Best Production-Ready Azure Architecture for a Secure Microsoft Foundry SQL Agent with 200M+ Records

Hello Microsoft Azure and Microsoft Foundry Community,

We are developing a production-grade Natural Language to SQL Agent using Microsoft Foundry Agent Service and Azure SQL Database.

Our entire infrastructure must be hosted within Microsoft Azure.

Our key requirements are:

1. Data Security: Maximum protection using Microsoft Entra ID, managed identities, private networking, and restricted database access.

Strict Data Privacy: No confidential SQL data, financial values, customer information, or query results should ever enter or be processed by an LLM.

High Accuracy: Reliable financial reporting using predefined business rules, validated SQL queries, and deterministic calculations.

Low Cost: Cost-effective Azure infrastructure, database processing, and LLM usage.

Scalability: Two large Azure SQL tables, each expected to scale to 100 million or more records (200M+ records combined).

Incremental Data Updates: Hourly inserts, updates, and deletes across both tables.

Performance: Efficient query execution and reporting with minimal impact on database performance.

We are currently using Microsoft Foundry UI for development and testing.

Questions for Microsoft Experts

What is the Microsoft-recommended production architecture for implementing this SQL Agent with our security, privacy, accuracy, scalability, and cost requirements?

Should we use Foundry Agent Service with Azure Container Apps, a private Python API, MCP, function calling, or another Azure-native approach?

How can we use the LLM only for natural-language understanding while ensuring that confidential SQL data, calculations, and query results never enter the LLM context?

Can Microsoft Foundry UI support our strict privacy requirements, or do we need a separate secure interface to display confidential query results?

What is the most cost-effective Azure SQL architecture for handling 200M+ records with hourly incremental inserts, updates, and deletes?

What is the recommended approach to ensure accurate financial reporting while preventing unauthorized database access and incorrect SQL calculations?

We would appreciate:

Official Microsoft documentation and reference architectures.

Microsoft-recommended security and privacy best practices.

Cost-optimization and performance recommendations.

Guidance based on real-world production implementations.

If this question is better suited to another Microsoft team or forum, please direct us to the appropriate Microsoft Foundry specialist or Azure Solutions Architect.

Thank you for your guidance.Hello Microsoft Azure and Microsoft Foundry Community,

We are developing a production-grade Natural Language to SQL Agent using Microsoft Foundry Agent Service and Azure SQL Database.

Our entire infrastructure must be hosted within Microsoft Azure.

Our key requirements are:

Data Security: Maximum protection using Microsoft Entra ID, managed identities, private networking, and restricted database access.

Strict Data Privacy: No confidential SQL data, financial values, customer information, or query results should ever enter or be processed by an LLM.

High Accuracy: Reliable financial reporting using predefined business rules, validated SQL queries, and deterministic calculations.

Low Cost: Cost-effective Azure infrastructure, database processing, and LLM usage.

Scalability: Two large Azure SQL tables, each expected to scale to 100 million or more records (200M+ records combined).

Incremental Data Updates: Hourly inserts, updates, and deletes across both tables.

Performance: Efficient query execution and reporting with minimal impact on database performance.

We are currently using Microsoft Foundry UI for development and testing.

Questions for Microsoft Experts

What is the Microsoft-recommended production architecture for implementing this SQL Agent with our security, privacy, accuracy, scalability, and cost requirements?

Should we use Foundry Agent Service with Azure Container Apps, a private Python API, MCP, function calling, or another Azure-native approach?

How can we use the LLM only for natural-language understanding while ensuring that confidential SQL data, calculations, and query results never enter the LLM context?

Can Microsoft Foundry UI support our strict privacy requirements, or do we need a separate secure interface to display confidential query results?

What is the most cost-effective Azure SQL architecture for handling 200M+ records with hourly incremental inserts, updates, and deletes?

What is the recommended approach to ensure accurate financial reporting while preventing unauthorized database access and incorrect SQL calculations?

We would appreciate:

Official Microsoft documentation and reference architectures.

Microsoft-recommended security and privacy best practices.

Cost-optimization and performance recommendations.

Guidance based on real-world production implementations.

If this question is better suited to another Microsoft team or forum, please direct us to the appropriate Microsoft Foundry specialist or Azure Solutions Architect.

Thank you for your guidance.