Free GES-C02 practice test

Cortex Analyst converts natural language into SQL against a governed semantic model to compute answers from structured tables, which is exactly this scenario. Cortex Search retrieves text, not computed aggregates. Calling AI_COMPLETE directly on rows would not reliably compute an aggregate like revenue by region. CoCo is a coding/admin assistant, not a business-question tool.

AI Studio is the low-code Snowsight experience for experimenting with prompts, Search services, and semantic models without writing code. Snowpark Python is a coding entry point for developers, not a no-code experience. The REST API is meant for external application integration, not in-Snowsight experimentation. A SQL stored procedure still requires writing code.

CoWork is the conversational work agent over governed enterprise data, aimed at non-SQL business users — matching this scenario exactly. CoCo is aimed at developers/admins writing and debugging code, the opposite audience. Snowpark Container Services hosts custom model workloads; it isn't a conversational agent. Model Registry manages custom models; it's also not a conversational business-user tool.

Cortex Code was renamed to Snowflake CoCo, the coding/administration agent. Snowflake CoWork is the rename of a different product, Snowflake Intelligence. Cortex AI Functions is the rename of Cortex AISQL, an unrelated capability. Cortex Agents is a distinct orchestration capability, not a renamed product.

TRANSLATE converts text between languages, exactly this use case. SENTIMENT scores tone, AI_CLASSIFY assigns categories, and EXTRACT_ANSWER answers a question from a passage — none translate.

EXTRACT_ANSWER returns the answer to a question given a source passage, ideal for pulling a specific fact from provided text. SENTIMENT scores tone, EMBED_TEXT_768 makes vectors, and COPY INTO loads data.

AI_AGG applies an LLM instruction across many rows of text and returns an aggregated result (e.g., a synthesized summary of complaints). COUNT is numeric aggregation, EMBED_TEXT_1024 makes vectors, and SENTIMENT scores single texts.

Cortex Agents most commonly combine Cortex Search (unstructured retrieval) and Cortex Analyst (structured computation) in one request. Fine-tuning adapts a model ahead of time, Autopilot helps author semantic views, and EMBED_TEXT_1024 is a low-level embedding function — none are the tools an Agent typically calls at request time.

EXTRACT_ANSWER performs extractive question-answering over a given passage. AI_COMPLETE needs a real prompt to be useful, SUMMARIZE condenses rather than answers, and AI_CLASSIFY assigns categories.

Cortex Search is a managed retrieval service providing hybrid (semantic + keyword) search over Snowflake data, purpose-built to back RAG applications without manually managing embeddings/indexes. The other options don't provide retrieval.

AI_FILTER's natural-language row-by-row boolean predicate has no older-style equivalent. AI_SENTIMENT, AI_CLASSIFY, and AI_COMPLETE each map to an older-style name (SENTIMENT, CLASSIFY_TEXT, and COMPLETE respectively).

CORTEX_ENABLED_CROSS_REGION controls whether Cortex may process a request in a different region when the model isn't available locally. Setting it appropriately enforces data-residency requirements. The other parameters are unrelated to cross-region inference.

Revoking SNOWFLAKE.CORTEX_USER is the direct, additive way to cut off Cortex access without altering table grants. Revoking SELECT on every table would remove far more than Cortex access and would break other unrelated workloads. Disabling warehouses would stop essentially all query activity, not just Cortex calls, and is a disproportionate action. Emptying the allowlist for the whole account would affect every role's Cortex access, not just the target role, which is broader than intended.

With cross-region inference disabled, an out-of-region model request fails outright — this fail-closed behavior is what protects residency. Automatic rerouting despite the setting being disabled would defeat the purpose of the control entirely. There's no described local-cache fallback that bypasses the region restriction. Indefinite queuing isn't the documented behavior; the call fails rather than waiting.

Cortex Guard screens model responses for unsafe or harmful content, adding a safety layer to generation. Time Travel, cloning, and search optimization address data recovery and query performance, not response safety.

Both statements accurately reflect the guide: Cortex is serverless credit consumption needing normal cost guardrails, and budgets offer finer scoping than an account-wide resource monitor. Access grants like SNOWFLAKE.CORTEX_USER are an access-control mechanism, not a cost-governance one, so cost governance is not achieved exclusively through it. Cortex spend can in fact be monitored and attributed through usage/metering views, contradicting the claim that it cannot be tracked at all.

The guide is explicit that RBAC, masking, and row access policies are inherited unchanged by Cortex pipelines — the same tables, same governance objects. Resource monitors address cost, not data-level governance, and aren't the only relevant control for a pipeline reading sensitive tables. The allowlist restricts model choice; it doesn't supersede or replace table-level access controls. It's explicitly false that existing controls stop applying just because a table feeds an AI pipeline — that's the exact misconception the guide warns against.

Document AI's model-training UI has been retired; the current, exam-relevant surface is the SQL-callable Cortex functions that accomplish the same goal. It hasn't merely been renamed to AI_EXTRACT, and there's no support-access workaround to revive a retired UI.

A document RAG pipeline parses files to text, chunks it, embeds the chunks, indexes/retrieves the relevant ones (Cortex Search or vector similarity), and passes them to COMPLETE as grounding context. The other orderings are nonsensical.

PARSE_DOCUMENT (AI_PARSE_DOCUMENT) extracts text from documents such as PDFs, including OCR and layout awareness, producing text you can then chunk and embed. The other functions score sentiment, compare vectors, and load data respectively.