Snowflake Cortex AI & Agents¶
Last reviewed: October 2026
What's changed recently
- Snowflake Intelligence is now Snowflake CoWork (renamed June 2026), the end-user surface for Cortex Agents. Cortex Code is the AI coding assistant and is replacing Snowflake Copilot.
- "Cortex AI Functions" is the official name for the SQL AI surface
(formerly marketed as Cortex AISQL). It now includes
AI_EXTRACT,AI_PARSE_DOCUMENT,AI_REDACT,AI_TRANSCRIBE,AI_SUMMARIZE, andAI_MULTI_EMBED. Document AI was decommissioned in March 2026 in favor ofAI_EXTRACT. - MCP works in both directions. A Snowflake-managed MCP server (GA November 2025) exposes Cortex tools to external clients, and agents can call remote MCP servers through MCP connectors (GA mid-2026).
- Agent operations matured: Cortex Agent evaluations reached GA in March 2026, and background agent runs (up to 6 hours) work with threads.
A deep, production-focused interview guide for Snowflake Cortex AI — the LLM/ML layer that runs next to your data, inside Snowflake's governance perimeter. This section goes beyond the overview on the Snowflake technology page: here we build mental models for Cortex Agents, Cortex Analyst (text-to-SQL over semantic views), Cortex Search (native RAG retrieval), the AISQL functions, and the security / cost / observability story an interviewer will probe.
Currency note
Cortex features ship fast and names change. The patterns here (agentic orchestration, semantic models, hybrid retrieval, governed-in-perimeter AI) are stable, but treat specific capability claims as "as of current Snowflake docs" and verify against Snowflake's official Cortex documentation (docs.snowflake.com) before an interview. This guide was checked against Snowflake's official Cortex Agents documentation. Content was rephrased for compliance with licensing restrictions.
The one-sentence pitch
Cortex lets you run LLM inference, semantic search, and full agentic workflows on governed data without moving it out of Snowflake — the same RBAC, masking policies, tags, and audit trail that protect your tables also protect every AI call.
The Cortex landscape¶
Cortex is a stack, not a single product. Interviewers want to see that you know which layer solves which problem.
flowchart TB
subgraph PERIMETER[Snowflake governance perimeter - RBAC, masking, tags, audit]
subgraph DATA[Your data]
STRUCT[(Structured tables)]
UNSTRUCT[(Docs, transcripts, PDFs)]
end
subgraph FUNC[Cortex AI Functions]
LLM[AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_AGG, AI_EXTRACT, AI_SENTIMENT ...]
EMB[AI_EMBED / vector ops]
end
subgraph RETR[Retrieval + NL query]
SEARCH[Cortex Search - hybrid retrieval over unstructured]
ANALYST[Cortex Analyst - text-to-SQL over semantic views]
end
subgraph AGENT[Cortex Agents - managed agentic platform]
ORCH[Orchestrator LLM: plan / use tools / reflect]
end
end
STRUCT --> ANALYST
UNSTRUCT --> SEARCH
UNSTRUCT --> FUNC
ANALYST --> AGENT
SEARCH --> AGENT
FUNC --> AGENT
AGENT --> APPS[Snowflake CoWork / Cortex Code / your app via REST API]
| Layer | What it is | When you reach for it |
|---|---|---|
| Cortex AI Functions (AISQL) | Call LLMs and embeddings directly in SQL — newer AI_* family (AI_COMPLETE, AI_CLASSIFY, AI_FILTER, AI_AGG, AI_EMBED, AI_EXTRACT, …) and legacy SNOWFLAKE.CORTEX.* (COMPLETE, SUMMARIZE, SENTIMENT, EMBED_TEXT_*). See the functions & API reference. |
Batch enrichment, classification, extraction over columns — set-based AI |
| Cortex Search | Managed hybrid (vector + keyword) retrieval service over text | The RAG retrieval engine; grounding for chatbots and agents |
| Cortex Analyst | Natural language → governed SQL over a semantic view/model | Self-serve analytics, "ask your data" over structured tables |
| Cortex Agents | Fully managed agentic platform: an orchestrator LLM that plans, calls tools, runs code, and responds | Multi-step questions that mix structured + unstructured data and actions |
| Snowflake CoWork / Cortex Code | CoWork (formerly Snowflake Intelligence) is the business-user agent chat; Cortex Code is the developer coding assistant | The UX layer; you also call agents from your own app via REST API or MCP |
The three questions that pick the layer¶
- Is the answer in structured tables? → Cortex Analyst (needs a semantic view).
- Is the answer in documents/text? → Cortex Search (native RAG).
- Does it need both, plus reasoning, tool calls, or actions? → a Cortex Agent that uses Analyst and Search as tools.
Why "AI next to the data" is the whole story¶
The differentiator versus bolting an external LLM app onto Snowflake:
- No data egress. Inference happens inside the account; sensitive columns never leave the perimeter to hit a third-party endpoint.
- Governance is inherited, not re-implemented. RBAC roles, dynamic data masking, row-access policies, and object tags apply to the data an AI call touches. You don't rebuild access control in an app tier.
- Auditability. Access History and query history cover AI calls and agent tool use, so you can answer "who asked what, over which data."
- Least data movement = less attack surface + lower cost (no duplicate vector store to secure, sync, and pay for separately).
Interview framing
When asked "why Cortex over a LangChain app on top of Snowflake?", lead with governance and data gravity, not model quality: "The models are comparable; the win is that the security and compliance boundary already exists around the data, so I inherit it instead of rebuilding it in an app tier — and I avoid egress." Then acknowledge the trade-off (below).
Honest trade-offs (say these before the interviewer does)¶
A senior candidate names the limits, not just the benefits:
- Model / provider choice is curated, not arbitrary. You pick from the models Snowflake offers (or let it auto-select), not any model on the internet.
- Region / availability — specific functions and models roll out by region; check availability for your account before promising a capability.
- Cost is credit-based and can surprise you — large-model calls over big columns are expensive; you design for it (batching, cheaper models, filters).
- You're inside Snowflake's runtime — great for governance, but heavy custom orchestration or non-SQL glue may still want an app tier calling the REST API.
What to read next in this section¶
-
The plan → use tools → reflect loop, tool types (Analyst, Search, code sandbox, custom tools, MCP), Threads, and how to design a governed agent.
-
Analyst, Semantic Models & Search (native RAG)
Semantic views, text-to-SQL, hybrid retrieval, Search↔Analyst integration, and combining structured + unstructured data.
-
AISQL Functions & the Agents API (technical reference)
The concrete layer: the
AI_*vsSNOWFLAKE.CORTEX.*function surface,AI_COMPLETEoptions + structured outputs, embeddings/vectors/chunking, and theagent:runREST API vsAGENT_RUN. -
Security, Governance, Cost & Observability
RBAC in the agent path, inaccessible-tool handling, audit, credit control, monitoring/evals, and production incidents with mitigations.
-
System Design + Mock Interview
A Cortex agentic system-design walkthrough and a timed mock interview.
-
Cheat Sheet + 30-Day Cortex Ramp
Rapid-fire facts, SQL snippets, and a focused study plan.
Related pages in this site¶
- Snowflake (technology overview) — architecture, SQL, performance, governance basics.
- Snowflake Interview Q&A — core warehousing/SQL interview bank.
- RAG topic and Agent Engineering — the vendor-neutral concepts Cortex implements.
- AI Security — the threat model that the governance story here defends against.