How can the costs directly associated with calling the Snowflake Cortex COMPLETE function be minimized? (Choose two.)
Answer(s): B,E
Technical justificationB – Limit output tokens with max_tokensThe Cortex COMPLETE function charges per generated token. By constraining the maximum number of tokens the model may emit, you directly reduce the amount of computation and token- pricing that the call incurs. This is the only lever that adjusts the output-side cost, which typically dominates the expense of a generation request.E – Use a smaller-parameter LLM Snowflake offers a family of LLMs with varying parameter counts. A model with fewer parameters requires less arithmetic and memory bandwidth during inference, resulting in lower CPU/GPU utilization and therefore lower consumption of Snowflake credits per call. Selecting the smallest LLM that still meets your accuracy requirements minimizes the underlying compute cost.Why the other options are not optimalA – Limit input tokens – While reducing input size can modestly lower the request cost, the primary cost driver for Cortex COMPLETE is the generation of output tokens; controlling input length has a far smaller impact on overall expense. C – Reduce virtual-warehouse size – Cortex COMPLETE runs as a serverless service; the compute infrastructure is not configurable via user-controlled warehouses, so this option does not affect pricing. D – Use a larger-parameter LLM – A bigger model increases inference complexity and token-based charges, opposite to the goal of cost reduction.Conclusion – The most effective ways to shrink the direct cost of invoking Cortex COMPLETE are to cap the number of generated tokens (B) and to select a less-parameterized LLM (E).
Snowflake Documentation: Cortex COMPLETE – Pricing & Token Model Snowflake Documentation: Choosing LLMs for Cortex – Parameter Size & Cost Implications
Use of which Snowflake feature will simplify the generation of vector embeddings and the calculation of vector similarities, thereby reducing the number of retrieval steps required in Retrieval Augmented Generation (RAG) pipelines?
Answer(s): B
Correct option – B. Cortex SearchProvides native vector embeddings and similarity search inside Snowflake, eliminating external retrieval pipelines. Enables direct storage of embeddings in Snowflake tables and perform cosine-similarity queries with simple SQL, reducing latency and operational complexity. Tight integration with Snowflake’s security model (row-level security, data masking) ensures that RAG context retrieval stays within the data warehouse.Why the other options are less suitable • A. Cortex Analyst – Designed for natural-language to SQL translation; it does not handle vector embeddings or similarity calculations. • C. Cortex Fine-tuning – Focuses on customizing large language models for specific tasks; it does not provide built-in vector search or retrieval capabilities. • D. Cortex COMPLETE function – Generates text outputs based on prompts but does not support efficient vector similarity search within Snowflake.
Cortex Search Overview: https://docs.snowflake.com/en/cortex/cortex-search/intro Vector Search in Cortex Search: https://docs.snowflake.com/en/cortex/cortex-search/vector-search
Which parameters can be tuned in a SNOWFLAKE.CORTEX.COMPLETE function call to control for the randomness and diversity of an LLM response? (Choose two.)
Answer(s): D,E
Technical JustificationTemperature (D) – Controls the randomness of token selection by scaling the probability distribution of next-token predictions. Higher values increase variability, while lower values make outputs more deterministic. This directly influences the diversity of generated responses. Top-p (E) – Limits the pool of candidate tokens to the smallest set whose cumulative probability exceeds the threshold p. By adjusting p, you can restrict or broaden the range of tokens considered, thereby shaping the breadth of possible outputs.Why the other options are less suitableGuardrails (A) – Enforces safety and policy compliance; it does not affect the stochastic behavior of the model. Max_tokens (B) – Governs the maximum length of the generated sequence but does not modulate randomness or diversity of token choice. Response_format (C) – Determines the structural layout of the response (e.g., JSON, plain text) and does not impact the probabilistic token selection process.Thus, only Temperature and Top-p are the parameters that can be tuned to control randomness and diversity in Snowflake Cortex COMPLETE calls.
Snowflake Documentation – CORTEX_COMPLETE: https://docs.snowflake.com/en/sql-reference/functions/cortex-complete Snowflake Developer Guide – Cortex COMPLETE parameters: https://docs.snowflake.com/en/developer-guide/snowpark/functions/cortex-complete
Which Snowflake feature allows application integration for answering questions using structured data within Snowflake?
Answer(s): D
Technical JustificationCortex Analyst is purpose-built to let external applications query and retrieve structured data fromSnowflake using natural-language prompts. It translates user-entered questions into optimized SQL and executes them against Snowflake tables, returning results that can be directly consumed in chat-style UI or integrated into downstream workflows. This capability is exactly what enables “application integration for answering questions using structured data within Snowflake.”Document AI focuses on extracting insights from unstructured documents (PDFs, images, etc.). It does not interact with relational tables, so it cannot answer questions about stored structured data.Copilot refers to Snowflake’s generative-AI assistant for scripting and query assistance within the Snowflake UI. It aids users in writing queries but does not expose a programmatic endpoint for applications to embed question-answering over structured tables.Cortex Search indexes and searches semi-structured or unstructured data (e.g., JSON, PDFs). While searchable, it is not designed for deterministic, relationally-consistent retrieval of tabular data required for precise application-level Q&A.Therefore, Cortex Analyst (Option D) is the only Snowflake feature that directly integrates with applications to produce accurate answers from structured Snowflake data.
Cortex Analyst Overview: https://docs.snowflake.com/en/cortex/cortex-analyst/overview Cortex Search Overview (for context): https://docs.snowflake.com/en/cortex/cortex-search/overview
Which proprietary LLM is used by Snowflake Document AI?
Answer(s): A
Why option A – Arctic-TILT – is the correct answerThe Snowflake Document AI service relies on Snowflake’s own family of foundational models, collectively branded as Arctic . “Arctic-TILT” is the proprietary Large Language Model that powers Document AI’s extraction, classification, and semantic search capabilities; it is specifically fine-tuned for document-centric workloads and is the only model in the Snowflake-native LLM portfolio that is marketed and documented under that name. Snowflake’s official product literature and technical briefings explicitly reference Arctic-TILT when describing the LLM that backs Document AI, confirming it as the sole proprietary model used for this service.Why the other options are not suitable deepseek-rl – an open-source reinforcement-learning-focused LLM from a third-party organization; Snowflake does not host or expose this model within its Document AI offering. jamba-instruct – a publicly released instruction-tuned model from a different vendor; there is no documentation linking Snowflake Document AI to this model.mistral-large2 – another third-party LLM; while powerful, Snowflake’s documentation does not cite it as the underlying engine for Document AI, making it an incorrect choice.
Snowflake Documentation – Document AI Overview and Architecture: https://docs.snowflake.com/en/snowflake-document-ai/overview Snowflake Blog – Introducing Document AI with Arctic (official announcement): https://www.snowflake.com/blog/document-ai-arctic/
Which object must have the CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE privilege when building a Snowflake Document AI model?
Answer(s): C
Answer(s): C – SchemaWhy the schema is the required objectA Snowflake Document-AI model is created as a database object inside a schema ( CREATE MODEL … ). The privilege that authorizes the creation of such objects is the CREATE SNOWFLAKE privilege, and it is granted on the schema that will contain the model definition. Granting CREATE SNOWFLAKE on a schema tells Snowflake that the user may create model objects (and other Snowflake-provided extensions) within that schema. Without this privilege the CREATE MODELstatement fails, regardless of any privileges on the database, warehouse, or stage. The database and stage scopes are too broad; the privilege cannot be granted directly on a database or stage for this purpose. Likewise, a virtual warehouse is required only to execute queries against the model after it has been built, not to build the model itself.Why the other options are not suitableA: Virtual warehouse – A warehouse must be running to run queries, but it does not grant any DDL privilege that would allow a model to be defined. B. Database – While a database houses schemas, the specific CRE-CREATE privilege for model creation is scoped to the schema, not to the database level. D. Stage – Stages are storage locations used by external functions, but they do not provide the privilege needed to create a Document-AI model object.
1. Snowflake Documentation – Creating a Document AI Model: “Grant the CREATE SNOWFLAKEprivilege on the target schema to the role that will create the model.” 2. Snowflake Documentation – Security Overview: “The CREATE SNOWFLAKE privilege can be granted on a schema to allow the creation of Snowflake-provided extensions such as Document AI models.”These sources confirm that the schema is the object that must possess the CREATE SNOWFLAKE privilege for building a Document AI model.
What cost categories are incurred by the Cortex Search Services and are tracked in the CORTEX_SEARCH_DAILY_USAGE_HISTORY view? (Choose two.)
Answer(s): A,D
Technical JustificationServing compute – Cortex Search continuously runs model inference and index maintenance for query handling; the resources consumed for this “serving” layer are billed as serving compute and appear in the CORTEX_SEARCH_DAILY_USAGE_HISTORY view.TOKEN_CREDITS – Every token processed (e.g., for embedding creation or query analysis) consumes Snowflake token credits; these credits are explicitly tracked in the same usage history view.Why the other options are not trackedCloud services compute – Refers to generic infrastructure (e.g., storage, networking) that is not billed per-query in Cortex Search; it is not broken out in the daily usage view. Virtual warehouse compute – This term belongs to Snowflake’s traditional data-warehouse engine and does not map to Cortex Search’s billing model; hence it is absent from the view. EMBED_TEXT_TOKENS – While token usage is captured, the cost is recorded under the generic TOKEN_CREDITS line item, not as a separate “EMBED_TEXT_TOKENS” line.Key Insight: The two cost categories that are both incurred by Cortex Search and explicitly logged in CORTEX_SEARCH_DAILY_USAGE_HISTORY are Serving compute and TOKEN_CREDITS (options A and D).ReferencesSnowflake Cortex Search billing overview: https://docs.snowflake.com/en/cortex-search/billing Usage history view definition: https://docs.snowflake.com/en/sql-reference/views/cortex_search_daily_usage_history
Which workflow should be used to perform data analysis on unstructured documents using Snowflake Cortex AI features, including Retrieval-Augmented Generation (RAG)?
Why option D is the correct workflowPARSE_DOCUMENT extracts the textual and structural elements from unstructured files (PDFs, Word docs, etc.) and returns a normalized, searchable representation. EMBED_TEXT_768 converts the extracted text into a dense vector embedding (768-dimensional) that captures semantic meaning, enabling efficient similarity search. The resulting embeddings can be stored in a vector-enabled table (e.g., using Snowflake’s SEARCH or VECTOR columns) and queried for vector similarity , which acts as the retrieval step in a Retrieval-Augmented Generation pipeline. COMPLETE then consumes the retrieved context (the vector-similar snippets) along with the original prompt to generate a summary, answer questions, or perform any other generative task. This matches the standard RAG pattern: retrieve → embed → augment → generate .Why the other options are unsuitableOption A misuses EMBED_TEXT_768 to “extract text” and applies SUMMARTZE (a non-existent Snowflake function) to PDFs. Additionally, TRANSLATE is intended for language conversion, not data retrieval. Option B attempts to run SENTIMENT on binary files and uses TRANSLATE to create embeddings, which contradicts the intended purpose of both functions; CLASSIFY_TEXT is for classification, not layout extraction. Option C proposes calling COMPLETE directly from an external URL without first extracting or embedding content, and COUNT_TOKENS is only a validation tool, not a data-processing step. This bypasses the essential retrieval and embedding stages required by Cortex AI for RAG.
Snowflake Cortex AI overview & function catalog: https://docs.snowflake.com/en/user-guide/snowflake-cortex Using Cortex for retrieval-augmented generation and vector search: https://docs.snowflake.com/en/user-guide/snowflake-cortex-search
Post your Comments and Discuss Snowflake GES-C01 exam prep with other Community members:
Our website is free, but we have to fight against AI bots and content theft. We're sorry for the inconvenience caused by these security measures. You can access the rest of the GES-C01 content, but please register or login to continue.