Snowflake GES-C02 Practice Exam
SnowPro Specialty: Gen AI GES-C02
The SnowPro Specialty: Gen AI certification validates the skills to build generative AI solutions on Snowflake: applying Snowflake Cortex AI functions and LLMs to business use cases, building retrieval with Cortex Search, governing AI cost and access, and parsing documents into structured data. The GES-C02 exam covers four domains: Snowflake for Gen AI Overview (18%), Snowflake Gen AI Functions (38%), Snowflake Gen AI Governance (29%) and Snowflake Document Processing (15%). This practice exam mirrors its scenario style with single-answer and multiple-select questions.
585 questions in the bank
Question bank updated : questions, answers, explanations and chapters.
Certification Overview
- Exam name: SnowPro Specialty Gen AI GES-C02
- Vendor: Snowflake
- Exam code: GES-C02
- Duration: 85 minutes
- Total questions: about 55
- Official passing score: 750 on a scale of 0-1000 (source)
- Our practice exam pass mark: 75%
Who Should Take This Exam?
AI Engineer, ML Engineer, Data Scientist, Data Engineer, Data Application Developer, Data Analyst, Snowflake Developer, Analytics Engineer
Prerequisites
No formal prerequisite. Snowflake recommends 1 or more years of Gen AI experience with Snowflake in an enterprise environment, with SQL and data engineering knowledge and some Python.
Topics Covered
- Snowflake for Gen AI Overview
- Snowflake Gen AI Functions
- Snowflake Gen AI Governance
- Snowflake Document Processing
Question Types
- Multiple Choice (Single Answer)
- Multiple Choice (Multiple Answers)
SnowPro Specialty Gen AI GES-C02 Practice Questions
Our question bank contains 585+ practice questions for this certification. Sample questions from each exam chapter. Expand a question to see the answer choices. With full access, you get unlimited practice exams with randomized questions from our full question bank.
Snowflake for Gen AI Overview
A data science team fine-tuned an open-source LLM outside Snowflake. It needs GPU inference, and SQL users must be able to call it. What should the team do?
- A. Upload the weights to an internal stage and pass the file path as the model argument of AI_COMPLETE
- B. Fine-tune a copy with Cortex Fine-tuning and call it through AI_COMPLETE
- C. Log it to the Model Registry and run it on a Snowpark-optimized warehouse
- D. Log it to the Model Registry and deploy it to a GPU compute pool in Snowpark Container Services Correct answer
Option D is correct because a model fine-tuned outside Snowflake is logged to the Model Registry, and deploying it to a GPU compute pool in Snowpark Container Services gives GPU inference that SQL users can call through the service function. Option A is wrong because AI_COMPLETE accepts only Cortex-hosted models, not staged weights. Option B discards the external model, and option C fails because Snowpark-optimized warehouses have no GPUs.
Regional sales managers use Snowflake Intelligence to ask questions about orders in a semantic view. Each manager must see rows for their own region only. Which control guarantees this?
- A. The CORTEX_ANALYST_USER database role granted to the regional manager roles
- B. A verified query for each region stored in the semantic view
- C. A row access policy on the underlying ORDERS table keyed to the querying role Correct answer
- D. A custom instruction in the semantic view that adds a region filter to each query
Option C is correct because a row access policy on the underlying table is enforced for every query, including the SQL that Cortex Analyst generates, so each manager sees only their own region. Option A only grants permission to use Cortex Analyst and filters nothing. Options B and D steer or instruct the model, which guides the SQL but does not guarantee row-level security.
A sales director asks one question: which three accounts have the highest ARR, and what do their latest call notes say about renewal risk? ARR sits in tables and the notes are PDFs on a stage. Which setup answers it in a single request?
- A. A Cortex Agent using a Cortex Analyst tool for ARR and a Cortex Search tool for the notes Correct answer
- B. A Cortex Search service built on the ARR table and queried with SEARCH_PREVIEW
- C. A single AI_EXTRACT query over the PDFs that also returns each account's ARR
- D. One Cortex Analyst request against a semantic view that also lists the PDF files as a column
Option A is correct because a Cortex Agent orchestrates tools: Cortex Analyst answers the structured ARR question and Cortex Search retrieves the unstructured call notes, all within one request. Option B fails because Cortex Search is built for unstructured text, not for ranking numeric tables. Option C cannot return ARR from PDFs, and option D cannot put files in a semantic view column.
A news publisher wants subscribers to question its licensed articles from their own Cortex Agents, without being able to read or copy the raw articles. Which feature is built for this?
- A. A reader account provisioned for each subscriber with SELECT on the articles
- B. A Cortex Knowledge Extension listed on Snowflake Marketplace Correct answer
- C. A secure share of the articles table, with each subscriber building a Cortex Search service
- D. A Cortex Analyst semantic view published as a listing
Option B is correct because a Cortex Knowledge Extension lets a publisher list content on Snowflake Marketplace so subscribers' Cortex Agents can query it without access to the raw articles. Options A and C expose the underlying articles to the subscriber. Option D shares a semantic view over structured data and is not designed to protect licensed text content.
Cortex Analyst misreads "fiscal quarter" in all of its answers because the company's fiscal year starts on February 1. It also writes slow, incorrect SQL for the recurring question about accounts at churn risk, which the data team has already hand-tuned. Which TWO changes fix both problems? (Select TWO).
- A. Fine-tune a base model on questions that mention fiscal quarters
- B. State the February 1 fiscal year start in the semantic view's custom instructions Correct answer
- C. Add the churn-risk SQL to the semantic view's custom instructions
- D. Save a verified query that hard-codes the current fiscal quarter's date range
- E. Save the hand-tuned churn-risk SQL as a verified query Correct answer
Options B and E are correct. Stating the February 1 fiscal year start in the semantic view's custom instructions fixes the fiscal quarter reading in all answers, and saving the hand-tuned churn-risk SQL as a verified query makes Cortex Analyst reuse it for that recurring question. Option A is unnecessary, option C puts a full query into general instructions, and option D hard-codes a date range that goes stale.
Snowflake Gen AI Functions
To cut storage and search cost, a team keeps 768-dimension embeddings from snowflake-arctic-embed-m-v1.5 as 256-dimension vectors and ranks them with VECTOR_INNER_PRODUCT so that the order matches cosine similarity. Which TWO steps give correct rankings? (Select TWO).
- A. Wrap each VECTOR_TRUNCATE(embedding, 256) result in VECTOR_NORMALIZE before storing it Correct answer
- B. Truncate the query embedding to 256 dimensions and normalize it before comparing Correct answer
- C. Sort the VECTOR_INNER_PRODUCT results in ascending order once the vectors are truncated
- D. Normalize the full 768-dimension vector once and keep that unit length after truncating it
- E. Cast the stored column to VECTOR(FLOAT, 256) so Snowflake drops the trailing elements
Options A and B are correct because truncating a unit vector breaks its unit length, so the stored vectors must be re-normalized after VECTOR_TRUNCATE and the query vector must get the same truncate-and-normalize treatment; only then does the inner product order match cosine. Normalizing the full vector first leaves the truncated vector shorter than 1, and an ascending sort or a cast does not restore the ranking.
Each support email may touch billing, legal and technical matters at once, and a report needs one output row per email and label pair. Which FROM clause produces those rows?
- A. FROM emails, LATERAL FLATTEN(input => AI_CLASSIFY(body, ['billing','legal','technical'], {'output_mode':'multi'}))
- B. FROM emails, LATERAL FLATTEN(input => AI_CLASSIFY(body, ['billing','legal','technical']):labels)
- C. FROM emails, LATERAL FLATTEN(input => AI_CLASSIFY(body, ['billing','legal','technical'], {'output_mode':'multi'}):labels) Correct answer
- D. FROM emails, LATERAL FLATTEN(input => AI_CLASSIFY(body, ['billing','legal','technical'], {'task_description':'return all matching labels'}):labels)
Option C is correct because AI_CLASSIFY in multi mode returns an object whose labels array holds every matching label, so FLATTEN must read the :labels path to get one row per email and label. Option A flattens the whole object, B omits multi mode so only one label comes back, and D has no multi-label setting.
Product photos on a stage must be sorted into shoes, bags or hats for a resale site, and the images contain no readable text. Which call labels each photo?
- A. AI_PARSE_DOCUMENT(TO_FILE('@photos', photo_path), {'mode':'OCR'})
- B. AI_SENTIMENT(TO_FILE('@photos', photo_path), ['shoes','bags','hats'])
- C. AI_SIMILARITY(TO_FILE('@photos', photo_path), 'shoes, bags or hats')
- D. AI_CLASSIFY(TO_FILE('@photos', photo_path), ['shoes','bags','hats']) Correct answer
Option D is correct because AI_CLASSIFY accepts image files and assigns one of the supplied categories. AI_PARSE_DOCUMENT reads text, AI_SENTIMENT scores sentiment and AI_SIMILARITY returns a numeric score, so none of them returns a category label.
Which key of the AI_CLASSIFY config object lets the caller state, in one sentence, what the labelling task is for?
- A. examples
- B. task_description Correct answer
- C. output_mode
- D. description
Option B is correct because task_description is the AI_CLASSIFY config key for describing the labelling task. examples and output_mode are real keys with other jobs, and description does not exist.
Two stages of an archive job use AI_PARSE_DOCUMENT: 840,000 typed pages that need text only, and 60,000 pages with pricing tables whose rows and columns must survive. OCR mode costs 0.5 credits per 1,000 pages and LAYOUT mode costs 3.33. Which total is the lowest that meets both needs?
- A. About 450 credits
- B. About 420 credits
- C. About 3,000 credits
- D. About 620 credits Correct answer
Option D is correct because the cheapest plan that meets both needs is OCR for the 840,000 text-only pages (420 credits) and LAYOUT for the 60,000 table pages (about 200 credits), about 620 in total. Using LAYOUT for everything costs about 3,000 and OCR for everything loses the table structure.
Snowflake Gen AI Governance
Which values can the CORTEX_MODELS_ALLOWLIST account parameter take?
- A. 'All', 'None', or a comma-separated list of model names Correct answer
- B. 'Enabled', 'Disabled', or a list of region names
- C. 'All', 'None', or the name of one application role
- D. 'PUBLIC', 'None', or a comma-separated list of role names
Option A is correct because CORTEX_MODELS_ALLOWLIST accepts 'All', 'None', or a comma-separated list of model names. Region names, a single application role or role names are not valid values; role-level control is done with model application roles instead.
Setting the account-wide model list with ALTER ACCOUNT SET CORTEX_MODELS_ALLOWLIST requires which role?
- A. ACCOUNTADMIN Correct answer
- B. SECURITYADMIN
- C. SYSADMIN
- D. USERADMIN
Option A is correct because only ACCOUNTADMIN sets the CORTEX_MODELS_ALLOWLIST account parameter with ALTER ACCOUNT. SECURITYADMIN, SYSADMIN and USERADMIN manage grants, objects and users, not this account-level Cortex setting.
Before planning model grants, an administrator wants to see which model-specific application roles exist in the account. Which statement lists them?
- A. SHOW MODELS IN SNOWFLAKE.MODELS
- B. SHOW APPLICATION ROLES IN APPLICATION SNOWFLAKE Correct answer
- C. SHOW ROLES LIKE 'CORTEX-MODEL-ROLE%'
- D. SHOW DATABASE ROLES LIKE 'CORTEX%' IN DATABASE SNOWFLAKE
Option B is correct because the model roles are application roles of the SNOWFLAKE application, so SHOW APPLICATION ROLES IN APPLICATION SNOWFLAKE lists them. They are not account roles, database roles or models, so the other statements do not return them.
With CORTEX_MODELS_ALLOWLIST set to 'None', what is true of a role that holds only the SNOWFLAKE.CORTEX_USER database role?
- A. It can call every model, because CORTEX_USER overrides the allowlist
- B. It cannot call any model until it receives a model application role Correct answer
- C. It can call only models hosted in the account's own region
- D. It can call only models named in the CORTEX_USER database role grants
Option B is correct because with the allowlist at 'None' no model is open to everyone, and CORTEX_USER only grants the functions. The role needs a model application role such as CORTEX-MODEL-ROLE-ALL or a per-model role before any call succeeds. CORTEX_USER does not override the allowlist.
A review must show which users sent text from the column CLINIC.NOTES.BODY to Cortex functions last week. Starting from ACCOUNT_USAGE views, which TWO steps used together produce that list? (Select TWO).
- A. Filter TAG_REFERENCES for the column and read which users referenced the tag during the week
- B. Filter ACCESS_HISTORY on base columns that include CLINIC.NOTES.BODY to collect the matching query IDs Correct answer
- C. Join those query IDs to QUERY_HISTORY and keep statements that call a Cortex function to read user and role Correct answer
- D. Read CORTEX_FUNCTIONS_USAGE_HISTORY, which lists the source columns that were passed to each function call
- E. Query LOGIN_HISTORY for sessions that ran COMPLETE and match them to the table through the client IP address
Options B and C are correct because ACCESS_HISTORY records the base columns each query read, giving the query IDs that touched CLINIC.NOTES.BODY, and joining them to QUERY_HISTORY shows which of those queries called a Cortex function and which user and role ran them. TAG_REFERENCES and LOGIN_HISTORY say nothing about column reads, and CORTEX_FUNCTIONS_USAGE_HISTORY does not list source columns.
Snowflake Document Processing
Task TASK_P, owned by role PIPE_R, runs INSERT INTO parsed SELECT relative_path, AI_PARSE_DOCUMENT(TO_FILE('@DOCS', relative_path)) FROM S_DOCS. Role INGEST_R owns stream S_DOCS and the internal stage DOCS. PIPE_R holds EXECUTE TASK, warehouse USAGE, INSERT on parsed and SNOWFLAKE.CORTEX_USER, yet runs fail with insufficient privileges. Which TWO grants fix this? (Select TWO).
- A. OPERATE on stream S_DOCS to PIPE_R
- B. READ on stage DOCS to PIPE_R Correct answer
- C. OWNERSHIP of stage DOCS to PIPE_R
- D. SELECT on stream S_DOCS to PIPE_R Correct answer
- E. USAGE on stage DOCS to PIPE_R
Options B and D are correct. Reading an internal stage needs the READ privilege, and the task selects from the stream, which needs SELECT on it. USAGE is the privilege for external stages, OPERATE is not needed to read a stream, and OWNERSHIP is far more than the task needs.
Compliance staff without SnowSQL must add a few Word files from their desktops to the named internal stage HR_DOCS, and a developer needs scripted bulk loads from a laptop shell. Which TWO methods place local files on the stage? (Select TWO).
- A. Run COPY INTO @HR_DOCS from a file:// folder on the laptop
- B. Run PUT in a Snowsight SQL worksheet
- C. Run COPY FILES INTO @HR_DOCS from the laptop folder
- D. Run PUT through the Snowflake CLI Correct answer
- E. Upload through the stage's file upload dialog in Snowsight Correct answer
Options D and E are correct. The Snowflake CLI can run PUT from a laptop shell for scripted loads, and the file upload dialog of a stage in Snowsight lets staff without SnowSQL add files. PUT does not run in a Snowsight worksheet, COPY INTO loads from a stage into a table, and COPY FILES copies between stages.
An older view calls SNOWFLAKE.CORTEX.PARSE_DOCUMENT with a stage name, a relative path and {'mode': 'LAYOUT'}. Table CONTRACTS now stores a FILE value for each document in column DOC, and the view must read from that column while the rows and columns of tables inside the documents stay intact. Which call does that?
- A. AI_PARSE_DOCUMENT(doc, {'mode': 'LAYOUT'}) Correct answer
- B. AI_PARSE_DOCUMENT(doc)
- C. SPLIT_TEXT_RECURSIVE_CHARACTER(doc, 'markdown', 1500, 150)
- D. AI_PARSE_DOCUMENT(TO_VARCHAR(doc), {'mode': 'LAYOUT'})
Option A is correct. AI_PARSE_DOCUMENT takes a FILE value as its first argument and an options object, and mode LAYOUT keeps the table structure. Option B would use the default OCR mode, which loses the tables, option C only chunks text, and option D passes a string where a FILE is required.
Finance wants the plain text of each scanned letter at the lowest per-page rate, plus the number of pages billed for it, from one AI_PARSE_DOCUMENT call per file with the options left at their defaults apart from mode. Which approach delivers both?
- A. Use mode LAYOUT and read the content and metadata:pageCount fields of the result
- B. Use mode OCR and take the page count from a PAGES column of the directory table
- C. Use mode OCR and count the entries in the pages array of the returned result
- D. Use mode OCR and read the content and metadata:pageCount fields of the result Correct answer
Option D is correct. OCR mode has the lowest per-page rate, and the result carries the text in content and the number of pages in metadata:pageCount, so one call gives both. LAYOUT costs more per page, the directory table has no pages column, and the pages array only exists when page_split is enabled.
Scanned loan pages (200,000 of them) go through OCR mode for their text and then through LAYOUT mode to recover tables. OCR bills 0.5 credits and LAYOUT 3.33 credits per 1,000 pages. Which change cuts the cost and still delivers both text and tables?
- A. Run OCR mode once and read the tables from the metadata field of its result
- B. Run LAYOUT mode once and use its content for the text and the tables Correct answer
- C. Run OCR mode with page_split enabled so the pages that hold tables come back apart
- D. Run LAYOUT mode on a sample of files and OCR mode on the rest
Option B is correct. LAYOUT mode already returns the text along with the tables, so a single LAYOUT pass at 3.33 credits per 1,000 pages is cheaper than OCR plus LAYOUT. OCR mode does not return table structure in its metadata, page_split only separates pages, and sampling would leave most files without tables.
Free Study Guide
Read the free Snowflake GES-C02 study guide: every exam topic explained in plain English, with diagrams, exam tips and the mistakes to avoid.
More Snowflake Practice Exams
Practice with realistic mock exams to prepare for your Snowflake certification.