies security policies and strips restricted columns from the output before returning data to the agent framework.

   • Single sign-on integration passes organizational roles and scopes directly to the database layer during agent sessions.
   • Dynamic data security rules filter result sets automatically before information transmits back to the agent interface.
   • Oracle provides Model Context Protocol resources and documentation via oracle.com/mcp.