Usage is calculated base on both input and output tokens for which Snowflake Cortex LLM functions? (Choose three.)
Answer(s): A,C,F
Why A, C, and F are correctSUMMARIZE generates a textual summary, so the service must read the input tokens and then produce output tokens for the summary. TRANSLATE takes an input text in one language and outputs a translated version; both the input and the generated translation consume token-based usage. CLASSIFY_TEXT returns a classification label (or multiple labels) that are generated tokens; therefore its usage is measured on both the incoming text and the produced label(s).Why the other options are not countedSENTIMENT returns a fixed-size sentiment score (often a single token); while it consumes input tokens, it does not generate a variable-length output text, so only input tokens are billed. EMBED_TEXT_768 and EMBED_TEXT_1024 compute vector embeddings. They only read input tokens to produce a numeric embedding; no output token stream is created, so usage is calculated on input tokens alone.
Snowflake Cortex LLM Functions – Usage billing details: https://docs.snowflake.com/en/snowflake-cortex/usage-billing Snowflake Documentation – SUMMARIZE, TRANSLATE, CLASSIFY_TEXT: https://docs.snowflake.com/en/snowflake-cortex/cortex-functions-llm-summarize-translate-classifyThese sources confirm that billing for Cortex LLM functions accounts for both the input and the generated output tokens only for the functions that actually generate text (SUMMARIZE, TRANSLATE, CLASSIFY_TEXT).
A Gen AI Specialist is building an internal chatbot for a company that will use Streamlit and Snowflake Cortex Analyst to interpret and respond to user queries against Snowflake data. Team members need to be able to ask a question like: Can you provide the sales trend for the last 6 months? The tool needs to be able to answer follow-up questions like: Can you break that down by region? What factors contributed to the dip in February? To ensure the tool interprets the follow-up questions correctly, what must be included in the API request to Cortex Analyst?
Answer(s): C
Why option C is the correct choiceCortex Analyst treats each incoming request as part of a multi-turn conversation . To produce accurate follow-up responses, the model must see the complete dialogue (the user’s original question, the model’s answer, and any subsequent clarifications). The API payload therefore requires a messages field containing the entire conversation history ; this context enables the model to understand implied references (“break that down by region”, “factors that caused the dip”). Without the full history, the model would have no way to infer which entities (“that”, “February”) refer to previous turns, leading to ambiguous or incorrect answers.Why the other options are not appropriateA – “The most recent user message” Supplying only the latest query ignores earlier context, so the analyst could not determine what “that” or “February” refer to.B – “A session token to maintain session state across API requests” Cortex Analyst does not rely on an external token for state; session context is passed explicitly via the messages field. A token would be unnecessary and would not convey the needed dialogue history.D – “A Snowflake session parameter to maintain conversation context among the previous API requests” Snowflake session parameters are scoped to a single SQL session and cannot be used to convey multi-turn conversation data to an external Gen AI endpoint. They are unrelated to the Cortex Analyst API contract.Hence, the specialist must include the entire conversation history in the messages field to ensure proper interpretation of follow-up questions.
Cortex Analyst API – Request format : https://docs.snowflake.com/en/developer-cortex-analyst/api-reference#request-payload Multi-turn conversation handling : https://docs.snowflake.com/en/developer-cortex-analyst/guided-analytics#contextual-prompting
What Snowflake Cortex AI observability metrics are used to evaluate responses generated by LLMs? (Choose two.)
Answer(s): B,C
Technical JustificationCoherence – This metric assesses the logical flow and structural integrity of the LLM’s output. In SnowflakeCortex AI, coherence is measured by evaluating whether the generated response maintains a consistent topic, proper sentence structure, and a natural progression of ideas. High coherence indicates that the model’s answer reads smoothly and is understandable to the user.Correctness – Also referred to as “groundedness” or “answer quality,” this metric verifies that the response accurately reflects the underlying data or knowledge base it was drawn from. It checks for factual alignment, proper citation of source material, and absence of hallucinated or contradictory statements. Correctness is essential for ensuring that the LLM’s output can be trusted for decision-making.Why Accuracy, F1 Score, and Precision are less suitableAccuracy in Snowflake Cortex AI is not used as a primary observability metric for LLM responses; it is more commonly applied to classification tasks rather than to evaluate the linguistic quality of generated text. F1 Score and Precision are statistical measures derived from classification confusion matrices (true/false positives/negatives). They are designed for discrete prediction problems and do not directly capture the semantic qualities of language generation such as flow or factual alignment, making them poor fits for LLM observability.Conclusion The two Snowflake Cortex AI observability metrics that specifically evaluate LLM-generated responses are Coherence and Correctness , as they directly address the structural and factual quality of the output, respectively.
Snowflake Cortex AI Observability Metrics – Cortex Observability Documentation Evaluating LLM Responses in Snowflake Cortex – Cortex LLM Evaluation Guide
A Gen AI Specialist needs a solution that will support asking questions in natural language and receive answers based on structured data in Snowflake without the use of SQL. How can this be accomplished? (Choose two.)
Answer(s): C,D
Technical JustificationCorrect options C and D – They leverage semantic models that map tables and columns to business concepts, and they expose a text-to-SQL interface via Snowflake’s AI endpoints. • Option C enables natural-language queries directly from the AI & ML Studio in Snowsight, eliminating the need for custom application code. • Option D extends the same capability by allowing developers to embed the query in a custom application that consumes the Snowflake Cortex Analyst API, giving flexible integration while still using a semantic model.Why the other options are unsuitable • A – “Snowflake Document AI” is designed for semi-structured textual documents, not for querying tabular data; custom data models and Document AI together do not provide a semantic-layer query interface. • B – “Cortex Agent API” with cross-region inference introduces unnecessary architectural complexity and does not address the semantic-model requirement; it is geared toward multi-modal inference rather than text-to-SQL over structured data. • E – Creating a Cortex Agent and fine-tuning an LLM with extracted data is an over-engineered approach; it does not harness the built-in Cortex Analyst API or semantic model features that are purpose-built for NL-to-SQL over structured tables.Summary – Options C and D combine Snowflake’s native semantic modeling with its approved AI endpoints (Cortex Analyst), providing the simplest, supported, and production-ready way to ask natural-language questions about structured data without writing SQL.
Semantic Models Overview : https://docs.snowflake.com/en/canada/latam/semantic-models Cortex Analyst API Documentation : https://docs.snowflake.com/en/developer-guide/cortex/api-analyst
A Gen AI Specialist is building a Gen AI application in Snowflake and needs to debug a specific record that returned an inaccurate response. The Specialist needs to understand how the input prompt was handled, what context was retrieved, which tools were invoked, and how the LLM responded, step-by-step. Which Snowflake AI observability feature should be used?
Technical justificationTracing captures the complete execution path of a Gen AI request, showing each stage: prompt handling, context retrieval, tool calls, and the LLM’s output, enabling precise debugging of inaccurate responses. Logging records discrete events but does not link them into a causal chain, so it cannot reveal the sequence of operations that led to the error. Metrics provide aggregate performance data (latency, error rates) but lack detail about the specific input-output flow for a single record. Monitoring offers real-time health overview and alerts, but it does not expose the granular step-by-step context needed to diagnose a particular faulty response.Therefore, Tracing is the appropriate observability feature for deep, step-wise debugging of a Gen AI record.
https://docs.snowflake.com/en/snowflake-cortex/observability/tracing https://docs.snowflake.com/en/snowflake-cortex/observability/logging
Review this query: SELECT DEMO_DB.DEMO_SCHEMA.EXM_DOC_AI_BUILD!PREDICT( GET_PRESIGNED_URL(@<stage_name>, '<relative_file_path>'), 9); What will be the output?
JustificationThe call GET_PRESIGNED_URL(@<stage_name>, '<relative_file_path>') returns a presigned URL that points to a file in a Snowflake internal stage ( DEMO_SCHEM ). Internal stages are owned by the user and do not require external storage credentials. EXM_DOC_AI_BUILD!PREDICT( … ) invokes a Document AI model-version that expects the file URL as its input and returns the AI inference result . Document AI outputs are semi-structured (e.g., JSON containing entities, key-value pairs, tables), not a flat relational table (structured) nor a free-form text blob. Because the source stage is internal, the correct description must reference an internal stage and the nature of the result must be semi-structured data .Why the other options are incorrectOption A – Incorrectly describes the output as structured data. Document AI does not return a tabular result set; it returns semi-structured JSON. Option B – Incorrectly claims the document resides in an external stage. The URL is generated from an internal stage ( DEMO_SCHEM ), so the source is not external.Option D – While it correctly identifies the output as semi-structured, it wrongly specifies an external stage, which does not match the query’s use of DEMO_SCHEM .Therefore, option C accurately captures both the storage context (internal stage) and the output type (semi-structured data).
Document AI Predict function: https://docs.snowflake.com/en/sql-reference/functions/document_ai_predict Presigned URL for internal stages: https://docs.snowflake.com/en/user-guide/data-load-unstructured#using-presigned-urls
A Gen AI Specialist creates this table:Which function will measure the relationship between vectors without considering their magnitudes?
Answer(s): D
D: VECTOR_COSINE_SIMILARITY is correct.ExplanationCosine similarity is designed specifically to measure the orientation (angle) between two vectors, effectively ignoring their magnitude.How it works: It calculates the cosine of the angle between two non-zero vectors in an inner product space. If the vectors point in the same direction, the cosine similarity is $1$; if they are orthogonal (perpendicular), it is $0$; and if they point in opposite directions, it is $-1$.Why it ignores magnitude: Because it normalizes the vectors by their lengths (norms) during the calculation, it focuses entirely on the "direction" or the relationship between the features, regardless of how large or small the feature values are.
Which arguments are required when using the SNOWFLAKE.CORTEX.COMPLETE function? (Choose two.)
Answer(s): A,E
JustificationA: model – This argument specifies the Cortex model to invoke. It is mandatory because the function must know which model to call. E. prompt_or_history – This argument supplies the text prompt (or chat history) that drives the model’s response. It is also mandatory as the function needs input data to generate output.The remaining arguments are optional enhancements:B: top_p – Controls sampling diversity; useful for controlling creativity but not required. C. response_format – Determines the output structure (e.g., JSON); optional for most use-cases. D. max_tokens – Limits the length of the generated text; optional and defaults to the model’s limit.Therefore, only A and E are required arguments for SNOWFLAKE.CORTEX.COMPLETE .
Snowflake Documentation – SNOWFLAKE.CORTEX.COMPLETE function: https://docs.snowflake.com/en/sql-reference/scalar-functions/cortex/complete Snowflake Cortex Developer Guide – Using the COMPLETE function: https://docs.snowflake.com/en/developer-guide/cortex/complete-function
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.