Microsoft Certified: Azure AI Engineer Associate flashcards

VNet Service Endpoints extend your virtual network's private address space to Azure services, allowing private access to Azure service resources from your VNet.

• Secures Azure service resources to your virtual network.
• Traffic travels over the Azure backbone network, not the public internet.
• Helps enforce data residency and network isolation.

Memory trick: Region-specific endpoints keep data home and secure.

Azure offers various encryption options, including Microsoft-managed keys (default), customer-managed keys (CMK) for data at rest, and Transport Layer Security (TLS) for data in transit.

• CMK gives customers full control over encryption keys.
• CMK keys are stored in Azure Key Vault.
• Compliance requirements often mandate CMK.

Memory trick: CMK gives you the 'Key' to your data's 'Lock' at rest.

Azure Private Link (via Private Endpoints) enables private, backbone-only access to Azure AI services from within a VNet. Network Security Groups (NSGs) then provide granular filtering of traffic within the VNet to control which resources can access these Private Endpoints.

• Private Link ensures traffic stays on Azure backbone.
• Private Link provides a private IP for the service in your VNet.
• NSGs filter traffic at the subnet or NIC level.
• Combined, they offer strong network isolation and access control.

Memory trick: Private Link is the private 'road', NSGs are the 'checkpoints' on that road.

A family of Azure OpenAI models specifically trained on a vast dataset of code and natural language, optimized for understanding and generating programming code.

• Excels at code generation, completion, and explanation.
• Supports multiple programming languages.
• Can be used for code refactoring and bug fixing.

Memory trick: Codex is the 'Code Expert' in the Azure OpenAI family.

Specialized models within Azure OpenAI Service designed to detect and categorize harmful content in text and images, helping developers build safer AI applications.

• Pre-trained for various harm categories (e.g., hate speech, violence).
• Provides confidence scores for detected categories.
• Crucial for responsible AI implementation in user-facing applications.

Memory trick: For 'Content Cops', you need the 'Moderation Force' models.

A parameter in Azure OpenAI Service that penalizes new tokens based on whether they appear in the text so far, reducing repetition.

• Discourages the model from repeating words/phrases.
• Higher values lead to more diverse, less repetitive text.
• Applied after frequency penalty.

Memory trick: TRP: To Really Produce diverse text, adjust your parameters.

The GPT-3.5 and GPT-4 model families are advanced large language models (LLMs) from OpenAI, optimized for natural language understanding and generation, making them highly effective for conversational AI applications like chatbots.

• Excels at human-like text generation.
• Capable of understanding context and generating coherent responses.
• Suitable for various NLP tasks, including summarization, translation, and dialogue.

Memory trick: GPT-4 is the 'General Purpose Talker' for your chatbot needs.

A prompt engineering technique where a few examples of the desired input-output behavior are provided to the model.

• Demonstrates desired patterns and formats.
• Improves model performance on specific tasks without fine-tuning.
• Effective for guiding tone, style, and content inclusion.

Memory trick: Few Examples Lead to Consistent Keywords: Give a 'few' good examples to hit your 'keywords'.

The 'temperature' parameter in Azure OpenAI Service controls the randomness and creativity of the generated text. A higher value leads to more diverse and surprising outputs, while a lower value results in more deterministic and focused outputs.

• Ranges typically from 0 to 2.
• Higher values (e.g., 0.7-1.0) increase creativity.
• Lower values (e.g., 0.2-0.5) increase determinism and focus.

Memory trick: Temperature 'heats up' or 'cools down' the AI's creativity.

A special message in the chat completions API that sets the behavior, persona, and overall instructions for the model.

• Sent once at the beginning of a conversation.
• Influences the model's tone, style, and general approach.
• Not visible to the end-user in typical chat interfaces.

Memory trick: Systematic Persona Crafting: Use System messages to craft a consistent personality.

A prompt engineering technique that explicitly limits the model's output to a predefined set of choices, format, or structure.

• Crucial for categorization, structured data extraction, and function calling.
• Often involves listing valid options or defining a strict schema.
• Reduces hallucinations and ensures parseable, predictable output.

Memory trick: Instruct with List, Lock the Output: Give clear 'instructions with a list' to 'lock' the output.

A prompt engineering technique that guides a large language model to generate output in a specific, machine-readable format, such as JSON or XML.

• Often uses few-shot examples to demonstrate the desired structure.
• Crucial for integrating LLMs with downstream systems and databases.
• Ensures consistency and parseability of generated data.

Memory trick: Few JSON Examples Structure Success: Give a 'few' JSON examples for 'structured' results.

A prompt engineering technique that combines explicit instructions with a few illustrative examples to guide an LLM to produce outputs that conform to specific formatting, style, or content requirements.

• Instructions define the rules, examples show application.
• Highly effective for complex or nuanced output constraints.
• Balances direct guidance with practical demonstrations.

Memory trick: Instructions and 'Few-Shot' examples are like a 'Recipe with Pictures' for AI style.

A prompt engineering strategy that combines explicit instructions for brevity and factual adherence with low-randomness generation parameters.

• Uses low 'temperature' or 'top_p' for deterministic output.
• Includes direct instructions like 'be concise', 'focus only on data', 'no fluff'.
• Ideal for factual summaries, data interpretation, and report generation.

Memory trick: Low Temp + Clear Instructions = Concise Data Power: Keep the 'temperature low' and 'instructions clear' for 'powerful, concise data'.

A feature within Azure Machine Learning that provides a comprehensive view and tools for assessing and mitigating responsible AI issues.

• Helps identify fairness issues, model errors, and interpretability gaps.
• Supports various AI models, including generative AI.
• Offers visualizations and mitigation techniques for responsible deployment.

Memory trick: AML Dashboard Guards AI Ethics: The 'AML Dashboard' is your 'guard' for 'AI ethics'.

Azure Monitor is a comprehensive monitoring solution that collects and analyzes telemetry from Azure resources, enabling continuous observation of AI model performance, inputs, and outputs to ensure responsible and ethical operation.

• Collects logs and metrics.
• Provides dashboards and alerting capabilities.
• Essential for detecting bias, drift, or harmful content over time.

Memory trick: Azure Monitor is the 'AI Watchdog' for your deployed models.

A technique for managing long conversation histories with LLMs by only including the most recent 'N' tokens or turns in the prompt.

• Addresses the LLM's fixed context window limitation.
• Keeps the conversation relevant to recent interactions.
• Can be combined with other techniques like summarization for critical information.

Memory trick: Slide the Window for Long Story Memory: 'Sliding window' helps your story's 'memory' last longer.

A parameter in Azure OpenAI Service that penalizes new tokens based on their frequency in the generated text so far, reducing the likelihood of the model repeating the same words or phrases.

• Values typically range from -2.0 to 2.0.
• Positive values discourage repetition.
• Helps generate more diverse and original text.

Memory trick: Frequency Penalty 'punishes' boring repetition, sparking new words.

Ensuring AI systems perform as intended, are robust against manipulation, and do not cause unintended harm or generate unsafe content.

• Crucial for high-stakes applications (e.g., finance, healthcare).
• Involves preventing inaccurate, biased, or harmful outputs.
• Includes robustness, security, and controlled content generation.

Memory trick: FFSAAT: For Future Safe AI, Always Act Transparently.

A prompt engineering strategy that involves providing specific source material directly within the prompt and explicitly instructing the LLM to generate responses *only* based on that provided context, thereby reducing the risk of 'hallucinations' or fabricated information.

• Crucial for factual accuracy in sensitive domains.
• Combines providing context with strong constraints.
• Limits the model's reliance on its general pre-trained knowledge.

Memory trick: Grounding is like giving the AI a 'Fact-Check Rulebook' for its sources.

A prompt engineering technique where explicit, detailed instructions are provided to guide the large language model (LLM) to produce a desired output format, content, or style.

• Involves clear, direct commands within the prompt.
• Effective for enforcing specific output constraints.
• Reduces ambiguity and improves consistency of results.

Memory trick: Instructions are like a 'GPS for the GPT' guiding its output precisely.

A parameter that controls the randomness of the output generated by the model. Higher values make the output more creative and diverse.

• Ranges from 0 to 2.
• Lower values (e.g., 0.2) make output more deterministic and focused.
• Higher values (e.g., 0.8-1.0+) make output more diverse, creative, and unpredictable.

Memory trick: Turn Up the Temp for Creative Sparks: Higher 'temperature' ignites 'creative' ideas.

The principle of Responsible AI that dictates AI systems should treat all individuals and groups equitably, without causing or reinforcing unfair bias or discrimination.

• Aims to prevent disparate impacts on different demographic groups.
• Requires careful evaluation of training data for biases.
• Crucial in sensitive domains like finance, healthcare, and hiring.

Memory trick: Fairness is about giving everyone a 'Fair Shot' with AI, especially in finance.

The process of removing or encrypting personally identifiable information (PII) from data before it is used as input for large language models.

• Crucial for protecting privacy and complying with regulations (e.g., HIPAA, GDPR).
• Prevents models from inadvertently generating or retaining PII.
• A proactive measure to mitigate data leakage risks.

Memory trick: Anonymize All Inputs, No PII Leaks: Make sure your inputs are anonymous to prevent any PII leaks.

A capability in Azure OpenAI Service that allows users to further train a pre-existing base model on their own proprietary dataset to specialize it for a particular task, domain, or style, leading to improved performance.

• Adapts models to specific use cases.
• Requires a high-quality, task-specific dataset.
• Can significantly enhance model accuracy and relevance.

Memory trick: Fine-tuning is like 'tailoring the AI' to fit your exact needs.

A parameter in Azure OpenAI Service that reduces the likelihood of the model repeating tokens that have already appeared in the generated text, making the output less repetitive.

• Penalizes tokens based on how many times they have already appeared.
• Higher values decrease repetition.
• Useful for generating more diverse and less redundant text.

Memory trick: Fine-tune output, avoid repeated words.

A technique to handle documents exceeding an LLM's token limit by breaking them into smaller, overlapping chunks, summarizing or processing each chunk, and then combining or recursively summarizing the results.

• Prevents information loss due to LLM context window constraints.
• Enables processing of arbitrarily long texts.
• Can involve 'map-reduce' or hierarchical summarization patterns.

Memory trick: Long text, chop it, map it, then reduce it to a summary.

A cloud-based Azure AI service that uses machine learning to extract text, key-value pairs, and structured data from documents.

• Handles various document types (forms, invoices, receipts, general documents).
• Supports both printed and handwritten text.
• Can extract data from complex tables and structured layouts.

Memory trick: Document Intelligence is like a smart librarian for your forms, finding exactly what you need.

A feature in Azure AI Search that enables type-ahead query suggestions, helping users discover relevant content more easily.

• Configured on specific fields in the index.
• Improves user experience by guiding search queries.
• Can be based on indexed content or query history.

Memory trick: To 'suggest' good ideas, you need a 'suggester' to whisper hints as you type.

Azure AI Document Intelligence provides advanced OCR capabilities to extract text from scanned documents, including image-based PDFs, making their content accessible for further processing.

• Handles both printed and handwritten text.
• Preserves document layout and reading order.
• Integrated into comprehensive document processing pipelines.

Memory trick: To 'read' scanned documents, you need a 'Document Intelligence' brain that can 'see' the text.

A feature in Azure AI Search that defines sets of equivalent terms, allowing queries for one term to automatically match documents containing its synonyms.

• Improves recall by expanding search queries.
• Can be global or specific to an index.
• Supports one-way or two-way synonym relationships.

Memory trick: To make your search 'broader', use a 'synonym map' to link all related words.

A Document Intelligence model trained by users on their own documents to extract specific data fields, tables, or structures unique to their business needs.

• Requires labeled training data (typically 5+ documents).
• Ideal for specialized document types or unique data extraction.
• Provides high accuracy for user-defined fields.

Memory trick: For unique documents, you need a 'custom model' to 'teach' the AI what to look for.

A configuration parameter for Azure AI Search indexers that specifies the maximum number of pages to extract from multi-page documents like PDFs.

• Default value is often low (e.g., 5 pages) to prevent excessive processing.
• Needs to be increased for large documents to ensure full content extraction.
• Configured within the `parameters` section of the indexer definition.

Memory trick: To process many pages, increase the max page count.

Built-in capabilities of Azure AI Document Intelligence models to prepare documents for optimal extraction, such as correcting orientation, enhancing image quality, and detecting language.

• Reduces the need for manual image clean-up.
• Improves accuracy of text and structure extraction.
• Includes automatic rotation, deskewing, and perspective correction.

Memory trick: Before the AI reads, it 'rotates' and 'straightens' the document, like a smart scanning assistant.

An Azure AI Document Intelligence model designed for high-accuracy optical character recognition (OCR) across various document types, including printed text, handwritten text, and symbols.

• Optimized for general document OCR.
• Excels at extracting handwritten text.
• Provides text, lines, words, locations, and confidence scores.

Memory trick: To read handwritten notes, use the Read model.

Specialized text analyzers in Azure AI Search that perform language-specific processing (e.g., stemming, tokenization, stop word removal) for both indexing and querying, improving relevance for multilingual content.

• Applied to `searchable` fields.
• Configured using `analyzer` for indexing and `searchAnalyzer` for querying.
• Supports numerous languages (e.g., `es.microsoft`, `de.microsoft`).
• Crucial for accurate multilingual search.

Memory trick: Analyze languages for better search results.

An Azure AI Search skill that identifies, categorizes, and optionally redacts personally identifiable information (PII) from unstructured text.

• Detects various types of PII (e.g., names, credit card numbers, phone numbers).
• Can be configured to redact or remove identified PII.
• Essential for privacy compliance and data security in search solutions.

Memory trick: PII is private, so PII Detection protects it.

A feature in Azure AI Search that customizes the relevance of search results by applying weights to fields, boosting terms, or using functions to influence the scoring algorithm.

• Allows you to define how search results are ranked.
• Can boost results based on field importance, recency, or other criteria.
• Uses a JSON structure to define the profile's rules.

Memory trick: Scoring profiles help you score better results.

A mechanism in Azure AI Search to implement document-level security by filtering search results based on user permissions or roles.

• Requires indexing security identifiers (SIDs) or roles with each document.
• Query time filters are applied based on the authenticated user's identity.
• Ensures users only see documents they are authorized to access.

Memory trick: To keep documents 'secure', you need 'security filters' like bouncers at a club, checking everyone's 'access list'.

An Azure AI Search built-in skill that extracts printed or handwritten text from images within documents.

• Part of an Azure AI Search skillset.
• Processes image content (e.g., within PDFs, JPEGs).
• Outputs extracted text which can then be indexed.

Memory trick: To search images, you need a 'skill' to 'see' the text inside them, like a super-powered OCR eye.

Properties applied to fields within an Azure AI Search index that determine how the field can be used in search operations.

• Common attributes include 'searchable', 'filterable', 'sortable', 'facetable', 'retrievable'.
• These attributes must be defined during index creation or update.
• They enable advanced features like faceted navigation and custom scoring profiles.

Memory trick: An index is like a book's table of contents, and field attributes are like special tags for each entry.

An Azure AI Search custom skill, typically implemented as an Azure Function or Web API, used to modify document content (e.g., redact PII) during the indexing pipeline.

• Allows for arbitrary code execution.
• Can be chained with other built-in or custom skills.
• Essential for data transformation or redaction not covered by built-in skills.

Memory trick: To hide secrets, you need a 'custom function' to 'scrub' the data clean before anyone searches.