GES-C01 Online Practice Questions

Explanation:
To guarantee a structured JSON output for entity extraction, (the updated version of 'COMPLETE()') with the response_format’ argument and a specified JSON schema is the most effective approach. This mechanism enforces that the LLM's output strictly conforms to the predefined structure, including data types and required fields, significantly reducing the need for post-processing and improving data quality within the pipeline.
Option A requires multiple calls and manual JSON assembly, which is less efficient.
Option B relies on the LLM's 'natural ability' to generate JSON, which might not be consistently structured without explicit ‘response_format’.
Option D uses, which is for generating summaries, not structured entity extraction.
Option E involves external LLM API calls and Python UDFs, which, while possible, is less direct than using native 'AI_COMPLETE structured outputs within a SQL pipeline in Snowflake Cortex for this specific goal.

Explanation:
StatementA is incorrect. While Snowflake recommends using only ‘conda_dependencies’ or only 'pip_requirements’ (not both) to avoid package conflicts, the scenario explicitly mentions PyPl packages. If using ‘pip_requirements’, all required packages should be listed there. The example incorrectly assumes ‘torchvision’ would necessarily be best sourced from Conda and dictates avoiding 'pip_requirements’ entirely, which is an oversimplification of the recommendation. Statement B is correct. To utilize GPU acceleration in SPCS, a compute pool configured with a GPU instance family (e.g., *GPU must be created and then referenced by name in the ‘service_compute_poor’ argument when creating the service. Statement C is correct. Snowflake's warehouse nodes have restricted directory access, and '/tmpP is recommended as a safe and writeable location for models that need to write files during execution. This principle extends to SPCS containers. Statement D is correct. The ‘create_service' method for deploying models to SPCS takes a gpu_requests argument, which specifies the number of GPUs to allocate to the service. Setting this (e.g., to "s) is crucial for ensuring the model runs on GPU hardware. Statement E is incorrect. The 'relax_version’ option, which modifies version constraints, defaults to 'True' in 'log_moder’ While often beneficial, it is not mandatory to explicitly set it to ‘True' for every deployment scenario.

Explanation:
Option A is correct because it demonstrates the proper use of the 'AI_COMPLETE function with the ‘response_format’ argument to specify a JSON schema and sets ‘temperature’ to 0 for consistent output, as per the documentation.
Option C is correct as the "required'’ field in the JSON schema ensures that specific fields must be extracted, and 'COMPLETE (or 'AI_COMPLETE) will raise an error if these fields cannot be found.
Option E is correct because for the most consistent results, setting the ‘temperature* option to O is recommended when calling 'COMPLETE (or "AI_COMPLETE) with structured outputs, regardless of the task or model.
Option B is incorrect because all models supported by support structured output, and specifying the ‘response_format’ is the direct mechanism to enforce a schema, although for complex tasks, adding 'Respond in JSON' to the prompt can improve accuracy.
Option D is incorrect as 'AI_COMPLETE Structured Outputs incurs compute cost based on the number of tokens processed, but it does not incur additional compute cost for the overhead of verifying each token against the supplied JSON schema.

Explanation:
The error message 'Unable to create a build on the specified database and schema' points directly to insufficient privileges for creating the model build itself. While the database role enables working on Document AI models and using SQL for extraction, specific object privileges are required to create the model ‘build’ in a given schema. The documentation explicitly states that 'CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE on the schema is required to create model builds (instances of the 'DOCUMENT INTELLIGENCE' class), along with 'CREATE MODEL’ on the schema.
Option A (USAGE on database) is a general prerequisite but not the direct cause of this specific 'unable to create a build' error.
Option B CCREATE TABLE) is for creating tables, not model builds.
Option D (s EXECUTE TASK) is required for running tasks in a pipeline, not for model build creation.
Option E is incorrect as warehouse size primarily impacts cost and performance, not the ability to create a model build due to privileges.

Explanation:
Option B is correct because Cortex Guard evaluates the LLM's responses, not just the prompts, to filter unsafe content. Additionally, only the input tokens for Cortex Guard are counted for billing, and this cost is in addition to the function's cost.
Option E is correct because Snowflake's trust and safety principles explicitly state that your Usage and Customer Data (including inputs and outputs of Snowflake AI Features) are NOT available to other customers and are NOT used to train, re-train, or fine-tune Models made available to others.
Option A is incorrect as customer data is not anonymised and used to improve general model performance for others.
Option C is incorrect as the sources do not mention fine-tuning a customer-owned guardrail model.
Option D is incorrect as the sources do not mention a ‘CORTEX GUARD ADMIN’ role or logging of filtered content to a customer-accessible audit table.