SnowPro Specialty: Gen AI (GES-C01) Practice Tests

Explanation

Supplying a JSON schema definition through the response_format parameter of SNOWFLAKE.CORTEX.COMPLETE enforces structural compliance at the model inference layer rather than in post-processing SQL. When a schema is provided through response_format, the model is constrained to produce output that matches the declared field names and data types on every invocation, eliminating the variability that prompt-only approaches introduce. This makes it the correct choice when zero data loss and zero malformed records are required. Relying on a detailed prompt instruction and filtering with TRY_COMPLETE does not guarantee schema compliance because the model can return syntactically valid JSON that omits required fields or uses incorrect types, and discarding those rows causes data loss in a pipeline that expects one output record per input transcript. Using EXTRACT_ANSWER three times per row is architecturally incorrect for this requirement. EXTRACT_ANSWER is designed to locate a specific answer span within a document context and always returns a text string; it cannot natively produce float or boolean values, so combining three separate extraction results with OBJECT_CONSTRUCT still depends on implicit type coercion through TRY_CAST that can fail silently and does not guarantee field presence. Combining CLASSIFY_TEXT for the categorical field with separate COMPLETE calls for the numeric and boolean fields introduces three separate inference calls per row, multiplying cost and latency, and still requires SQL string concatenation to assemble the final JSON object without any structural guarantee from the model layer.

Explanation

Streamlit in Snowflake does not automatically retain conversation history between user interactions. The developer must explicitly use Streamlit's session_state object to accumulate all prior messages and then pass the full conversation history to each COMPLETE call. COMPLETE has no enable_memory parameter — every call is stateless by design. Cortex Agents manage their own orchestration state for multi-step tool workflows but do not solve the fundamental Streamlit session management problem. Storing messages in a table is technically possible but does not trigger any automatic context injection — the developer would still need to query and format the history explicitly for each COMPLETE invocation.

Explanation

TruLens SDK integration is Snowflake's recommended AI observability approach for production RAG pipelines. TruLens instruments the pipeline at each stage, evaluating answer relevance (does the answer address the question?), context relevance (do the retrieved chunks relate to the question?), and groundedness (is the answer supported by the retrieved context?). Results are stored directly in Snowflake tables, enabling historical trending, threshold alerting, and audit trails. Cortex Guard is a content safety filter designed to detect harmful or policy-violating content — it is not a quality evaluation tool and cannot assess factual accuracy or hallucination rates. Query execution time has no correlation with answer quality; fast responses are routinely hallucinated. BLEU scores are a machine translation quality metric developed for comparing fixed-length candidate translations against reference translations, making them inappropriate for evaluating open-ended generative RAG responses.

Explanation

Research on RAG pipeline design consistently shows that smaller chunks of approximately 512 tokens or fewer yield higher retrieval quality compared to larger chunks, even when the embedding model technically supports much longer inputs. When chunks are smaller, each embedding more precisely represents a focused concept or passage, enabling vector similarity search to identify the most relevant passage for a given query with high accuracy. Large chunks blend multiple topics into a single embedding vector, making retrieval less targeted because the embedding must represent a broader range of content simultaneously. The 8,192-token context window of the snowflake-arctic-embed-l-v2.0-8k model indicates the maximum input it can process without truncation, not the recommended chunk size for optimal retrieval performance. The 2,048-token option falls between extremes but does not represent the research-backed recommendation for RAG retrieval quality. Chunk size is independent of a model's output embedding dimension — the 1,024 dimension count describes the size of the generated vector, not the recommended input size, and aligning chunk tokens to dimension count reflects a fundamental misunderstanding of how embedding models work.

Explanation

Snowflake Copilot is an in-product AI assistant integrated directly within the Snowsight interface, designed to help users write SQL, understand queries, and explore data through natural language without any additional configuration or setup. Its embedded nature and zero-configuration design make it the appropriate tool for ad-hoc SQL assistance for general analysts. Cortex Analyst is a developer-facing API feature that requires a semantic model defined in a YAML specification file, which captures business context, logical tables, dimensions, metrics, facts, and join relationships specific to the customer's data schema. This schema-aware configuration allows Cortex Analyst to generate accurate SQL against proprietary data models, making it the correct choice for a production analytics API where precision against known business metrics is critical. Copilot does not support custom semantic model YAML files — it works from general Snowflake SQL knowledge only. Cortex Analyst is not embedded in Snowsight and is not intended for ad-hoc user assistance. The two tools are architecturally distinct and serve complementary rather than overlapping roles.