This post is a part of the DP-800: Developing AI-Enabled Database Solutions Exam Prep Hub.
This topic falls under these sections:
Implement AI capabilities in database solutions (25–30%)
   --> Design and implement models and embeddings
      --> Evaluate external models, including multimodal, multilanguage, sizes, and structured output

Note that there are 10 practice questions (with answers) at the end of each section to help you solidify your knowledge of the material. Also, there are 4 practice tests with 30 questions each available from the hub's main page below the exam topics section.

Introduction

One of the most important responsibilities of a SQL AI Developer is selecting the appropriate AI model for a given business problem. Microsoft SQL Server, Azure SQL Database, Azure SQL Managed Instance, and Azure AI services increasingly integrate with external Large Language Models (LLMs) and embedding models to provide intelligent capabilities such as natural language querying, document summarization, semantic search, recommendation engines, and Retrieval-Augmented Generation (RAG).

Not every model is suitable for every workload. Larger models generally provide better reasoning but incur higher costs and latency. Smaller models offer faster responses and lower costs but may lack advanced reasoning capabilities. Some models support images and audio (multimodal), while others specialize in text or code. Additionally, many enterprise applications require structured outputs such as JSON rather than free-form text.

For the DP-800 exam, candidates should understand how to evaluate external models based on business requirements, performance, cost, scalability, and AI capabilities.

What Are External Models?

An external model is an AI model that runs outside the database engine and is accessed through an API or AI service.

Examples include:

• Azure OpenAI models
• Azure AI Foundry-hosted models
• Open-source models hosted on Azure AI Foundry or Kubernetes
• Other cloud-hosted foundation models exposed through REST APIs

Instead of performing AI inference inside SQL Server, the application or database calls an external service.

Example architecture:

Application
      │
      ▼
Azure SQL Database
      │
      ▼
Azure OpenAI Service
      │
      ▼
AI Model
      │
      ▼
Generated Response

This approach allows SQL-based applications to leverage continuously improving AI models without modifying the database engine.

Factors When Evaluating External Models

Several characteristics should be considered before selecting a model.

These include:

• Accuracy
• Reasoning capability
• Response quality
• Cost
• Latency
• Throughput
• Context window size
• Structured output support
• Multilingual capability
• Multimodal capability
• Security and compliance
• Availability
• Scalability

Selecting the right model is often a balance between these factors rather than maximizing any single characteristic.

Evaluating Multimodal Models

What Is a Multimodal Model?

A multimodal model can process multiple types of input rather than only text.

Common input types include:

• Text
• Images
• Documents
• Charts
• Audio
• Video (supported by some models)

Example:

A customer uploads:

• Invoice PDF
• Photograph of damaged goods
• Written description

A multimodal model can analyze all three inputs together.

Business Scenarios

Multimodal models are useful for:

• Document analysis
• Invoice processing
• Insurance claims
• Medical imaging
• Manufacturing quality inspections
• Product recognition
• OCR-enhanced workflows
• Diagram interpretation

Example:

Instead of asking:

“Describe this invoice.”

The application uploads the invoice itself.

The model extracts:

• Vendor
• Invoice number
• Total
• Purchase date
• Line items

Advantages

Multimodal models:

• Reduce preprocessing
• Improve accuracy
• Handle real-world data
• Simplify AI workflows
• Support richer user experiences

Limitations

They typically:

• Cost more
• Require more compute resources
• Have higher latency
• Process larger payloads
• May not be necessary for text-only applications

Evaluating Multilingual Models

Many enterprise applications serve users around the world.

A multilingual model understands and generates responses in multiple languages without requiring translation.

Example languages include:

• English
• Spanish
• French
• German
• Portuguese
• Japanese
• Chinese
• Korean
• Arabic

Example

Customer question:

Spanish:

¿Cuál es el estado de mi pedido?

The AI responds correctly in Spanish.

Business Benefits

Multilingual models:

• Improve customer experience
• Eliminate translation pipelines
• Simplify global deployments
• Maintain conversational context across languages
• Reduce development complexity

Evaluation Criteria

When comparing multilingual models, evaluate:

• Number of supported languages
• Translation quality
• Cultural understanding
• Domain-specific terminology
• Consistency across languages
• Response quality

Common Use Cases

• Global customer support
• International e-commerce
• Government services
• Travel applications
• Healthcare portals
• Financial institutions

Evaluating Model Size

Model size generally refers to the relative complexity and capability of an AI model. While parameter counts are not always publicly disclosed for commercial models, larger models typically provide stronger reasoning at the cost of increased compute requirements.

Generally:

Small model

• Faster
• Lower cost
• Lower latency

Large model

• Better reasoning
• Better code generation
• Better summarization
• Higher cost
• Higher latency

Small Models

Ideal for:

• Chatbots
• Classification
• Data extraction
• Intent detection
• Basic summarization

Advantages:

• Fast responses
• Low operational cost
• High throughput
• Efficient scaling

Medium Models

Good balance between:

• Performance
• Cost
• Accuracy

Typical uses:

• Customer support
• SQL generation
• Business assistants
• Document summarization

Large Models

Best for:

• Complex reasoning
• Long documents
• Advanced coding
• RAG
• Planning
• Agentic AI

Trade-offs include:

• Higher inference costs
• Greater latency
• Increased resource consumption

Latency vs. Accuracy

Every AI solution involves balancing response speed and output quality.

Example:

Customer chatbot

Acceptable latency:

2–3 seconds

Scientific research assistant

Acceptable latency:

10–20 seconds

because answer quality matters more than speed.

Trade-Off Example

Context Window Size

The context window defines how much information the model can process in a single request.

A larger context window allows the model to consider more text simultaneously.

Examples include:

• Long contracts
• Large knowledge bases
• Entire manuals
• Meeting transcripts
• Large SQL schemas

Benefits

Larger context windows reduce the need to split documents into smaller chunks and help preserve context across lengthy inputs.

Limitations

Larger contexts generally:

• Increase processing time
• Increase inference cost
• Consume more tokens

Applications should include only relevant information rather than maximizing context size unnecessarily.

Structured Output

Many enterprise applications require machine-readable responses instead of conversational text.

Example:

Instead of:

“The customer’s order total is $425 and ships tomorrow.”

Return:

{
  "customer":"John Smith",
  "orderTotal":425,
  "shipDate":"2026-07-29"
}

Structured output allows applications to parse responses reliably.

Why Structured Output Matters

Applications can:

• Deserialize JSON
• Populate SQL tables
• Call stored procedures
• Trigger workflows
• Validate data
• Build dashboards

without performing fragile text parsing.

Common Structured Formats

• JSON
• JSON arrays
• Objects
• Lists
• Tables
• XML (less common)
• Markdown tables (for presentation)

JSON remains the most common structured format for modern AI integrations.

Function Calling and Tool Use

Many modern models support function calling (also called tool calling), where the model requests that the application invoke predefined functions or APIs instead of generating all information directly.

Example workflow:

User
  │
  ▼
LLM
  │
Calls:
GetCustomerOrders()
  │
Application
  │
SQL Database
  │
Results
  │
LLM
  │
Final Answer

This approach improves accuracy by combining model reasoning with authoritative business data.

Cost Considerations

AI model selection has a direct impact on operational cost.

Factors affecting cost include:

• Model complexity
• Input tokens
• Output tokens
• Images processed
• Audio processed
• Request volume
• Concurrency
• Context window size

A higher-capability model should only be selected when its additional reasoning or multimodal features provide measurable business value.

Benchmarking Models

Before deploying an external model into production, evaluate it against representative workloads.

Typical metrics include:

• Response accuracy
• Hallucination rate
• Latency
• Cost per request
• Throughput
• Reliability
• Structured output validity
• Multilingual quality
• Safety and policy compliance

Use realistic prompts and datasets that reflect production scenarios.

Security and Responsible AI

When integrating external models with SQL-based applications:

• Protect sensitive data.
• Apply the principle of least privilege.
• Use managed identities where possible.
• Store secrets securely (for example, in Azure Key Vault).
• Validate AI-generated outputs before acting on them.
• Avoid sending unnecessary personally identifiable information (PII) to external services.
• Monitor prompts and responses for safety, quality, and compliance.

Azure OpenAI Model Selection Guidance

Although Microsoft’s available models evolve over time, the evaluation process remains consistent.

When choosing a model, consider:

• Does the workload require multimodal input?
• Is multilingual support necessary?
• What response latency is acceptable?
• How much reasoning capability is required?
• Is structured JSON output needed?
• Will the model participate in a RAG workflow?
• What are the expected request volumes?
• What is the available budget?

The best model is the one that satisfies the business requirements while meeting performance, cost, and governance objectives.

Best Practices

• Match model capability to business requirements.
• Avoid selecting the largest model unless its advanced capabilities are needed.
• Use structured outputs whenever applications consume AI responses programmatically.
• Benchmark multiple models using representative production scenarios.
• Minimize token usage to reduce costs and improve response times.
• Use multimodal models only when image, audio, or document understanding is required.
• Validate generated content before updating databases or executing business processes.
• Monitor quality, latency, and cost continuously after deployment.

DP-800 Exam Tips

Remember these key distinctions for the exam:

• Multimodal models process multiple input types, such as text and images.
• Multilingual models understand and generate content in multiple languages without requiring separate translation services.
• Smaller models typically provide lower latency and lower cost, making them suitable for high-volume, straightforward tasks.
• Larger models generally provide stronger reasoning, summarization, and code generation but require more compute resources and incur higher costs.
• Structured outputs, particularly JSON, are preferred when AI responses must be consumed by applications, APIs, or SQL processes.
• Function calling allows models to invoke trusted business logic or database operations instead of relying solely on generated responses.
• Model selection should always balance accuracy, latency, scalability, cost, security, and maintainability.

Summary

Selecting an external AI model is one of the most important architectural decisions in AI-enabled database solutions. The ideal model depends on the workload, whether that involves multilingual customer support, multimodal document analysis, structured data extraction, or advanced reasoning over enterprise data.

For the DP-800 exam, focus on understanding the trade-offs among model capabilities rather than memorizing specific model names. Be prepared to evaluate models based on multimodal support, multilingual performance, reasoning quality, latency, cost, context window size, and structured output capabilities. Equally important is understanding how these models integrate with Azure SQL and Azure AI services to build scalable, secure, and maintainable AI-enabled database solutions.

Practice Exam Questions

Question 1

You are developing an AI-enabled application that summarizes support tickets stored in Azure SQL Database. The application must support English, Spanish, French, German, and Japanese without deploying separate models for each language.

Which type of model best satisfies this requirement?

A. A monolingual English language model with prompt translation
B. A multilingual language model trained on multiple languages
C. A computer vision model with OCR capabilities
D. A speech recognition model

Correct Answer: B

Explanation:
Multilingual large language models (LLMs) are specifically trained to understand and generate text in many languages, eliminating the need to deploy separate models for each supported language. While prompt translation can work, it introduces additional latency and possible translation inaccuracies. Computer vision and speech models are not designed for multilingual text generation.

Question 2

An organization wants an AI model that can analyze scanned invoices, extract tables, understand handwritten notes, and answer user questions about the document.

Which model capability is required?

A. Structured output only
B. Text embedding generation
C. Multimodal processing
D. Sentiment analysis

Correct Answer: C

Explanation:
Multimodal models process multiple input types—including images, documents, handwritten text, and natural language—allowing them to interpret invoices and answer questions. Embedding models create vector representations but do not analyze images directly.

Question 3

You need an AI model that consistently returns data in valid JSON matching a predefined schema for direct insertion into a SQL table.

Which capability should you prioritize?

A. Long context window
B. Large parameter count
C. Function calling only
D. Structured output support

Correct Answer: D

Explanation:
Structured output capabilities ensure responses conform to predefined schemas such as JSON, reducing parsing errors and simplifying database integration. Function calling invokes external operations but does not guarantee JSON schema compliance.

Question 4

Your application performs simple product categorization and sentiment analysis on thousands of customer reviews every minute. Response time and operational cost are more important than handling complex reasoning tasks.

Which model size is the most appropriate?

A. The largest available reasoning model
B. A medium-sized multimodal model
C. A small language model optimized for classification tasks
D. A vision-language model

Correct Answer: C

Explanation:
Simple classification workloads generally do not require large reasoning models. Smaller models provide lower latency, reduced infrastructure costs, and sufficient accuracy for routine categorization and sentiment analysis.

Question 5

A financial institution evaluates several external AI models before deployment.

Which factor should receive the highest priority when handling confidential customer information?

A. Number of supported programming languages
B. Data privacy and regulatory compliance
C. Maximum context window size
D. Availability of image generation

Correct Answer: B

Explanation:
For regulated industries, protecting sensitive information and complying with regulations are primary evaluation criteria. Features such as image generation or larger context windows are secondary if the model cannot satisfy organizational security and compliance requirements.

Question 6

Your organization must choose between two external language models.

Model A produces slightly more accurate answers but averages 8 seconds per response.

Model B is slightly less accurate but consistently responds in under one second.

Which consideration is being evaluated?

A. Tokenization strategy
B. Embedding dimensions
C. Latency versus accuracy tradeoff
D. Database normalization

Correct Answer: C

Explanation:
Model evaluation frequently involves balancing response quality against latency. Interactive applications often prioritize faster responses, while analytical workloads may tolerate longer processing times for greater accuracy.

Question 7

A development team is comparing two embedding models.

One produces 768-dimensional vectors while another produces 3,072-dimensional vectors.

What is generally true?

A. Higher-dimensional embeddings always guarantee better search results.
B. Larger embeddings often improve semantic representation but require more storage and computation.
C. Embedding dimensions have no effect on vector databases.
D. Smaller embeddings always produce higher recall.

Correct Answer: B

Explanation:
Higher-dimensional vectors can capture richer semantic information but increase storage requirements, indexing costs, and similarity search computation. Larger dimensions do not automatically produce better search quality.

Question 8

A healthcare application requires AI-generated discharge summaries that follow a strict template so they can be automatically imported into Azure SQL Database.

Which model feature is most important?

A. Image generation capabilities
B. Speech synthesis support
C. Larger token limits only
D. Structured output generation

Correct Answer: D

Explanation:
Structured outputs enable AI-generated responses to consistently match required formats, such as JSON or predefined schemas, simplifying automated ingestion into databases and reducing validation errors.

Question 9

Why might an organization intentionally choose a smaller external language model instead of the newest, largest model?

A. Smaller models are always more accurate.
B. Smaller models always support more languages.
C. Smaller models often provide lower cost, reduced latency, and sufficient performance for many workloads.
D. Smaller models eliminate the need for prompt engineering.

Correct Answer: C

Explanation:
Many enterprise workloads involve straightforward tasks where the largest model offers minimal additional benefit. Smaller models frequently provide faster responses, lower inference costs, and simpler deployment while meeting performance requirements.

Question 10

An AI-enabled SQL application must process both text and uploaded product images to answer customer questions.

Which model should be recommended?

A. A multimodal language model
B. A text embedding model only
C. A relational database engine
D. A recommendation engine

Correct Answer: A

Explanation:
Multimodal models can simultaneously process textual and visual information, enabling users to ask questions about images and receive context-aware responses. Text embedding models only generate vector representations and cannot directly analyze images.

Exam Tips

For the DP-800 exam, remember these key evaluation principles when selecting external AI models:

• Select multilingual models when supporting multiple languages without translation pipelines.
• Choose multimodal models whenever applications must process images, documents, audio, or mixed media.
• Prefer structured output capabilities when AI responses must populate SQL tables or APIs reliably.
• Evaluate model size based on workload complexity, balancing cost, latency, throughput, and reasoning ability.
• Consider privacy, compliance, and data residency before selecting external AI services.
• Compare models using multiple metrics, including accuracy, latency, throughput, token limits, context window size, scalability, and operational cost.
• Remember that larger models are not always the best choice—the optimal model is the one that best satisfies the application’s functional, performance, security, and budget requirements.

Go to the DP-800 Exam Prep Hub main page