CertArea
All Snowflake exams

Free GES-C02 practice test: 30 questions with answers

30 original questions for SnowPro Specialty: Gen AI, drawn from every topic area. Answer each one to see whether you got it right and why. No sign-up.

Free questions
30
Full bank
100 questions

GES-C02 exam guide: topics, cost and format

All 30 GES-C02 questions with answers

The same questions as the test above, grouped by exam topic. Open any question to see the answer and explanation.

Cortex LLM functions (6 questions)

  1. Question 1

    A data engineer wants to generate a text summary of each customer review directly in a SQL query using a hosted model. Which Snowflake Cortex function is designed for this?

    1. AGET_DDL
    2. BSHOW TABLES
    3. CSNOWFLAKE.CORTEX.EMBED_TEXT only
    4. DSNOWFLAKE.CORTEX.COMPLETE
    Show answer and explanation
    Correct answer: D

    CORTEX.COMPLETE calls a hosted language model from SQL to generate text such as summaries. EMBED_TEXT creates embeddings, SHOW TABLES lists tables, and GET_DDL returns object definitions.

  2. Question 2

    Which Cortex function creates vector embeddings from text so that the text can be compared by similarity?

    1. ASNOWFLAKE.CORTEX.EMBED_TEXT_768 or the current EMBED_TEXT function family
    2. BSNOWFLAKE.CORTEX.TRANSLATE with a source and target language
    3. CSNOWFLAKE.CORTEX.SENTIMENT to measure the emotional polarity
    4. DSNOWFLAKE.CORTEX.SUMMARIZE to condense lengthy product documents
    Show answer and explanation
    Correct answer: A

    The EMBED_TEXT function family creates embeddings for similarity search. Sentiment, translation, and summarization functions produce other outputs, not embeddings.

  3. Question 3

    A team wants to classify support tickets into a fixed set of categories with a model, in SQL. Which Cortex capability fits best?

    1. AA stored procedure that deletes tickets from the support table nightly
    2. BA manual GROUP BY on the ticket text with exact string matching
    3. CA Cortex LLM function prompted with the category list and the ticket text
    4. DA table partition by category name that stores each ticket separately
    Show answer and explanation
    Correct answer: C

    A Cortex LLM function with the category list and ticket text performs classification in SQL. A GROUP BY on raw text does not classify meaning, partitions do not classify, and deleting tickets is unrelated.

  4. Question 4

    A query calls a model for every row of a large table, and costs are higher than expected. Which optimization is most effective?

    1. AFilter rows before calling the model and process only the rows that need it
    2. BStore the results in a text file on a laptop and copy the file to a shared drive
    3. CDisable the warehouse to cut costs and rerun the query after hours
    4. DCall the model for every row of every table in the schema
    Show answer and explanation
    Correct answer: A

    Filtering before the model call reduces the number of billable model invocations. Calling for every row, disabling the warehouse, and local text files do not reduce cost appropriately.

  5. Question 5

    A team wants to choose a model for a Cortex function. What factor should guide the choice?

    1. AThe number of columns in the table and its declared width
    2. BThe colour of the query editor theme chosen in the account preferences
    3. CThe model's name length in characters, including its version suffix and prefix
    4. DTask quality, cost, latency, and region availability for the model
    Show answer and explanation
    Correct answer: D

    Model choice depends on task quality, cost, latency, and availability in the region. Name length, editor colour, and column counts do not determine suitability.

  6. Question 6

    Which Cortex function translates text from one language to another directly in SQL?

    1. ASNOWFLAKE.CORTEX.SENTIMENT
    2. BSNOWFLAKE.CORTEX.TRANSLATE
    3. CSNOWFLAKE.CORTEX.EXTRACT_ANSWER
    4. DSNOWFLAKE.CORTEX.EMBED_TEXT
    Show answer and explanation
    Correct answer: B

    TRANSLATE performs language translation in SQL. Sentiment scores opinions, extract-answer finds answers in text, and EMBED_TEXT creates embeddings.

Cortex Search and retrieval (6 questions)

  1. Question 7

    A team wants to build a search service over support articles that returns results by meaning and keeps the index updated as articles change. Which Snowflake feature fits?

    1. AA materialized view of article titles only
    2. BAn external spreadsheet filter
    3. CCortex Search service on the source table
    4. DA manual full-text index rebuilt once a year
    Show answer and explanation
    Correct answer: C

    Cortex Search provides managed semantic and keyword retrieval over tables and keeps its index refreshed. Yearly rebuilds, title-only views, and spreadsheet filters do not provide meaning-based search with updates.

  2. Question 8

    A RAG application needs the most relevant passages for each question, and the retriever must filter by product line. Which capability supports this?

    1. AAttribute filters on the search service combined with the query
    2. BFiltering the retrieved passages by the character length of the question
    3. CSending the full passage set to the model on each question without ranking
    4. DIgnoring the product line attribute and retrieving passages by similarity
    Show answer and explanation
    Correct answer: A

    Attribute filters limit retrieval to the relevant product line while the query ranks results. Sending everything, ignoring the filter, and filtering by question length do not produce relevant passages.

  3. Question 9

    Why do RAG pipelines often split long documents into chunks before indexing them?

    1. AChunks make documents larger for storage because each chunk is duplicated
    2. BSmaller chunks give more precise retrieval and fit the model's context limit
    3. CChunks remove the need for metadata since each chunk is self-describing
    4. DChunks are mainly required for scanned image files stored in stages
    Show answer and explanation
    Correct answer: B

    Chunks improve precision and fit context limits, which improves answers. Chunks do not enlarge documents, metadata remains useful, and chunking applies to text.

  4. Question 10

    Which two factors affect the quality of retrieval in a RAG system? (Choose two.)

    1. AHow the documents are chunked and enriched with metadata
    2. BThe colour of the search results page
    3. CThe quality of the embedding model for the content language
    4. DThe number of users logged in at the same time
    5. EThe name of the database schema
    Show answer and explanation
    Correct answer: A, C

    Chunking, metadata, and the embedding model's language quality strongly affect retrieval. Page colour, user counts, and schema names do not affect retrieval relevance.

  5. Question 11

    A developer builds a RAG app and notices that some answers cite outdated policies. Which fix addresses the root cause?

    1. AIgnore the policy changes and let the model answer from its cached documents
    2. BRefresh the search index as source documents change, and check the update schedule
    3. CIncrease the model's temperature and top-p for more varied answers
    4. DDisable retrieval so the model answers from its parametric memory alone
    Show answer and explanation
    Correct answer: B

    Stale answers come from an index that has not caught up with source changes, so refreshing and checking schedules fixes the cause. Temperature, disabling retrieval, and ignoring changes do not.

  6. Question 12

    Which measure best tells the team whether the retriever returns the right passages for test questions?

    1. ARecall of relevant passages in the top results
    2. BThe number of columns in the table holding the test set
    3. CThe average length of the Cortex function names in use
    4. DThe total number of virtual warehouses in the account
    Show answer and explanation
    Correct answer: A

    Recall in the top results measures whether relevant passages are retrieved, which is the retriever's job. Warehouse counts, column counts, and name lengths do not measure retrieval quality.

Generative AI governance and security (6 questions)

  1. Question 13

    A company must ensure that a Cortex AI function does not expose data outside the organization's region. Which consideration applies?

    1. AUse whichever region the developer prefers when setting up the Cortex function
    2. BChoose a model and deployment available in a compliant region and review data handling terms
    3. CIgnore regions because Cortex models are global and served from one shared endpoint in each account
    4. DCopy the data to a public region so that a cross-region replication job can serve analysts much faster
    Show answer and explanation
    Correct answer: B

    Choosing compliant regions and reviewing data handling terms addresses residency and privacy. Ignoring regions, copying data publicly, and personal choices violate that requirement.

  2. Question 14

    Which access control approach protects data used by Cortex functions?

    1. AGranting every analyst the ACCOUNTADMIN role so they can manage grants
    2. BSharing one service login for the analytics team
    3. CRole-based privileges on the source data and the functions
    4. DDisabling access control for AI workloads across the warehouse
    Show answer and explanation
    Correct answer: C

    Role-based privileges on sources and functions enforce least privilege for AI workloads. Admin rights for all, shared logins, and disabled controls expose data.

  3. Question 15

    A team wants to make sure that prompts and responses for an AI application are auditable. Which practice supports this?

    1. APrint every response to a paper log binder kept by the support team
    2. BDelete query history nightly after each call
    3. CLog function calls and retain query history for review
    4. DKeep responses in a chat window session and rely on the transcript
    Show answer and explanation
    Correct answer: C

    Logging function calls and retaining query history supports audit and investigation. Deleting history, chat-only storage, and printing are not reliable audit trails.

  4. Question 16

    A business wants to prevent a model from answering questions about topics it is not approved to discuss. Which control addresses this?

    1. AGuardrails or instructions that restrict allowed topics, with testing
    2. BRemoving the system instructions from the prompt so the model relies on defaults
    3. CAllowing unrestricted answers to reduce latency during peak hours
    4. DHoping the model refuses by itself without extra guardrails or instructions
    Show answer and explanation
    Correct answer: A

    Guardrails and restricting instructions, validated by testing, limit topics in a controlled way. Hoping for refusals, unrestricted answers, and removing instructions do not enforce the boundary.

  5. Question 17

    Which two practices help protect sensitive data sent to language models? (Choose two.)

    1. AMinimize sensitive fields in prompts to what the task requires
    2. BMask or exclude personal identifiers before calling the model
    3. CSend entire customer records to the model for every request
    4. DStore prompts in a public location for review
    5. EShare model responses with all staff regardless of permissions
    Show answer and explanation
    Correct answer: A, B

    Minimizing and masking sensitive fields reduces exposure in model calls. Sending entire records, public prompt storage, and unrestricted sharing increase risk.

  6. Question 18

    An AI application must restrict who can call a specific Cortex function in production. Which control applies?

    1. ADisable access logging for the function in production
    2. BGrant the function privilege only to the roles that need it
    3. CGrant it to PUBLIC so that every role in the account can call the function
    4. DShare a single service login for the application pool
    Show answer and explanation
    Correct answer: B

    Granting function privileges only to necessary roles enforces least privilege. Public grants, shared logins, and disabled logging remove the restriction or visibility.

Cortex Analyst and agents (6 questions)

  1. Question 19

    Business users want to ask natural-language questions of a governed semantic model and receive SQL-backed answers. Which Snowflake capability fits?

    1. AAsking an administrator to write each ad hoc SQL query
    2. BA printed data dictionary of table columns
    3. CExporting the tables to spreadsheets for offline review
    4. DCortex Analyst using a semantic model definition
    Show answer and explanation
    Correct answer: D

    Cortex Analyst translates natural language into SQL against a governed semantic model. Printed dictionaries, spreadsheet exports, and manual query requests do not provide governed self-service answers.

  2. Question 20

    Why is a semantic model used with Cortex Analyst?

    1. AIt defines business terms, relationships, and metrics so generated SQL is accurate
    2. BIt is a cosmetic colour theme that changes how results are displayed in Cortex Analyst sessions
    3. CIt replaces the warehouse entirely, so queries run on the semantic layer's own compute pool
    4. DIt stores the user's password and session token for authenticating queries
    Show answer and explanation
    Correct answer: A

    A semantic model maps business language to tables, joins, and metrics, which improves SQL accuracy. Passwords, warehouse replacement, and colour themes are not its purpose.

  3. Question 21

    An agent must answer questions and also take an action in an external system when the user approves it. Which design is appropriate?

    1. AAn agent with defined tools and a human approval step before the action runs
    2. BA chatbot with no tools that claims to act on the user's behalf in CRM
    3. CA manual process where staff send an email request for every action taken
    4. DAn agent that acts on every request automatically without approval
    Show answer and explanation
    Correct answer: A

    Defined tools with human approval keep actions controlled and accountable. Unapproved actions, tool-less claims, and manual emails do not provide a governed agent.

  4. Question 22

    A natural-language question about revenue returns a result that differs from the finance report. What should the team check first?

    1. AThe display name of the chat window and the title of the conversation thread in the interface
    2. BThe colour and font size of the answer bubble, along with the icon and the background theme chosen
    3. CThe length of the question text and the number of words the user typed
    4. DThe semantic model's metric definitions and filters compared with the finance definition
    Show answer and explanation
    Correct answer: D

    Metric definitions and filters in the semantic model are the usual source of mismatches. Colour, question length, and window names do not explain numeric differences.

  5. Question 23

    A business wants Cortex Analyst to answer questions about a new table. What must be updated first?

    1. AThe semantic model should describe the new table, its columns, and relationships
    2. BIncreasing the warehouse size to a larger tier and enabling multi-cluster scaling for the team
    3. CThe user's password and the multi-factor authentication settings for login
    4. DNothing, because the analyst reads the tables automatically without definitions
    Show answer and explanation
    Correct answer: A

    The semantic model must describe new tables and relationships so the analyst can use them correctly. Automatic reading without definitions, warehouse size, and passwords are not the requirement.

  6. Question 24

    An agent must follow a multi-step plan to answer a complex question, using several data sources. Which design element helps keep the steps organized?

    1. ARandom tool selection on each turn without a plan, so the agent picks functions at random
    2. BAsking the user to carry out each step manually in the chat window
    3. CDefined tools for each data source and an orchestration plan the agent follows
    4. DA single prompt with no tools and no intermediate state, asking the model to reply in one pass
    Show answer and explanation
    Correct answer: C

    Defined tools and an orchestration plan organize multi-step reasoning across data sources. A single prompt without tools, random selection, and manual steps do not provide structured orchestration.

Generative AI fundamentals (6 questions)

  1. Question 25

    A team wants a language model to answer questions using its own sales documents stored in Snowflake. Which pattern is most appropriate?

    1. AAsking the model to recall the sales documents from its pretraining data at answer time
    2. BRetrieval-augmented generation that searches the documents and passes relevant passages to the model
    3. CCopying the text of every sales document into the model prompt on each request, regardless of relevance to the question
    4. DFine-tuning a model on the company documents before it answers each incoming user question
    Show answer and explanation
    Correct answer: B

    Retrieval-augmented generation finds relevant passages and includes them in the prompt, which keeps answers grounded and current. Fine-tuning per question, full-document prompts, and relying on training memory do not meet the need efficiently or reliably.

  2. Question 26

    Which term describes a numeric representation of text that captures its meaning, used for similarity search?

    1. AData mask
    2. BPartition key
    3. CChecksum
    4. DEmbedding
    Show answer and explanation
    Correct answer: D

    Embeddings are vectors that capture semantic meaning so similar texts can be found. Checksums verify integrity, partition keys organize storage, and data masks hide values.

  3. Question 27

    What is a token in the context of language models?

    1. AA login credential like a username and password for the warehouse
    2. BA unit of text, such as a word or word fragment, processed by the model
    3. CA table partition that splits rows into storage segments by date
    4. DA billing invoice line item for monthly storage charges
    Show answer and explanation
    Correct answer: B

    Tokens are the text units models read and generate, and they drive input and output limits and cost. Credentials, partitions, and invoice lines are unrelated.

  4. Question 28

    A chatbot gives confident answers that include facts not present in the source documents. What is this called?

    1. AHallucination
    2. BOverfitting of the warehouse
    3. CIndex fragmentation
    4. DQuery caching
    Show answer and explanation
    Correct answer: A

    Hallucination describes fluent output that is unsupported by sources or facts. Overfitting relates to training, index fragmentation to storage, and caching to performance.

  5. Question 29

    Which technique reduces hallucinations by requiring the model to answer only from retrieved context and say when the context is insufficient?

    1. AIncreasing the sampling temperature to its maximum value
    2. BAsking the model to be creative and imaginative in phrasing
    3. CRemoving the retrieved context passages from the user prompt
    4. DGrounding the prompt in retrieved context with explicit instructions
    Show answer and explanation
    Correct answer: D

    Grounding with explicit instructions limits answers to supplied context and encourages honest uncertainty. Higher temperature, removed context, and creative prompts increase hallucination risk.

  6. Question 30

    Which two statements about embeddings are correct? (Choose two.)

    1. ATexts with similar meaning have vectors that are close in the embedding space
    2. BEmbeddings must be decrypted before similarity can be computed
    3. CEmbeddings are only useful for images
    4. DChanging the embedding model can change similarity results
    5. EEmbeddings are the same as primary keys
    Show answer and explanation
    Correct answer: A, D

    Similar meanings produce nearby vectors, and different embedding models produce different spaces, so results can change. Embeddings do not need decryption to compare, they apply to text as well as other data, and they are not primary keys.