Generative Artificial Intelligence

Azure OpenAI Service provides access to advanced generative AI models developed by OpenAI.

Available Models:
- GPT (Generative Pre-trained Transformer): Language models for text generation, summarization, and chat
- DALL-E: Models for image generation from textual descriptions
- Codex: Models specialized in code generation and understanding
- Whisper: Models for audio transcription and translation

GPT Series:
- GPT-3.5: Versatile model for chat, writing, and analysis
- GPT-4: More advanced model with better reasoning and multimodal capabilities
- GPT-4 Turbo: Optimized version with higher speed and efficiency

Model Features:
- Scalability: Can handle tasks of different complexities
- Multilingual: Support for multiple languages
- Contextual: Maintains context in long conversations
- Customizable: Fine-tuning for specific use cases

Azure OpenAI Service is Microsoft's platform for securely accessing and managing OpenAI models.

Azure Studio:
- Web interface for interacting with OpenAI models
- Tools for experimentation and prototyping
- Resource and cost management

Playgrounds:
- Chat Playground: For conversations and chatbots
- Completions Playground: For text generation
- DALL-E Playground: For image generation
- Code Playground: For code work

Basic Parameters:
- Temperature: Controls creativity (0-2, higher = more creative)
- Top P: Controls response diversity (0-1, higher = more diverse)
- Max Tokens: Limits response length
- Stop Sequences: Defines when to stop generation

Security and Compliance:
- Data encryption in transit and at rest
- Compliance with privacy standards
- Content filters for responsible use
- Usage monitoring and logging

Prompt engineering is the art of designing effective instructions to get better responses from generative AI models.

Basic Techniques:
- Be specific: Give clear and detailed instructions
- Provide context: Include relevant background information
- Define format: Specify how the response should be structured
- Use examples: Show examples of desired input and output

Structure of a Good Prompt:
1. Role: Define the role the model should assume
2. Task: Clearly describe what it should do
3. Context: Provide necessary background information
4. Instructions: Give specific guidelines on format and style
5. Examples: Include examples when appropriate

Types of Prompts:
- Closed prompts: For specific and structured responses
- Open prompts: For creative and exploratory generation
- Few-shot prompts: With examples to guide behavior
- Chain-of-thought prompts: For step-by-step reasoning

Copilots are AI assistants integrated into Microsoft applications that help users be more productive.

Copilot Concept:
- Intelligent assistants that work alongside the user
- Integrated into familiar applications like Word, Excel, Teams
- Use generative AI to help with complex tasks
- Learn from user's context and behavior

Examples of Microsoft Copilots:
- GitHub Copilot: Programming assistant in VS Code
- Microsoft Copilot: General assistant in Windows and Edge
- Copilot in Office: Help in Word, Excel, PowerPoint
- Copilot in Teams: Improves meetings and collaboration

Features:
- Contextual: Understands the context of the current task
- Proactive: Suggests actions before they are requested
- Customizable: Adapts to user preferences
- Secure: Respects data privacy and security

Benefits:
- Increases productivity by automating repetitive tasks
- Reduces learning curve of complex tools
- Provides real-time expert assistance
- Improves creativity and problem-solving

Generative AI poses unique responsibility challenges that require specific security and ethical measures.

Content Filters:
- Jailbreak prevention: Blocks attempts to manipulate the model
- Content moderation: Filters harmful or inappropriate content
- Factual verification: Reduces hallucinations and incorrect information

Specific GenAI Challenges:
- Hallucinations: Generation of false but convincing information
- Amplified biases: Biases in training data propagate
- Disinformation: Ability to generate deceptive content at scale
- Deepfake abuse: Creation of realistic fake content

Mitigation Measures:
- Responsible fine-tuning: Additional training to reduce biases
- Continuous monitoring: Real-time detection of misuse
- Transparency: Documentation of limitations and capabilities
- Access controls: Restriction of use for high-risk cases
- User education: Awareness of AI limitations

Microsoft Framework:
- Six AI Principles: Applied specifically to generative models
- Content Safety: Automatic filters and human moderation
- Fairness in AI: Regular bias audits
- Accountability: Usage tracking and clear responsibility

One of the most important capabilities in the exam is the ability to connect language models (like GPT-4) to your own enterprise data without needing to retrain the model.

Concept (RAG - Retrieval Augmented Generation):
1. Ingestion: Azure indexes your documents (PDFs, Word, HTML) using Azure AI Search.
2. Retrieval: When a user asks a question, the system searches for relevant information in your documents.
3. Generation: Sends that information to the GPT model as 'context' to generate an answer based *only* on your data.

Benefits:
- Accurate answers: Prevents the model from making up information (hallucinations).
- Up-to-date data: Does not depend on the model's training cut-off date.
- Security: Private data is not used to train the base public model.
- Citations: The model can cite which document it used to answer.

To understand how these models work and are billed, it is necessary to know two fundamental units.

Tokens (The unit of text):
- Models do not read whole words, they read 'tokens' (character fragments).
- Rule of thumb: 1,000 tokens ≈ 750 words.
- Importance: Azure OpenAI bills for every 1,000 tokens processed (input + output). There are also token limits per request (Context Window).

Embeddings (The unit of meaning):
- A special data representation format that converts text into numerical vectors (lists of numbers).
- Usage: Allows computers to understand the semantic relationship between words.
- Example: In a vector space, the word 'Queen' will be mathematically close to 'King' and 'Woman'. Used for intelligent searches.