Colleague exam-experience notes (topics + scenario flavor, not an answer key)

Two more phData colleagues' feedback, forwarded same day as colleague-pass-checklist-kaulab-basu.md. These are question prompts only — no answer choices or correct answers are recorded anywhere here. Treat as signal for which mechanics/topics to know cold and what scenario framing looks like, not as a memorize-the-answer-key resource (see third-party-resources.md's standing note on why brain-dump-style prep is both fragile and against Snowflake's exam policy — this stays on the "what to study" side of that line since no answers are included).

Exam character (from the overview slide)

• Heavily scenario-based and syntax-focused
• Multiple questions covered: VQR, Cortex Guard, cross-region inference, model performance metrics
• Comparatively lengthy exam
• Process of elimination is a useful technique
• Complete syntax given in one question can help answer another question about the same service (functions/services get asked about from multiple angles)

Topics confirmed tested (from the raw question list — paraphrased prompts, no answers)

• Document AI: max PDF count it can process at a time
• CLASSIFY_TEXT: predicting query output given a specific call
• AI_COMPLETE with guardrail=true: predicting query output
• Improving CLASSIFY_TEXT accuracy when it misclassifies (multi-select)
• Which functions charge for both input AND output tokens (vs. input-only)
• Data type of the input parameter in VECTOR_COSINE_SIMILARITY
• Compulsory (required) parameters in CORTEX_COMPLETE — same question pattern likely recurs for other functions
• Ways CORTEX_SEARCH_SERVICES can be called (multiple interfaces)
• Scenario: table has group-specific allergy data; which Cortex function tells a GenAI developer what foods a group should avoid (function-selection judgment call, not syntax recall)

Additional specifics (from the rewritten/expanded version — some genuinely new vs. the raw list above)

1. Where an uploaded file is stored by default when using the Snowflake Intelligence chat interface
2. Cortex Analyst semantic model: which component of module_custom_instructions restricts/blocks specific topics before SQL generation
3. Choosing the right capability for NL-over-structured-data via REST API (feature-selection, same genre as Basu's #1 topic)
4. AI_PARSE_DOCUMENT: best way to extract specific fields (customer name, invoice number) from parsed invoice content
5. AI_PARSE_DOCUMENT error modes: a 110MB PDF fails with a file-size error; a PNG screenshot fails when processed with page_split=TRUE — both explained + how to address
6. Cortex Search: which cost component to examine when a service keeps generating cost with zero active queries (idle/indexed-data cost, distinct from query-volume cost — ties to Basu's cost-driver notes)
7. FINETUNE on an arctic-extract model: how the number of training epochs gets determined
8. Model Registry: interface/mechanism to invoke a registered custom model from an external application (distinct from the internal-SQL-invocation path Basu's notes cover)

What this adds to the study hub

New surface area not in colleague-pass-checklist-kaulab-basu.md or the existing study-*.md files: module_custom_instructions topic-restriction component, AI_PARSE_DOCUMENT's two specific failure modes (110MB size cap, page_split=TRUE + image input conflict), and external-application invocation of a Model Registry model. Worth adding hands-on reps for these three during Day 3-4 of study-plan.md.

Related

• colleague-pass-checklist-kaulab-basu.md — the fuller tiered checklist, same forwarding session
• study-plan.md §Spend-extra-time topics — same genre of community signal, now corroborated
• third-party-resources.md — brain-dump policy note this file stays on the right side of