Colleague pass checklist — Kaulab Basu (phData internal)

Highest-signal resource in this folder: a phData colleague's real study checklist + exam-experience notes, posted internally right after passing. Labeled GES-C01 (predecessor cert) but per the folder's standing note, C01 content is ~90% reusable for C02 — domain rebalance and document-processing reframing are the main deltas (see README.md).

New vs. our existing study files (verified 2026-09-16 by grepping study-cortex-ai.md and study-governance.md): Cortex Analyst depth (VQR mechanics, multi-turn REST API fields, semantic-model relationship syntax, observability event tables), the exact Document AI privilege grant (CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE at schema level), and the SPCS GPU_ compute-pool prefix requirement. Our RBAC/allowlist coverage (CORTEX_USER, CORTEX_MODELS_ALLOWLIST) already matches his.

Tier 1 — Highest weight (master these first)

1. Cortex Analyst

Natural-language-to-SQL over structured data; when to choose Analyst vs. Search vs. CLASSIFY_TEXT vs. PARSE_DOCUMENT.
Semantic model YAML: logical tables/columns, relationships (join_type, relationship_type, left_table/right_table, relationship_columns), and the join a model generates.
Verified Query Repository (VQR): use_as_onboarding_question, custom_instructions, and how suggested questions are generated (LLM fallback → up to 3; VQR present → up to 5 by semantic similarity).
Multi-turn conversations: passing full conversation history in the messages field; required fields in the Cortex Analyst REST API request body (content, role of speaker, path to semantic model YAML).
Observability with event tables (logs questions asked + generated SQL) and latency-optimization strategies.
Keeping the semantic model concise for high-cardinality fields (Cortex Search Service reference vs. Search Optimization Service).

2. Cortex Search (RAG)

The Snowflake feature for RAG over unstructured data; how it handles embeddings and vector similarity internally.
CREATE CORTEX SEARCH SERVICE: serverless (no dedicated warehouse), CHANGE_TRACKING = TRUE requirement for incremental refresh, ATTRIBUTES for filtering, embedding-model cost trade-offs (Arctic vs. Voyage).
Query interfaces: SEARCH_PREVIEW SQL function, Snowflake Python APIs, Snowflake REST APIs.
Integrating Cortex Search with AI_COMPLETE for a RAG chatbot: passing history, formatting retrieved context, enforcing structured output with response_format.
Cost drivers (query volume and indexed data volume) and cost/performance tuning for hourly-refreshed services.

3. Security, Roles & Privileges

Granting access to Cortex: SNOWFLAKE.CORTEX_USER database role; restricting access via REVOKE DATABASE ROLE ... FROM ROLE PUBLIC and REVOKE IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE.
Document AI privileges: CREATE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE granted at the schema level; role grants for using an existing pipeline.
GRANT statements to allow a role to create a Streamlit app (CREATE STREAMLIT ON SCHEMA, USAGE ON STAGE).
Two-part access control: RBAC role AND account-level model allowlist; SPCS public-endpoint auth (key-pair → JWT → OAuth token).

4. Cortex COMPLETE / AI_COMPLETE

What COMPLETE does (generates a response from a specified model + prompt); it is stateless between calls.
Required arguments: model and prompt_or_history; optional: top_p, max_tokens, response_format.
Parameters controlling randomness/diversity: temperature and top_p.
Structured JSON output via response_format with a JSON schema; extracting fields with dot notation into new columns.
Streaming with stream=True; the options/prompt_or_history array-of-objects requirement when using guardrails.
Multimodal COMPLETE use cases: labeling image files and classifying landmarks (not audio/in-painting/object replacement).

5. Fine-tuning

SNOWFLAKE.CORTEX.FINETUNE: training query must return PROMPT and COMPLETION columns.
Fine-tuning to a smaller, task-specific model to cut production cost without hurting performance.
Managing/monitoring fine-tuned models via FINETUNE functions and account usage views (not a fictitious ML_TEST view).

6. Document AI / PARSE_DOCUMENT

PARSE_DOCUMENT / AI_PARSE_DOCUMENT for extracting text/tables from PDFs and scanned images (built-in OCR).
Model !PREDICT with BUILD_SCOPED_FILE_URL from an internal stage; output is semi-structured JSON.
Supported actions/limits of Document AI (concurrent users on a model build); the standard directory-table → Document AI → streams/tasks pipeline pattern.

Tier 2 — Medium weight (solidify after Tier 1)

7. Model Access & Governance

CORTEX_MODELS_ALLOWLIST: comma-separated string syntax ('mistral-large2, llama3.1-70b'), not an array; 'ALL' vs 'None' behavior.
Cross-region inference: grant CORTEX_USER + enable AWS region + allowlist the model; data stays on the cloud provider's private backbone and is not cached.
403 error when calling a model not on the allowlist even with the right model role.

8. Task-Specific LLM Functions

Which functions are task-specific: SUMMARIZE, TRANSLATE, SENTIMENT, CLASSIFY_TEXT, EXTRACT_ANSWER, ENTITY_SENTIMENT (vs. general-purpose COMPLETE/TRY_COMPLETE).
Sentiment output shapes: SENTIMENT returns a float −1 to +1; ENTITY_SENTIMENT returns categorized labels.
Choosing COMPLETE for open-ended generation vs. EXTRACT_ANSWER/CLASSIFY_TEXT for extraction/labeling.

9. Embeddings & Vectors

EMBED_TEXT_768 / EMBED_TEXT_1024; billed on input tokens; 512-token context and chunking for RAG quality.
VECTOR data type (native, not VARIANT/OBJECT) and distance functions: VECTOR_L1_DISTANCE (Manhattan), VECTOR_L2_DISTANCE (Euclidean), VECTOR_COSINE_SIMILARITY, VECTOR_INNER_PRODUCT.
Computing document similarity with EMBED_TEXT_768 + VECTOR_L2_DISTANCE.

10. Model Registry

Required log_model arguments: model and model_name (version_name/code_paths/python_version optional).
Registry supports versioning and direct SQL invocation of registered models; it is language-agnostic and models stay internal.
Commands to list available Cortex LLMs: SHOW MODELS; and SHOW MODELS IN SNOWFLAKE.MODELS;.

Tier 3 — Lower weight (don't skip, but study last)

11. Snowpark Container Services (SPCS)

Deploying custom models: a compute pool (GPU_ prefix, e.g. GPU_NV_S) plus a service specification; requesting GPU via gpu_requests in create_service.
Choosing SPCS + GPU for low-latency custom/fine-tuned model inference.

12. Cost, Credits & Usage Views

CORTEX_FUNCTIONS_QUERY_USAGE_HISTORY for hourly token-cost aggregation; CORTEX_DOCUMENT_PROCESSING_USAGE_HISTORY for Document AI.
Design factors driving performance per credit (query complexity, compute availability).

13. Text Chunking

SPLIT_TEXT_RECURSIVE_CHARACTER returns an array of text chunks to fit LLM context windows; ordering it correctly in a pipeline for oversized documents.

14. Cortex Agents

Agentic paradigm with a single API for chatbots over mixed CSV/PDF data (Document AI + Cortex Search + Cortex Agents); Snowflake-hosted LLMs keep data in the governance boundary; cost = orchestration tokens + tool usage.

15. AI Safety & Guardrails

Aligned models refuse harmful requests (e.g., "hacking a system") rather than failing or hallucinating; Cortex Guard is optional/configurable, not mandatory.

16. AI Observability / TruLens

Enabling logging/traces for a Python RAG app: TRULENS_OTEL_TRACING=1 and installing trulens-core, trulens-connectors-snowflake, etc.

17. Streamlit & App Integration

Building chatbots with Streamlit + snowflake.cortex.complete; the snowflake.cortex.complete Python function is the Gen AI capability exposed via a REST API.

Suggested study sequence (Basu's own recommendation)

Work top-to-bottom: Tier 1 covers the bulk of the exam — budget most time on Cortex Analyst, Cortex Search, security/privileges, COMPLETE, fine-tuning, and Document AI.
For every function, memorize required vs. optional arguments and exact SQL syntax — many questions hinge on syntax details (allowlist string format, scoped file URL, PROMPT/COMPLETION columns).
Practice feature-selection questions (Analyst vs. Search vs. Agents vs. Document AI) — these recur constantly.
Learn the governance model cold: CORTEX_USER role, allowlist, cross-region behavior, schema-level Document AI privileges.

README.md — official C02 domain weights (Functions 38%, Governance 29%, Overview 18%, Document Processing 15%)
study-cortex-ai.md — our Domain 2 deep dive; Cortex Analyst section here is thinner than Basu's, worth merging his VQR/REST-API detail in
study-governance.md — our Domain 3 deep dive; RBAC/allowlist coverage already matches Basu's
study-plan.md — Day 2-3 (Cortex functions) and Day 6 (Governance) are where this slots in
third-party-resources.md — the broader curated-links list this supplements

colleague pass checklist kaulab basu