Select Two (Multi-Select) Questions

Recall

Included in evaluation.

Precision

Included in evaluation.

tags

Tags can be combined with other features in a single request.

read

Read OCR can be combined with tags or objects in a single analysis call.

The choice of embedding model (e.g., text-embedding-3-large vs text-embedding-3-small) and its dimensionality.

The embedding model determines how accurately semantic meaning is captured in vector space.

The chunking strategy and chunk size used to segment source documents.

Chunking strategy determines the granularity and contextual coherence of retrieved passages.

Configure Azure Private Endpoints to restrict access to your Virtual Network.

Private endpoints keep traffic off the public internet.

Enable diagnostic logging and stream audit records to Log Analytics.

Diagnostic logging ensures traceability and auditing.

Implement Microsoft Entra ID role-based access control (RBAC) instead of relying solely on API keys.

Entra ID RBAC provides fine-grained identity-based access control.

Resize extremely large images to optimal dimensions while maintaining readability.

Reducing massive file sizes speeds up network transfer and processing.

Implement proper polling intervals (e.g., waiting 1 second between checks) rather than aggressive busy-waiting.

Proper polling intervals reduce unnecessary HTTP calls and rate limit risk.

Co-locate the application hosting and the Azure AI Vision resource in the same Azure region.

Co-location minimizes network round-trip time.

text

The text property holds the string content of the detected line or word.

lines

The lines array contains extracted text lines.

JPEG

JPEG is fully supported by Azure AI Vision.

PNG

PNG is fully supported by Azure AI Vision.

Use strict structural delimiters (like XML tags) to isolate retrieved context from system instructions

XML tags help the model distinguish untrusted retrieved data from trusted system instructions.

Implement output validation filters to check model completions for unauthorized actions or exfiltration attempts

Output validation catches any successful injection attempts before responses reach users.

Enable Prompt Shields for indirect prompt injection in Azure AI Content Safety

Prompt Shields detect malicious instructions hidden within retrieved web pages or documents.

Extraction of key-value pairs and fields defined during training

Custom Neural models extract user-defined fields and key-value pairs based on training labels.

Extraction of tables from documents with variable layouts

Custom Neural models support table extraction across variable layouts.

Reflection pattern where the agent critiques its own intermediate outputs and refines them.

Self-reflection allows agents to catch errors and improve output quality iteratively.

ReAct (Reason and Act) pattern combining thought generation with tool execution steps.

The ReAct pattern alternates between reasoning about a problem and acting via tools.

Plan-and-Solve prompt strategy where the agent creates a complete execution plan before stepping through tasks.

Plan-and-solve structures multi-step problem solving effectively.

Run automated evaluations in Azure AI Foundry to measure metrics such as groundedness, relevance, and protected material presence.

Automated evaluation tools in Azure AI Foundry help quantify model performance and safety metrics against validation datasets.

Perform adversarial testing (red teaming) to identify potential prompt injections, jailbreaks, and unintended behaviors.

Red teaming uncovers vulnerabilities and edge cases before the application is released to production users.

Specify the desired output format explicitly (e.g., JSON schema or markdown tables).

Explicit formatting instructions ensure parsable and structured outputs.

Use delimiters such as triple backticks or XML tags to clearly separate instructions from user-provided data.

Delimiters prevent user input from being misinterpreted as system instructions.

Incorporate few-shot examples of edge cases and expected behaviors.

Few-shot examples significantly improve performance on specialized tasks.

Few-shot prompting providing high-quality input-output demonstration examples

Demonstration examples set clear expectations for format, tone, and accuracy.

Chain-of-thought (CoT) prompting instructing step-by-step reasoning

CoT prompts guide the model to break down complex problems into manageable logical steps.

Azure OpenAI Batch API for asynchronous high-volume jobs at reduced costs

The Batch API processes large volumes of asynchronous requests with dedicated quota and cost benefits.

Provisioned Throughput Units (PTUs) for guaranteed model capacity and predictable latency

PTUs reserve dedicated model capacity, ensuring predictable performance and high throughput.

Embedding model dimension size and consistency with source generation

Query and document vectors must use the exact same embedding model and dimensionality.

HNSW (Hierarchical Navigable Small World) algorithm parameters such as m and efSearch

HNSW index parameters balance search speed and recall accuracy.

Choice of vector similarity metric (Cosine, Dot Product, or Euclidean)

The similarity metric determines how vector proximity is calculated during search.

Configure system messages to explicitly define boundaries

System messages are the first line of defense in guiding AI behavior safely.

Enable Azure AI Content Safety filtering

Content filtering is a core pillar of responsible AI in Azure.

Implement manual review cycles for high-impact outputs

Human-in-the-loop is critical for high-stakes AI decisions.

Tool use and external API invocation capabilities

Agents dynamically select and execute tools based on intermediate reasoning.

Iterative reasoning and self-reflection loops

Agents evaluate their progress and correct errors across multiple execution turns.

Autonomous planning and task decomposition

Agents can break down high-level goals into sequential steps and sub-tasks.

Request duration (latency) in milliseconds

Latency metrics help monitor model responsiveness and performance SLAs.

Prompt tokens and completion tokens consumed per request

Token counts track resource consumption and quota utilization.

Evaluate whether document chunk sizes and overlaps align with query granularity

Poor chunking sizes can split vital context or bundle unrelated topics, ruining retrieval precision.

Inspect embedding vector distances and similarity scores of retrieved documents

Checking similarity scores helps determine if embeddings accurately represent query-document semantic closeness.

Enable semantic ranking in Azure AI Search to re-rank results based on deep language understanding

Semantic ranking significantly improves search precision by re-ranking initial results using advanced AI models.

Use Managed Identities with Azure RBAC for service-to-service authentication

Managed identities eliminate hardcoded credentials and enforce role-based access control.

Configure Azure Private Endpoints for Azure OpenAI and Azure AI Search

Private endpoints secure data traffic within the virtual network, preventing public internet exposure.

Implement Customer-Managed Keys (CMK) via Azure Key Vault for data encryption at rest

CMK gives organizations full control over encryption keys used to protect indexes and model deployments.