Snowflake GES-C01 Exam Prep
SnowPro Specialty Gen AI (Page 10 )

Explanation:

The Snowflake Cortex LLM REST API follows the same authentication model as Snowflake’s native APIs and supports OAuth 2.0 and key-pair (asymmetric) authentication .
OAuth 2.0 is the recommended method for user-driven access; it allows a client to obtain a short-lived access token by exchanging client-id/secret or via the JWT-bearer flow, which can be scoped to the CORTEX role.
Key-pair authentication (asymmetric JWT signing with a private key) enables service-to-service or automated scripts to call the endpoint without interactive user interaction, and it is also listed as a supported method in the Snowflake documentation.
The other options are not applicable:
SAML 2.0 and SSO are enterprise-wide identity mechanisms used for web-based applications, but they are not directly exposed as authentication schemes for the Cortex LLM REST endpoints. Multi-Factor Authentication (MFA) is a security attribute enforced by Snowflake’s security policies, but it is not an API authentication protocol itself; the API does not issue separate MFA credentials.
Therefore, the two supported methods are OAuth 2.0 and Key-pair authentication (options B and D ).

Reference:

Snowflake Cortex LLM REST API Authentication Overview – Snowflake Documentation https://docs.snowflake.com/en/developer-guide/snowflake-cortex/rest-api-auth
OAuth 2.0 and Key-Pair Authentication for Snowflake Cortex – Snowflake Community https://community.snowflake.com/s/article/OAuth-2-0-and-Key-Pair-Authentication-for-Cortex-LLM-REST-API

Explanation:

Explanation:

Why BCE is the right choice
B – Streams and tasks – Streams capture new files as they land in the stage, and tasks schedule the stored-procedure workflow (download  →  load  →  transform) on an hourly cadence, ensuring a fully automated, repeatable pipeline. C – External access integration – Allows a stored procedure to invoke an HTTP endpoint directly from Snowflake, enabling the “download PDFs from a website once an hour” step without moving data out of Snowflake. E – Cortex TRY_COMPLETE – Generates natural-language output based on a prompt that includes the extracted text and a persona instruction; this is the only Cortex function that can produce custom opinions, making it essential for the persona-driven opinion step.
Why the other options are not suitable
A – Snowpipe – Designed for continuous bulk ingestion from external stages; here the ingestion is driven by a scheduled HTTP fetch and can be handled by tasks, so Snowpipe adds unnecessary complexity. D – Cortex SENTIMENT – Provides only sentiment scores, which are not tailored to a persona-based opinion;
the requirement calls for a generative response, best served by TRY_COMPLETE. F – Dynamic tables using incremental refreshes – Dynamic tables are useful for materialized views that change frequently, but the pipeline described is orchestrated through stored procedures and tasks; dynamic tables are not required to meet the hourly automation and opinion-generation goals.

Reference:

1. Snowflake Documentation – Creating and Using Streams and Tasks https://docs.snowflake.com/en/user-guide/data-streams-tasks
2. Snowflake Documentation – External Functions and External Access Integrations https://docs.snowflake.com/en/sql-reference/external-function
3. Snowflake Documentation – Cortex: TRY_COMPLETE
https://docs.snowflake.com/en/cortex/cortex-try_complete
4. Snowflake Documentation – Dynamic Tables Overview https://docs.snowflake.com/en/sql-reference/sql/create-dynamic-table

Explanation:

Explanation:

Technical justification model (A) – The model argument supplies the actual trained artefact (e.g., a PyTorch or TensorFlow object) that
Snowflake needs to store in the Model Registry. Without providing the model object the function has nothing to register, so this argument is mandatory. model_name (B) – The model_name argument uniquely identifies the target model within the registry namespace. Snowflake requires a name to create or update a model entry; omitting it would leave the registry without a reference point, making this argument mandatory.
The remaining arguments are optional:
version_name (D) can be omitted; if not supplied Snowflake auto-generates a sequential version.
code_paths (C) are only needed when you want to register external source files alongside the model, which is not required for basic logging. python_version (E) merely informs Snowflake of the runtime version and does not affect the registration process, so it is not required.
Thus, the only required arguments are A and B .

Reference:

Snowflake Documentation – Log a model in Snowflake Model Registry: https://docs.snowflake.com/en/developer-guide/snowpark-python/machine-learning/log-model Snowflake Documentation – Snowpark Python API Reference – log_model: https://docs.snowflake.com/en/developer-guide/snowpark-python/reference/ml_functions#log_model

Explanation:

Technical Justification
Deploying a smaller, task-specific variant of the base model through Snowflake Cortex Fine-tuning reduces the model’s parameter count and computational footprint, directly lowering cost per inference while preserving accuracy on the narrowly defined proof-of-concept workloads. Tailoring the model to the specific semantics of the target tasks eliminates unnecessary processing of generic knowledge, which means fewer tokens are needed for inference and storage costs drop accordingly. Approaches that merely restrict context length or rely on prompting tricks do not alter the underlying resource consumption of the original large model; they only affect input size and therefore cannot achieve the same level of cost reduction. Introducing a semantic layer or using conversational agents may improve interaction quality but still executes the same high-capacity model under the hood, so operational expenses remain unchanged. Consequently, fine-tuning the model to a more compact, domain-relevant version is the only technique that simultaneously cuts production costs and maintains the required performance level.

Reference:

Snowflake Cortex Overview: https://docs.snowflake.com/en/snowflake-cortex Fine-tuning Models in Snowflake Cortex: https://docs.snowflake.com/en/snowflake-cortex/finetuning

Explanation:

Technical justification
A – Compute pool is mandatory because Snowpark Container Services (S-CS) runs the container on a Snowflake compute pool (i.e., a set of Snowflake virtual warehouses). The pool provides the compute resources and scaling behavior required for the container to start, execute, and be monitored. E – Service specification is required to define the runtime behavior of the container service. It contains the image location, startup command, exposed ports, resource limits, and other configuration options that tell Snowflake how to launch and manage the container.
These two items (A and E) are the core building blocks needed to deploy and run a container on S-CS. Without a pool you have no compute to host the container; without a service spec Snowflake does not know how to treat the container image or what expectations to enforce.
Why the other options are not required for deployment/runtime
B – External stage linked to Amazon S3 is only needed when you load data from external files; it is not part of the container-deployment workflow in S-CS. C – Schedule task assigned to a virtual warehouse is related to scheduled data loading jobs, not to the execution of a container service. D – File format registered for JSON or CSV ingestion applies to the Snowpipe ingestion pipeline, not to container execution.
Thus, only A and E satisfy the requirements for provisioning a container with Snowpark Container Services.

Reference:

Snowflake Documentation – Deploying Containers with Snowpark Container Services : https://docs.snowflake.com/en/snowpark-container-services/deploying-containers Snowflake Documentation – Service Specification for Snowpark Container Services : https://docs.snowflake.com/en/snowpark-container-services/service-specification

Explanation:

Why option D is correct
Snowflake Cortex Guard is billed for the billable tokens that are actually consumed when a query accesses the protected fields ( from_text and question ).
Each token processed contributes to the compute cost; there is no per-row, per-user, or flat-rate charge. The pricing model therefore aligns usage-based billing with the amount of textual data that Cortex Guard analyzes.
Why the other options are unsuitable

A: “Compute charges are based on the number of rows scanned in a table.” Snowflake’s elastic compute is token-centric for Cortex Guard, not row-count dependent, so row scans do not drive the cost.
B: “Compute charges are based on the number of input tokens processed.” This statement is partially true but incomplete; the cost is tied specifically to tokens in the from_text and question fields, not to all input tokens of an arbitrary query.
C: “Charges are based on the number of users accessing the data.” User count does not affect Cortex Guard pricing; charges are purely data-processing (token) based.

Reference:

Snowflake Cortex Guard Overview & Pricing: https://docs.snowflake.com/en/cortex-guide/guard-intro Cortex Guard Billing Details (Pricing model): https://docs.snowflake.com/en/cortex-guide/cortex-guard-pricing