CertArea
All Snowflake exams

Free SPS-C01 practice test: 30 questions with answers

30 original questions for SnowPro Specialty: Snowpark, 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

SPS-C01 exam guide: topics, cost and format

All 30 SPS-C01 questions with answers

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

Generative AI fundamentals (6 questions)

  1. Question 1

    A team wants a large language model to answer questions using only company documents, reducing made-up answers. Which pattern is most appropriate?

    1. ALowering the context window to zero tokens so the model draws on its pretrained weights, which keeps replies consistent
    2. BAsking the model to answer from its pretrained weights, with no source text
    3. CRetrieval-augmented generation that retrieves relevant passages and passes them to the model
    4. DFine-tuning the model on the company's logo and brand colour palette to improve replies
    Show answer and explanation
    Correct answer: C

    Retrieval-augmented generation grounds answers in retrieved company content. Logos, zero context, and unsourced answers do not ground the model.

  2. Question 2

    A developer notices that a model's answers vary between runs. Which inference setting most directly reduces this variation?

    1. ALowering the temperature
    2. BRemoving the system prompt entirely
    3. CIncreasing the temperature
    4. DAdding more random seeds to the prompt
    Show answer and explanation
    Correct answer: A

    Lower temperature makes outputs more deterministic and less varied. Higher temperature increases variation, and extra random seeds in the prompt are not a control.

  3. Question 3

    What does a token represent in a large language model's input and output?

    1. AA complete source document, like a PDF, that the model reads in one request
    2. BA unit of text, such as a word piece, that the model processes
    3. CA single letter of the alphabet, such as a lowercase k, used as an input unit
    4. DA user account record holding login credentials and role grants for staff
    Show answer and explanation
    Correct answer: B

    Tokens are sub-word units of text that models process and that drive context length and cost. Documents, single characters only, and user accounts are not tokens.

  4. Question 4

    A prompt is longer than the model's context window. What is the most likely result?

    1. AThe prompt is stored permanently in the model's weights, so later requests can reference it
    2. BThe model automatically grows its context window to fit the prompt by allocating extra memory
    3. CThe model answers using just the first word of the prompt and ignores later tokens
    4. DThe input must be shortened or split, because the model cannot attend to all of it at once
    Show answer and explanation
    Correct answer: D

    Inputs beyond the context window must be shortened, split, or retrieved selectively. Models do not expand automatically, and the other statements are incorrect.

  5. Question 5

    Which description best matches the term hallucination in the context of language models?

    1. AA system error that stops the model from running after an out-of-memory fault
    2. BA type of database index that speeds up lookups on a column
    3. CA confident output that is not supported by the source data or facts
    4. DA deliberate narrative feature that adds invented characters to a story
    Show answer and explanation
    Correct answer: C

    Hallucination is an unsupported output presented confidently. The other options describe unrelated features or errors.

  6. Question 6

    Which two techniques help reduce hallucinations in a generative AI application? (Choose two.)

    1. AGround answers in retrieved, authoritative sources
    2. BAsk the model to cite the passages its answer relies on
    3. CIncrease the temperature to make responses more creative
    4. DRemove all instructions so that the model is free to answer
    5. EUse the model's memory as the only source of truth
    Show answer and explanation
    Correct answer: A, B

    Grounding and citations tie answers to sources and make errors visible. Higher temperature, removing instructions, and relying on memory alone increase hallucination risk.

Snowflake Cortex AI functions (6 questions)

  1. Question 7

    A data analyst wants to classify customer support messages by sentiment directly in SQL, without training a model. Which Snowflake capability is most appropriate?

    1. ACreating a materialised view of the messages that refreshes on a schedule
    2. BA Cortex LLM function such as SENTIMENT or a COMPLETE prompt applied in the query
    3. CWriting a custom neural network in a spreadsheet macro to score each row
    4. DExporting the messages to a plain text file for manual reading by support team
    Show answer and explanation
    Correct answer: B

    Cortex LLM functions run inside SQL without model training. Spreadsheets, manual reading, and materialised views do not perform sentiment analysis.

  2. Question 8

    A team needs a summary of long text fields in a table, produced by a model and stored alongside the rows. Which approach fits Snowflake Cortex?

    1. AEmail each long text field, row by row, to an external contractor who writes a summary in a spreadsheet
    2. BTruncate each text value to its first ten characters, discarding the remainder before the table
    3. CCall a Cortex text function such as SUMMARIZE in a query that writes its output to a table
    4. DStore the full text in a comment attached to the table as metadata for later review by the analysts who query it
    Show answer and explanation
    Correct answer: C

    Calling a Cortex text function in a query writes the summaries to a table in-platform. Email, truncation, and table comments do not produce model-generated summaries.

  3. Question 9

    Which Snowflake capability is designed to extract text and structured fields from documents such as invoices and forms?

    1. ADynamic tables
    2. BTime Travel
    3. CResource monitors
    4. DDocument AI
    Show answer and explanation
    Correct answer: D

    Document AI extracts values from documents such as invoices. Time Travel, resource monitors, and dynamic tables do not extract document content.

  4. Question 10

    A chatbot must search a large set of company documents by meaning rather than exact keywords. Which Snowflake service supports this?

    1. AA masking policy that replaces sensitive column values with redacted output for unauthorised roles
    2. BCortex Search, which provides hybrid semantic and keyword search over indexed text
    3. CA file format that compresses the documents into columnar blocks for faster bulk loading
    4. DA standard B-tree index on the document title column, which matches exact string values
    Show answer and explanation
    Correct answer: B

    Cortex Search provides semantic and keyword search over indexed text for retrieval. Title indexes match exact text, file formats compress data, and masking policies restrict visibility.

  5. Question 11

    A business user wants to ask questions in plain English about governed sales data and receive SQL-backed answers. Which Snowflake capability fits?

    1. ACortex Analyst using a semantic model that describes the data
    2. BGranting the user the ACCOUNTADMIN role so they can open the sales tables
    3. CA static PDF export of the sales schema diagram, saved for reference
    4. DPrinting each sales record to paper for a manual reconciliation spreadsheet
    Show answer and explanation
    Correct answer: A

    Cortex Analyst translates natural-language questions into queries over a semantic model. PDFs, printing, and admin roles do not answer questions over governed data.

  6. Question 12

    Which two practices help control the cost of running LLM functions over a large table in Snowflake? (Choose two.)

    1. ARun the model only on rows that need it, using filters
    2. BTest the prompt on a small sample before running it over the full table
    3. CCall the model for every row in the table, including duplicates, every time
    4. DUse the largest model for every task regardless of need
    5. EDisable cost monitoring for AI workloads
    Show answer and explanation
    Correct answer: A, B

    Filtering rows and sampling first control spend on model calls. Calling on duplicate rows, always using the largest model, and disabling cost monitoring increase cost.

Retrieval and grounding (6 questions)

  1. Question 13

    In a retrieval-augmented generation pipeline, what is the role of the embedding step?

    1. AIt converts text into numerical vectors so that similar passages can be found by meaning
    2. BIt deletes duplicate documents from the corpus before indexing, keeping the newest revision of each file
    3. CIt translates the generated answer into another language using a separate machine translation model
    4. DIt sets the user's row-level access permissions on the index before each search request is run
    Show answer and explanation
    Correct answer: A

    Embeddings convert text to vectors that capture meaning for similarity search. Translation, deletion, and permissions are different functions.

  2. Question 14

    A RAG system returns irrelevant passages, and the answers are poor. Which change is most likely to improve retrieval quality?

    1. AAsk users to write longer, more conversational queries with no specific terms or product names
    2. BIncrease the model's temperature to maximum so the retriever explores more varied passages
    3. CRemove the source documents from the index after each refresh to keep storage costs low
    4. DImprove how documents are chunked and the quality of the embeddings and ranking
    Show answer and explanation
    Correct answer: D

    Chunking, embedding quality, and ranking drive retrieval relevance. Raising temperature, removing sources, and vaguer queries make answers worse.

  3. Question 15

    A RAG pipeline splits long documents into chunks. What is a common reason for overlapping chunks?

    1. ATo prevent search results from being returned when the query has a stop word
    2. BTo preserve context that spans the boundary between two chunks
    3. CTo remove the need for metadata fields by embedding attributes in text
    4. DTo make the index twice as large for no clear reason
    Show answer and explanation
    Correct answer: B

    Overlap keeps context that crosses chunk boundaries available for retrieval. Enlarging the index without reason, removing metadata, and preventing results are not purposes of overlap.

  4. Question 16

    A RAG application must return only passages the current user is allowed to see. Where should this control be enforced?

    1. AIn the model's system prompt, by instructing it to hide passages it judges as sensitive
    2. BIn the browser, by hiding the returned results with CSS display rules
    3. CIn the retrieval layer, using the data's access policies for the requesting user
    4. DNowhere in the stack, because the model itself will decide what is visible
    Show answer and explanation
    Correct answer: C

    Enforcing access in the retrieval layer means unauthorised passages never reach the model. Prompt instructions, CSS hiding, and model judgement do not enforce access.

  5. Question 17

    A search over documents returns relevant results only when users type the exact product code. What capability would improve matching of natural-language queries?

    1. AIncreasing the number of exact-match filters applied before ranking
    2. BRemoving the body text from the index and keeping just document titles
    3. CSemantic search using embeddings alongside keyword matching
    4. DStoring the documents as scanned page images with no OCR text layer
    Show answer and explanation
    Correct answer: C

    Semantic search with embeddings matches meaning, and combining it with keywords handles exact codes too. Removing text, images only, and more exact-match filters do not improve natural-language matching.

  6. Question 18

    Which two factors most affect the quality of a RAG system's answers? (Choose two.)

    1. AHow documents are chunked and enriched with metadata
    2. BThe relevance of the retrieved passages to the question
    3. CThe colour of the user interface
    4. DThe number of users who have logged in this month
    5. EThe length of the application's name
    Show answer and explanation
    Correct answer: A, B

    Chunking, metadata, and retrieval relevance strongly shape answer quality. The interface colour, user counts, and application name do not affect the answers.

Responsible AI and governance (6 questions)

  1. Question 19

    A company wants to prevent Cortex AI features from exposing sensitive data to users who are not authorised to see it. Which control applies to the data the model can access?

    1. ARole-based access control and policies on the underlying tables and documents
    2. BSharing one administrator login with the AI team so its service account can read every table
    3. CHoping the model will refuse sensitive questions and keep figures private
    4. DRemoving row-level security from the source data tables for the AI feature's own service role
    Show answer and explanation
    Correct answer: A

    Access controls and policies on underlying data limit what the AI can retrieve for each user. Relying on the model to refuse, removing security, and shared logins do not enforce access.

  2. Question 20

    A company must document how its AI feature uses customer data. Which artifact best serves this need?

    1. AA marketing slogan and launch brochure for the feature, written for prospective customers
    2. BA list of the developers' favourite tools and IDE plug-ins used during model training and evaluation
    3. CThe model's temperature and top-p sampling settings for each deployment environment in production
    4. DA data and model use record that describes the data sources, purposes, and controls
    Show answer and explanation
    Correct answer: D

    A record of data sources, purposes, and controls supports transparency and compliance. A slogan, a single setting, and tool preferences do not document data use.

  3. Question 21

    Which two practices help a team evaluate the safety of a generative AI feature before release? (Choose two.)

    1. ATest the feature with adversarial and edge-case prompts
    2. BMeasure how often the feature gives unsupported answers on a representative set
    3. CRelease the feature to all users and monitor only complaints
    4. DSkip testing because the model vendor is reputable
    5. ERemove logging so that unsafe outputs are not recorded
    Show answer and explanation
    Correct answer: A, B

    Adversarial testing and measured evaluation on representative data reveal safety issues before release. Releasing without testing, skipping tests on vendor trust, and removing logs hide problems.

  4. Question 22

    A company wants its AI feature to avoid producing harmful content. Which layer of protection is most effective when combined with the others?

    1. AInput and output filtering alongside clear system instructions and monitoring
    2. BTurning off request and response logging for the feature to reduce storage costs at scale
    3. CRemoving content review from the inference pipeline to cut response time
    4. DRelying on users to report harmful content through the feedback form
    Show answer and explanation
    Correct answer: A

    Layered controls with filtering, instructions, and monitoring reduce harmful content. Relying on reports, disabling logs, and removing review leave gaps.

  5. Question 23

    A user types an instruction that tries to override the application's rules, such as ignore your previous instructions. What is this attack called?

    1. AIndex scanning
    2. BPrompt injection
    3. CQuery caching
    4. DSchema evolution
    Show answer and explanation
    Correct answer: B

    Prompt injection attempts to override an application's instructions through user input. The other terms refer to database or platform features.

  6. Question 24

    Which control reduces the risk that a prompt injection causes a model-driven tool to take an unauthorised action?

    1. ARemoving the tool's logging so calls are not recorded for auditors
    2. BGiving the tool full administrator rights so it can act on other users
    3. CTrusting the model's output without checks before running SQL
    4. DLimiting the tool's permissions and requiring approval for sensitive actions
    Show answer and explanation
    Correct answer: D

    Least-privilege tools and approval for sensitive actions limit the harm an injection can cause. Admin rights, disabled logging, and unchecked outputs increase risk.

Cost and performance (6 questions)

  1. Question 25

    A generative AI feature is expensive to run because it calls a large model for simple classification tasks. What is the best change?

    1. AUse a smaller model for simple tasks and reserve the large model for complex ones
    2. BCall the model twice for every request and keep the second answer as a backup
    3. CRemove the classification feature from the product and its dashboards
    4. DUse the largest model for every task to be safe and very accurate
    Show answer and explanation
    Correct answer: A

    Routing simple tasks to a smaller model reduces cost while keeping quality for complex tasks. The largest model everywhere, duplicate calls, and removing the feature are poor responses.

  2. Question 26

    A team caches the answers to repeated identical questions in a generative AI application. What is the main benefit?

    1. ALower cost and latency for repeated requests, provided the cache is refreshed when the data changes
    2. BHigher model temperature, so that answers vary more across repeated identical questions
    3. CFewer security controls because cached answers skip role checks across user sessions
    4. DThe model learns from every cached answer permanently and stores them in its weights
    Show answer and explanation
    Correct answer: A

    Caching lowers cost and latency for repeats, but stale answers must be refreshed. Temperature, weaker security, and permanent learning are not benefits of caching.

  3. Question 27

    A batch job uses an LLM to process a large table overnight, and the results are needed by morning. What is the most important planning factor?

    1. ANaming the job with a long title so that dashboards and schedulers can list it under the correct owner
    2. BChoosing a colour for the output table and a font for the report so that the table looks consistent
    3. CRunning the full batch without a trial run on a small sample of rows first
    4. DEstimating the token volume and the compute or credit cost before running the full job
    Show answer and explanation
    Correct answer: D

    Estimating tokens and cost before a large run prevents surprises and supports planning. Colours, names, and untested runs are not planning factors.

  4. Question 28

    Which two actions help reduce the latency of a RAG application? (Choose two.)

    1. ARetrieve fewer but more relevant passages
    2. BCache frequent queries or embeddings where the data allows
    3. CSend the entire document collection to the model with each question
    4. DUse the largest model for the shortest possible answers
    5. EDisable retrieval to avoid the index lookup
    Show answer and explanation
    Correct answer: A, B

    Fewer relevant passages and caching shorten the path to an answer. Sending everything, using the largest model for short answers, and dropping retrieval reduce quality or increase latency.

  5. Question 29

    A team compares the cost of two designs for the same feature: one calls the model for every row, and the other calls the model only for new or changed rows. Which design is usually cheaper and why?

    1. AThey cost the same because the model provider bills per table scanned, so row count barely matters
    2. BThe incremental design costs more because it adds extra pipeline steps
    3. CThe incremental design, because it avoids repeating model calls on unchanged data
    4. DThe every-row design, because it uses fewer tokens per call overall
    Show answer and explanation
    Correct answer: C

    Incremental processing avoids repeated model calls on unchanged rows, reducing tokens and cost. The other statements misstate the cost drivers.

  6. Question 30

    A Cortex function processes a large table, and the job is slow. Which change is most likely to improve throughput without changing the results?

    1. APartition the work and run it on a suitably sized warehouse where the function supports it
    2. BRun the job during the busiest hour of the day, when interactive dashboards compete
    3. CProcess one row per minute by calling the function serially on a small warehouse
    4. DDisable the warehouse and rely on the result cache alone, which refreshes weekly
    Show answer and explanation
    Correct answer: A

    Partitioning and suitable compute improve throughput without changing outputs. Slow single-row processing, busy-hour scheduling, and disabling the warehouse do not help.