Free GES-C01 Exam Braindumps
SnowPro Specialty Gen AI Questions 41–48 (page: 7)

Explanation:

Why the correct options are best
Option A – SNOWFLAKE.CORTEX.COMPLETE with a RESPONSE_FORMAT
CORTEX.COMPLETE can be instructed to return a structured response (e.g., JSON) by setting
RESPONSE_FORMAT='JSON' .
The resulting JSON can be accessed with dot notation (e.g., result:fields.column1 , result:fields.column2 ,
result:fields.column3 ) and materialized as separate columns in a SELECT statement.
This approach guarantees that each generated column contains a deterministic, typed value derived from the free-form text, which is ideal for producing multiple derived fields in a single pipeline step.
Option D – SNOWFLAKE.CORTEX.EXTRACT_ANSWER
CORTEX.EXTRACT_ANSWER is purpose-built to extract pre-defined answer keys from LLM output.
When called in a SELECT , it returns a result object whose fields can be referenced with dot notation to populate distinct columns (e.g., extract_ans1 , extract_ans2 , extract_ans3 ).
Because the extraction keys are supplied up front, the function produces a clean, tabular output without the need for additional parsing logic, making it the most direct way to generate three separate columns from a single text column.
Why the other options are less suitable
Option B – CORTEX.COMPLETE with TOP_P
TOP_P only controls the randomness of the generated text; it does not enforce a structured schema.
Consequently, extracting deterministic column values would require custom parsing of free-form output, increasing maintenance overhead and risking inconsistencies.
Option C – CORTEX.ENTITY_SENTIMENT per column
ENTITY_SENTIMENT is designed for sentiment analysis on identified entities, not for extracting arbitrary user-defined fields. Using it three times would only yield sentiment scores, not the varied information needed for three distinct new columns.
Option E (duplicate of D)
Although technically valid, the duplicate entry does not add a new distinct method; the certification expects the two unique choices (A and D).
Implementation sketch
SELECT text_col, cortex.complete( 'SELECT * FROM my_table WHERE id = :id', 'response_format': 'JSON'
or, using EXTRACT_ANSWER :
SELECT text_col, cortex.extract_answer(text_col, 'ANSWER_1') AS col1, cortex.extract_answer(text_col, 'ANSWER
Both patterns materialize the three derived columns directly in SQL, satisfying the pipeline requirement.

Reference:

Snowflake Cortex – COMPLETE function Snowflake Cortex – EXTRACT_ANSWER function

Explanation:

Technical justification
D – Suspend serving when it is not necessary to run the service Snowflake Cortex Search is billed on a per-second basis for both compute and storage used while the search service is active. By pausing (suspending) the service when there is no workload or when indexing/realtime updates are scheduled only for off-peak windows, you eliminate idle compute charges and storage-related fees. This directly reduces the overall cost while still preserving the full search capability when needed.
A – Make frequent but small updates to the source data Frequent incremental changes increase the refresh frequency of the Cortex index. Each refresh incurs additional compute cycles and may trigger extra storage write-amplification, leading to higher per-second billing. Continuous small updates therefore raise, rather than lower, costs.
B – Split the text search columns into chunks of no more than 612 tokens Token-level chunking primarily influences indexing latency and the size of individual indexing jobs.
While it can help manage processing limits, it does not affect the amount of compute time Cortex runs; the service still incurs charges for the total duration it is active. Consequently, this approach yields negligible cost savings.
C – Reduce the target tag from 5 minutes to 1 minute for real-time indexing Shortening the target tag forces more frequent index rebuilds, which consumes additional compute resources and storage writes. This results in higher continuous compute usage and thus higher costs, making it the opposite of a cost-saving measure.
Conclusion The only option that actively reduces billable compute time by halting the Cortex Search engine when not in use is D . All other options either increase processing frequency or simply adjust indexing granularity without impacting the core billing model.

Reference:

Snowflake Cortex Search Overview https://docs.snowflake.com/en/snowflake-cortex/search
Pricing for Snowflake Cortex Services (pay-as-you-go) https://docs.snowflake.com/en/snowflake-cortex/pricing

Explanation:

Explanation:

Why options A and B are the correct governance principles
A – Customer data isolation – Snowflake’s architecture guarantees that a customer’s data is stored in a separate, encrypted logical layer and is never exposed to other customers or to the model developers that train or serve the underlying LLMs. Data can only be accessed through the customer’s own Snowflake account and permissions.
B – Role-Based Access Control (RBAC) for models – Access to Cortex models (including foundation and custom models) is governed by Snowflake RBAC. Permissions are granted at the role level, ensuring that only authorized roles can query, fine-tune, or deploy models, thereby enforcing least-privilege access to the underlying AI capabilities.
Why the remaining options are not appropriate
C – Sharing fine-tuned models with other customers – Snowflake does not automatically permit a customer’s fine-tuned model to be shared with other customers; sharing requires explicit permission through Snowflake shares or private model-sharing mechanisms, and the governance model does not guarantee that such sharing will occur.
D – Third-party model access limited to the Model Registry – Access to third-party models is managed through Snowflake’s model-catalog and authentication mechanisms, but there is no governance rule that restricts usage solely to the Model Registry; models can also be called via APIs, Cortex Functions, etc.
E – Mandatory use of Cortex Guard – Cortex Guard is an optional safety layer that can be enabled to filter unsafe responses, but it is not a mandatory governance guarantee for all Cortex COMPLETE calls; usage depends on the user’s configuration.
Conclusion Governance in Snowflake’s Gen AI offering centers on data isolation (A) and RBAC-driven model access (B) . These two principles directly protect customer data and enforce controlled, role-based interactions with AI models, which aligns with Snowflake’s security and compliance commitments.

Reference:

Snowflake Security Overview – Data Isolation – https://docs.snowflake.com/en/snowflake-security/security-overview Role-Based Access Control (RBAC) in Snowflake – https://docs.snowflake.com/en/sql-reference/sql/create-role Cortex Guard – Safety Filtering (optional feature) – https://docs.snowflake.com/en/cortex/cortex-guard Model Registry and Model Access Controls – https://docs.snowflake.com/en/cortex/cortex-security#model-registry-and-access-controls

Explanation:

In many AI SDKs and API implementations, specific parameters have strict requirements regarding the data structure they expect.
When you use the options argument—particularly in a context involving guardrails or advanced configuration—the system often requires the input (the prompt_or_history) to be formatted as an array of objects rather than a simple string.
Why D is correct: This structure allows the API to maintain context, roles (e.g., "user," "assistant," "system"), and the specific state of the conversation, which is necessary when applying guardrails that need to evaluate the history and intent of the prompt.
Why the others are likely incorrect:
A & C: Temperature and max_tokens are generally independent configuration parameters.
While they influence the output, they are rarely mutually exclusive with guardrails in standard API designs.
B: While function parameter order matters in many programming languages, modern SDKs often use keyword arguments (named parameters), making the order of the arguments secondary to the structural validity of the data passed to them.

Explanation:

Technical Justification
Option D – snowflake.cortex.complete Python function – This Snowpark-provided function is exposed through Snowflake’s external API as a REST endpoint that accepts JSON payloads (prompt, parameters, etc.) and returns completions from the Cortex LLM. Because the function is callable via a documented HTTP API, it is the only choice that offers programmatic connectivity through an exposed REST API.
Option A – Copilot – A conversational UI for Snowflake users; it does not expose a programmable REST endpoint.
Option B – Cortex LLM Playground – An interactive web UI for experimenting with prompts; it is not intended for external programmatic access via API.
Option C – “A SNOWFLAKE” – Nonsensical placeholder and does not represent an actual feature.
Option E – CORTEX.COMPLETE SQL function – Executes the LLM completion entirely within Snowflake’s
SQL engine; the result is returned as part of the query result set, not as a callable external REST service.
Therefore, the only Snowflake Gen AI capability that explicitly provides connectivity via an exposed REST API is the snowflake.cortex.complete Python function .

Reference:

1. Snowflake Cortex Python API – Snowpark documentation:
https://docs.snowflake.com/en/snowpark/python-functions/cortex-complete 2. Snowflake External REST APIs – Developer guide: https://docs.snowflake.com/en/developer-
guide/snowflake-components/rest-api

Explanation:

Technical justification
C: When no Verified Query Repository (VQR) is present, Cortex Analyst uses the underlying LLMs to generate up to three suggested questions. The service falls back to its built-in LLM-driven query synthesis when a VQR is unavailable, limiting the output to three question candidates for usability and performance reasons.
D: When a Verified Query Repository (VQR) is defined, Cortex Analyst suggests up to five questions based on semantic similarity to the user's input. With an active VQR, the engine expands the suggestion pool to five items and ranks them by semantic relevance to the provided natural-language input, allowing more tailored options.
Why the other options are incorrect
A – Cortex Analyst does not guarantee that every suggested question returns results; it may propose questions that are syntactically valid but could yield empty result sets. B – Onboarding questions flagged in VQR are shown based on semantic similarity, not only when the user supplies no input. E – Prioritization is driven by semantic similarity, not by the frequency of logical columns referenced in the repository.

Reference:

Snowflake Documentation – Cortex Analyst: Suggested Questions Overview https://docs.snowflake.com/en/cortex/cortex-analyze/cortex-analyst-suggested-questions Snowflake Documentation – Verified Query Repository (VQR) Best Practices https://docs.snowflake.com/en/cortex/cortex-analyze/vqr-best-practices

Explanation:

Why CE are the correct choices
E – Model Registry supports versioning The registry stores metadata for every model version, allowing multiple numbered versions of the same model to coexist. This enables traceability, rollback, and controlled promotion of newer capabilities without affecting production workloads.
C – Models can be deployed and invoked directly using SQL commands in Snowflake Once a model is registered, Snowflake allows it to be called from SQL (e.g., via the PREDICT function or registered functions that wrap the model). No external deployment step or code change is required; the model runs inside the Snowflake engine, delivering predictions as part of regular queries.
Why the other options are inaccurate
A – “Once a model is registered, it cannot be updated or deleted.” The registry is designed for mutable lifecycle management: you can register a new version, replace an existing one, or delete obsolete versions while preserving earlier ones for audit purposes.
B – “Models need to be exported to external storage before they are registered.” Registration does not require external export; models built within Snowflake (e.g., via Snowpark, SnowflakeML, or external functions) can be registered directly, keeping all artifacts internal to the platform.
D – “Models registered … can only be used with Snowpark for Python.” The registry is language-agnostic. Registered models can be consumed by Snowpark (Python, Scala), native
SQL predicates, or other Snowflake ML APIs; they are not confined to a single language runtime.

Reference:

Snowflake Model Registry Overview – Explains versioning, lifecycle, and SQL invocation capabilities. https://docs.snowflake.com/en/user-guide/snowflake-by-data-cloud/model-registry Registering and Deploying Models in Snowflake – Details how to register a model, manage versions, and call it via SQL. https://docs.snowflake.com/en/developer-guide/snowflake-model-registry/register-models