✅ Option A (Correct) Reasoning: Snowflake Cortex Search enables users to ask questions in natural language against their enterprise data. It leverages Retrieval Augmented Generation (RAG) to find relevant information and provide direct, conversational answers without requiring users to write SQL queries. This directly addresses the question’s requirement.
❌ Why the other choices are incorrect:

Option B is incorrect: Cortex Analyst is not a recognized feature or function within Snowflake Cortex offerings. There is no specific product named ‘Cortex Analyst’ that provides direct natural language question answering.

Option C is incorrect: The Cortex CLASSIFY_TEXT function is designed for categorizing text into predefined labels, not for generating direct answers to natural language questions. Its purpose is text classification.

Option D is incorrect: The Cortex PARSE_DOCUMENT function is used for extracting structured information and entities from unstructured documents. It facilitates data extraction rather than providing conversational answers to user questions.

Reference: https://docs.snowflake.com/en/developer-guide/snowflake-cortex/search-overview

✅ Option A (Correct) Reasoning: Snowflake guarantees strict data isolation, ensuring customer data processed with Gen AI features remains within the customer’s account and is never shared with other customers or used to train public models by developers.✅ Option B (Correct) Reasoning: Snowflake’s robust Role-Based Access Control (RBAC) is foundational, extending to Gen AI features to ensure only authorized users and roles can access and utilize models and related resources.❌ Why the other choices are incorrect:

Option C is incorrect: Snowflake does not generally allow customers to share fine-tuned models with other customers as a governance principle; sharing is highly controlled and specific.

Option D is incorrect: While the Model Registry is key for model management, saying it’s the only way third-party models access accounts is too restrictive and not a universal governance guarantee.

Option E is incorrect: While Cortex Guard is a critical safety feature, Snowflake does not mandate its use for all Cortex COMPLETE functions; customers retain some control over safety implementations.

Reference: https://www.snowflake.com/en/data-cloud/solutions/gen-ai/

✅ Option A (Correct) Reasoning: The commands GRANT USAGE ON GENAI MODEL llama3.1-70b TO ROLE GENAI_DEVELOPER; and GRANT USAGE ON FUNCTION COMPLETE (TABLE(VARIANT)) TO ROLE GENAI_DEVELOPER; correctly provide the minimal required USAGE privilege for both the specific GenAI model and the COMPLETE function with its precise signature, enabling the role to use the LLM.
❌ Why the other choices are incorrect:

Option B is incorrect: It uses LLM MODEL instead of the correct GENAI MODEL object type and omits the required function signature (TABLE(VARIANT)) for the COMPLETE function.

Option C is incorrect: It incorrectly uses the EXECUTE privilege instead of the USAGE privilege, which is required for using GenAI models and functions.

Option D is incorrect: While GRANT ALL PRIVILEGES would technically enable usage, it grants broader permissions than necessary. USAGE is the principle of least privilege and the accurate grant for this scenario.

Reference: https://docs.snowflake.com/en/user-guide/ml-powered-features/genai/genai-functions-sql#granting-privileges-to-genai-functions-and-models

✅ Option C (Correct) Reasoning: The SNOWFLAKE.CORTEX_USER database role grants access to Snowflake Cortex functions. Revoking this database role from PUBLIC ensures only explicitly granted roles can use Cortex.
✅ Option D (Correct) Reasoning: Revoking USAGE on the SNOWFLAKE.CORTEX schema from PUBLIC directly prevents general access to the functions defined within that schema, thereby restricting Cortex function usage.
❌ Why the other choices are incorrect:

Option A is incorrect: SNOWFLAKE.CORTEX_FUNCTIONS is a database role. The correct syntax for revoking a database role is REVOKE DATABASE ROLE, not REVOKE ROLE.

Option B is incorrect: SNOWFLAKE.CORTEX_USER is a database role. The correct syntax for revoking a database role is REVOKE DATABASE ROLE, not REVOKE ROLE.

Option E is incorrect: REVOKE IMPORTED PRIVILEGES typically pertains to secure data sharing or objects shared into a database, not directly to restricting access to native Cortex functions.

Reference: https://docs.snowflake.com/en/user-guide/security-access-control-overview#managing-database-roles

✅ Option B (Correct) Reasoning: The VECTOR_L1_DISTANCE function in Snowflake calculates the L1 norm (Manhattan distance) between two vectors. Manhattan distance is the sum of the absolute differences of their Cartesian coordinates. It is also known as taxicab or city block distance.

❌ Why the other choices are incorrect:

Option A is incorrect: VECTOR_INNER_PRODUCT calculates the dot product of two vectors, which is not the Manhattan distance. It measures how much two vectors point in the same direction.

Option C is incorrect: VECTOR_L2_DISTANCE calculates the Euclidean distance (L2 norm) between two vectors, which is the shortest straight-line path, not the Manhattan distance.

Option D is incorrect: VECTOR_COSINE_SIMILARITY measures the cosine of the angle between two vectors, indicating their directional similarity, not a direct distance metric like Manhattan distance.

Reference: https://docs.snowflake.com/en/sql-reference/functions/vector_l1_distance

✅ Option D (Correct) Reasoning: This workflow accurately describes the RAG process using Snowflake Cortex AI. PARSE_DOCUMENT correctly extracts text from unstructured documents. EMBED_TEXT_768 then vectorizes this content for efficient storage and similarity search. Finally, vector similarity retrieves relevant data, which the COMPLETE function uses for LLM-powered analysis and generation, leveraging Snowflake Cortex AI RAG capabilities.❌ Why the other choices are incorrect:

Option A is incorrect: EMBED_TEXT_768 is for vectorization, not text extraction. SUMMARIZE requires text, not direct PDF files. TRANSLATE is for language conversion, not data retrieval in RAG.

Option B is incorrect: SENTIMENT and CLASSIFY_TEXT are for text analysis, not initial binary file processing or layout extraction. TRANSLATE does not generate embedding vectors for RAG.

Option C is incorrect: COMPLETE is for LLM generation with context, not direct summarization from URLs as an initial RAG ingestion step. COUNT_TOKENS is for token management, not a core RAG workflow step.

Reference: https://docs.snowflake.com/en/user-guide/cortex-ai/

✅ Option A (Correct) Reasoning: Serving compute represents the compute resources consumed by the search service for tasks like indexing data, running queries, and managing search indexes. This is a direct cost category tracked in the CORTEX_SEARCH_DAILY_USAGE_HISTORY view.✅ Option E (Correct) Reasoning: EMBED_TEXT_TOKENS represents the number of tokens used for generating embeddings from text data, which is a fundamental part of the Cortex Search Service and explicitly tracked in the CORTEX_SEARCH_DAILY_USAGE_HISTORY view.❌ Why the other choices are incorrect:

Option B is incorrect: Cloud services compute is a broader platform component and not a specific usage type tracked for Cortex Search Service costs.

Option C is incorrect: Cortex Search Services are serverless and do not utilize customer-provisioned virtual warehouses for their compute.

Option D is incorrect: While related to token usage, EMBED_TEXT_TOKENS is the specific and granular usage type tracked within the CORTEX_SEARCH_DAILY_USAGE_HISTORY view for token-related costs.

Reference: https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-search-service/cortex-search-service-usage

✅ Option C (Correct) Reasoning: To build a Snowflake Document AI model, the CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE privilege must be granted on the specific schema where the model will be created. Schemas are the logical containers for database objects like ML models.❌ Why the other choices are incorrect:

Option A is incorrect: Virtual warehouses provide compute resources for queries and operations but do not hold privileges for creating schema-level objects such as Document AI models.

Option B is incorrect: While schemas reside within a database, the privilege for creating specific objects within a schema is typically granted directly on the schema itself for granular control.

Option D is incorrect: Stages are used for storing data files that are loaded into or unloaded from tables. They are not where ML models are created or where creation privileges for such models are granted.

Reference: https://docs.snowflake.com/en/user-guide/ml-document-intelligence-sql#prerequisites-for-using-snowflake-document-intelligence

✅ Option A (Correct) Reasoning: Snowflake Document AI leverages Arctic-TILT, which is Snowflake’s own proprietary Text Information and Layout Transformer (TILT) model. This model is specifically designed for document understanding tasks, making it integral to Snowflake’s Document AI capabilities.❌ Why the other choices are incorrect:

Option B is incorrect: deepseek-rl is a model developed by Deepseek AI, not Snowflake.

Option C is incorrect: jamba-instruct is a model developed by AI21 Labs, not proprietary to Snowflake.

Option D is incorrect: mistral-large2 is a model from Mistral AI, not a proprietary Snowflake LLM.

Reference: https://www.snowflake.com/en/data-cloud/snowflake-ai/document-ai/

✅ Option C (Correct) Reasoning: Snowflake Cortex Search enables applications to answer questions using structured data. It performs semantic search over tables and views, identifying relevant information for Retrieval Augmented Generation (RAG) to generate accurate answers directly from data within Snowflake.

❌ Why the other choices are incorrect:

Option A is incorrect: Document AI focuses on extracting information from unstructured documents like invoices and contracts, not on integrating applications to answer questions from structured tabular data.

Option B is incorrect: "Copilot" is a general term for an AI assistant, not a specific Snowflake feature for application integration to query structured data directly.

Option D is incorrect: While Snowflake Cortex offers various analytical functions, "Cortex Analyst" is not a distinct, named feature for application integration to answer questions using structured data. Cortex Search is the specific capability for this use case.

Reference: https://docs.snowflake.com/en/developer-guide/snowflake-cortex/search

✅ Option D (Correct) Reasoning: The temperature parameter directly controls the apparent randomness and creativity of the LLM’s response. Higher values produce more diverse and unpredictable outputs, while lower values result in more deterministic and focused responses.✅ Option E (Correct) Reasoning: The top_p parameter affects the diversity of the LLM’s output by sampling from a probability distribution. It selects the smallest set of tokens whose cumulative probability exceeds the top_p value, influencing the range of possible words and thus the response’s diversity.❌ Why the other choices are incorrect:

Option A is incorrect: guardrails are used to enforce safety policies and prevent the generation of harmful or undesirable content, not to control the randomness or diversity of the response style.

Option B is incorrect: max_tokens specifies the maximum number of tokens the LLM can generate in its response. It controls the length of the output, not its randomness or diversity.

Option C is incorrect: response_format dictates the desired structure of the LLM’s output, such as JSON. It influences the format, not the inherent randomness or diversity of the generated content.

Reference: https://docs.snowflake.com/en/sql-reference/functions/snowflake.cortex.complete

✅ Option B (Correct) Reasoning: Snowflake Cortex provides features like VECTOR_EMBEDDINGS for generating vector representations and VECTOR_COSINE_SIMILARITY for comparing them. These are integral to vector search capabilities, collectively simplified by what can be termed ‘Cortex Search,’ directly reducing RAG pipeline retrieval steps.
❌ Why the other choices are incorrect:

Option A is incorrect: ‘Cortex Analyst’ is not a specific, recognized Snowflake feature for generating vector embeddings or calculating similarities.

Option C is incorrect: ‘Cortex Fine-tuning’ refers to adapting models, not directly simplifying vector embedding generation or similarity calculations for RAG.

Option D is incorrect: The ‘Cortex COMPLETE function’ is for text generation via LLMs, not for vector embedding generation or similarity calculation.

Reference: https://docs.snowflake.com/en/user-guide/snowflake-cortex/llm-functions

✅ Option B (Correct) Reasoning: The max_tokens parameter for Snowflake Cortex COMPLETE directly limits the number of tokens generated in the response. Since Cortex pricing is token-based for both input and output, reducing the output tokens directly minimizes costs associated with the function call.✅ Option E (Correct) Reasoning: Using an LLM with fewer parameters, available through Snowflake Cortex, generally results in a lower per-token cost. Selecting a smaller, more cost-effective model suitable for the task directly reduces the overall expenditure for the function invocation.❌ Why the other choices are incorrect:

Option A is incorrect: The max_tokens parameter for Snowflake Cortex COMPLETE specifically controls the maximum output tokens. To minimize input token costs, one should focus on writing more concise and efficient prompts, not by setting max_tokens for input.

Option C is incorrect: Snowflake Cortex LLM functions are serverless services managed by Snowflake. Their costs are based on token usage, not on the size or usage of user-provisioned virtual warehouses.

Option D is incorrect: LLMs with more parameters are typically larger and more complex, which generally translates to higher per-token costs due to increased computational demands, rather than minimizing them.

Reference: https://docs.snowflake.com/en/user-guide/ml-ai/llm-functions/llm-functions-overview

✅ Option C (Correct) Reasoning: Models registered in Snowflake Model Registry can be deployed as user-defined functions or stored procedures, enabling direct invocation using standard SQL commands for inference within Snowflake.

✅ Option E (Correct) Reasoning: The Snowflake Model Registry offers robust versioning capabilities, allowing users to track, manage, and deploy multiple iterations of a model while maintaining a comprehensive history of changes.

❌ Why the other choices are incorrect:

Option A is incorrect: Registered models and their versions can be updated, promoted to different stages, or deleted from the registry, providing complete lifecycle management.

Option B is incorrect: Models are typically registered directly into the Snowflake Model Registry, often leveraging internal stages for storage, without requiring prior export to external systems.

Option D is incorrect: While Snowpark for Python is commonly used for model development and registration, registered models, once deployed, can be consumed via SQL and potentially other Snowpark languages, offering broad accessibility.

Reference: https://docs.snowflake.com/en/developer-guide/snowpark-ml/how-to/manage-models-registry

✅ Option C (Correct) Reasoning: When no Verified Query Repository (VQR) is configured, Snowflake Cortex Analyst leverages its underlying Large Language Models (LLMs) to generate up to three suggested questions to assist the user.✅ Option D (Correct) Reasoning: When a Verified Query Repository (VQR) is defined, Cortex Analyst intelligently suggests up to five questions from the VQR, basing these suggestions on their semantic similarity to the user’s input.❌ Why the other choices are incorrect:

Option A is incorrect: Cortex Analyst aims to provide useful suggestions, but there is no guarantee that all suggested questions will return results. The documentation does not state this as a strict rule.

Option B is incorrect: While onboarding questions from the VQR are indeed shown when the user provides no input, this describes a display condition for a specific type of VQR question, not a general generation mechanism like LLMs or semantic similarity from user input.

Option E is incorrect: The documentation primarily highlights semantic similarity to user input and onboarding questions as VQR suggestion mechanisms. Prioritization based on most-frequently used logical columns is not explicitly stated as a primary method for suggesting questions.

Reference: https://docs.snowflake.com/en/user-guide/snowflake-cortex-analyst

✅ Option C (Correct) Reasoning: The SNOWFLAKE.CORTEX.COMPLETE SQL function can be invoked programmatically via the Snowflake SQL API, which is an exposed REST API. This allows external applications to interact with Snowflake’s Gen AI capabilities by submitting SQL queries that utilize Cortex functions.
❌ Why the other choices are incorrect:

Option A is incorrect: Copilot is a user interface assistant within Snowsight, designed for natural language interaction to generate SQL. It does not offer a direct REST API for programmatic access to Gen AI features.

Option B is incorrect: The Cortex LLM Playground is an interactive UI tool within Snowsight for experimentation with large language models. It is a visual development environment, not an exposed REST API.

Option D is incorrect: A snowflake.cortex.complete Python function is part of the Snowpark library, designed for execution within Snowflake (e.g., in a Python UDF or stored procedure). It is a programmatic interface within Snowpark, not an externally exposed REST API itself.

Reference: https://docs.snowflake.com/en/developer-guide/sql-api/index