Visual guides

Visual Guides on Snowflake, Data & More

Cortex AI, inside Snowflake LLMs, search and text-to-SQL that run where the data already lives Your data tables · docs · PDFs tickets · call notes Snowflake Cortex LLM functions COMPLETE · SUMMARIZE TRANSLATE · SENTIMENT CLASSIFY · EXTRACT Cortex Search hybrid vector + keyword the "R" in RAG auto-refreshes with data Cortex Analyst question → SQL guided by a semantic model you define Document AI pull fields out of invoices, forms, scans into columns Cortex Agents plan → Search or Analyst → run tools → answer with citations one API for chat apps, Slack bots, Snowflake Intelligence Call it from ▸ plain SQL ▸ Python / Snowpark ▸ REST API ▸ Streamlit apps ▸ Snowflake Intelligence Models hosted in-platform Anthropic · OpenAI · Meta · Mistral no keys, no egress, billed in credits Data never leaves the governance boundary RBAC, masking policies, row access and lineage apply to AI calls too sketch 01 · snowflake, drawn out
AI

What Cortex is and how I use it

The parts of Snowflake Cortex I actually reach for, and how a RAG app fits together without moving data out.
Adaptive Warehouses Set a spend ceiling. Snowflake picks the compute for each query. Before one warehouse per workload XS BI dashboards M ELT jobs 2XL "the month-end one" idle 27 days a month queued ✗ you pick the size up front ✗ resize by hand when it queues ✗ idle time = credits burned ✗ one big query jams the small one same SQL After one adaptive warehouse ADAPTIVE_WH q1 small q2 big, briefly q3 medium ✓ Snowflake sizes compute per query ✓ scales out when queries queue ✓ shared pool, so far less idle you set: max credits that's the whole config What stays the same connection strings keep the warehouse name, roles and grants carry over, queries are unchanged. start with dev and BI; leave tuned ELT on classic warehouses at first sketch 02 · snowflake, drawn out
Compute

Adaptive warehouses, and what changes when you switch

Snowflake can now size and scale compute per query. What changes, what stays the same, and how I'd move a customer over.
Where do Snowflake credits go? Six things I check, and the queries I use to find them compute ~ 3 of every 4 credits typical mix, yours will differ virtual warehouses (running, and idle) storage: tables + Time Travel + Fail-safe cloud services + serverless (tasks, search, Cortex) Six things I check, in order 1 Auto-suspend at 60s, not 10 minutes cache loss is cheaper than paying for silence · AUTO_SUSPEND = 60 2 Find the top 10 queries, fix those first QUERY_ATTRIBUTION_HISTORY · a few queries burn most of the credits 3 Right-size: start small, grow only on spill each size up doubles cost; bytes spilled to remote storage is your signal 4 Put guardrails on: monitors, budgets, timeouts RESOURCE MONITOR · BUDGET · STATEMENT_TIMEOUT on every warehouse 5 Trim storage you never read transient tables for staging · Time Travel 1 day, not 90 · drop the clones nobody owns 6 Consolidate warehouses, then go adaptive fewer, busier warehouses idle less · adaptive warehouses do the sizing for you Rule of thumb: 20% of warehouses and queries cause 80% of the spend. sketch 03 · snowflake, drawn out
Cost

Six things I check when a Snowflake bill goes up

Where credits go, the six things I look at in order, and the queries I use to find them.