--- title: "What Is Snowflake Cortex? | Dawiso" description: "Snowflake Cortex is Snowflake's suite of managed AI features: LLM functions, Cortex Analyst, Search and Agents, running on governed data in the platform." url: "https://www.dawiso.com/glossary/snowflake-cortex" language: "en" --- June 8, 2026 snowflake cortexcortex AIcortex analystcortex searchcortex agentsLLM functionssnowflake AI # What Is Snowflake Cortex? **Snowflake Cortex** is Snowflake's suite of fully managed AI and machine-learning features that run directly inside the [Snowflake](https://www.dawiso.com/glossary/snowflake) platform - on the data you already store there, under the access controls you already manage. Instead of exporting data to an external model provider or standing up a separate ML stack, you call Cortex with SQL or Python and the work happens where the data lives. Cortex spans serverless [large language model](https://www.dawiso.com/glossary/large-language-model) functions, a text-to-SQL analytics service, an enterprise search service, and an agent layer that orchestrates them. Cortex matters because it collapses the distance between data and AI. The persistent problem with enterprise AI is not the model - it is getting governed, trustworthy data to the model safely. By embedding AI functions inside the data platform, Cortex inherits Snowflake's role-based access control, so an AI feature cannot quietly read data a user was never allowed to see. That security inheritance is real and valuable - but it is access control, not business meaning. Cortex governs *which rows* a model may read; it does not, on its own, decide what those rows *mean*, trace where they came from across the wider data stack, or settle which definition of a metric is the agreed one. Those remain the work of governance - and they are what determine whether a Cortex answer is actually right. TL;DR **Snowflake Cortex** is Snowflake's managed AI suite that runs on governed data inside the platform. It includes **Cortex AISQL / LLM Functions** (call models like Claude, Llama, and Mistral with SQL), **Cortex Analyst** (ask questions in plain English and get governed SQL answers from [semantic views](https://www.dawiso.com/glossary/snowflake-semantic-views)), **Cortex Search** (hybrid vector + keyword [RAG](https://www.dawiso.com/glossary/what-is-rag-in-ai) over text), and **Cortex Agents** (orchestrate the above to answer complex questions). Its enterprise advantage is that every call inherits Snowflake's access control - but the quality of the answers still depends on the business context behind the data. ## Snowflake Cortex Defined Snowflake Cortex is a collection of serverless AI capabilities exposed as native Snowflake functions and services. "Serverless" means there is no infrastructure to provision: you do not size GPUs or manage model endpoints. You invoke a function, Snowflake runs it on managed compute, and you pay per use. "Native" means the features are part of the platform's surface area - callable from a SQL statement, a Snowpark Python program, or the Snowflake UI - and subject to the same governance as the underlying tables. The defining characteristics of Cortex: - **Data stays in place** - AI runs against data inside Snowflake's security boundary, not in an external service. - **Access control is inherited** - Cortex calls respect existing role-based access control, masking policies, and row-level security. This is security governance; the semantic governance that decides whether an answer is *correct* still has to be supplied. - **SQL-first access** - Most capabilities are reachable from a SQL query, which means analysts, not just ML engineers, can use them. - **Managed models** - Snowflake hosts and updates a catalog of foundation models, removing the operational burden of running them. ## The Components of Cortex Cortex is best understood as four layers that build on each other, with a set of foundation models underneath and an orchestration layer on top. ## Cortex LLM Functions (AISQL) The foundation of Cortex is a set of LLM functions - increasingly branded "AISQL" - that let you call a large language model from inside a SQL query. Instead of writing application code to hit a model API, you run a function over a column and get results back as rows. Common functions include: - **COMPLETE** - Send a prompt to a chosen model and get a generated completion. The general-purpose function behind most custom use cases. - **SUMMARIZE** - Condense long text into a short summary. - **TRANSLATE** - Translate text between languages. - **SENTIMENT** - Score the sentiment of text. - **CLASSIFY_TEXT** - Assign text to categories you define. - **EXTRACT_ANSWER** - Pull a specific answer out of a passage. - **EMBED_TEXT** - Generate [vector embeddings](https://www.dawiso.com/glossary/vector-database) for similarity search and RAG. Because these run as SQL, they compose naturally with the rest of your queries: classify ten thousand support tickets, summarize each long description, and join the results to your existing tables in one statement. Snowflake hosts a catalog of [foundation models](https://www.dawiso.com/glossary/large-language-model) - including Anthropic's Claude, Meta's Llama, Mistral, OpenAI models, and Snowflake's own Arctic - and you select the model per call. ## Cortex Analyst, Search & Agents On top of the raw functions, Cortex provides three higher-level capabilities that solve specific enterprise problems. ### Cortex Analyst Cortex Analyst is a managed text-to-SQL service for structured data. A user asks a question in plain English - "what was net revenue by region last quarter?" - and Analyst generates and runs the SQL, returning a governed answer. Critically, it does not read raw tables directly; it reads a [semantic view](https://www.dawiso.com/glossary/snowflake-semantic-views), which carries the metric definitions, join paths, and synonyms that tell the model what the business actually means. The semantic view is what separates a reliable answer from a plausible-looking wrong one. ### Cortex Search Cortex Search is a hybrid retrieval service combining vector (semantic) and keyword (lexical) search over text. It is the retrieval engine for [retrieval-augmented generation](https://www.dawiso.com/glossary/what-is-rag-in-ai) inside Snowflake - index your documents, and an application or agent can fetch the most relevant passages to ground an LLM's answer, reducing [hallucination](https://www.dawiso.com/glossary/ai-hallucination). ### Cortex Agents Cortex Agents orchestrate the layers below. An agent answers a complex business question by planning which tools to call - Cortex Analyst for the structured numbers, Cortex Search for the supporting documents - and then composing a final answer. This is the basis of [agentic](https://www.dawiso.com/glossary/multi-agent-systems) "talk to your data" experiences, where each department might run its own agent scoped to its own semantic views. ## Why Governance Matters for Cortex Cortex gives an AI agent two things: a capable model and Snowflake's access control. What it does **not** give the agent is context. A Cortex agent has no inherent knowledge of what your data means - it does not know that "net revenue" excludes returns, that a "customer" is deduplicated across three source systems, or which of four similarly-named tables is the certified one. An agent is only ever as good as the context it is handed, and Cortex does not author that context for you. So where is an agent supposed to get its context? Cortex has designated places for it - the [semantic view](https://www.dawiso.com/glossary/snowflake-semantic-views) for structured questions, and indexed, classified content for Cortex Search. But those are *containers*, not sources. A semantic view is only correct if something fills it with the right metric definitions and join paths; Cortex Search only retrieves well if the content underneath is organized and classified. Point an agent at an empty or wrong semantic view and it will answer just as confidently - and just as wrongly. This is the whole reason a governed [data catalog](https://www.dawiso.com/glossary/data-catalog) belongs underneath an AI agent. The catalog, [business glossary](https://www.dawiso.com/glossary/what-is-a-business-glossary), and [data lineage](https://www.dawiso.com/glossary/data-lineage) are where an agent's context actually lives - and where it is kept correct: - The **business glossary** holds the agreed definitions - what "net revenue" or "active customer" actually means - that a semantic view must encode for Cortex Analyst to answer correctly. Without it, every team hand-rewrites definitions and the agents quietly disagree with each other. - **Data lineage** answers the question every compliance team eventually asks - which data can this agent actually see, and what breaks if an upstream table changes - across platforms, not just inside Snowflake. - The **data catalog** is what makes the tables and documents behind the agent discoverable, classified, and trustworthy in the first place - so the agent reasons over curated context, not whatever happened to be lying around. This is exactly the role [data governance](https://www.dawiso.com/glossary/data-governance) - and Dawiso - plays for Cortex. Dawiso connects to Snowflake and scans Cortex agents and their semantic views into one interactive lineage, so you can trace exactly which agent reads which view and which tables those views touch. And because a governed Dawiso [data product](https://www.dawiso.com/glossary/data-product) already carries the definitions, glossary terms, and mappings a semantic view needs, Dawiso can generate that semantic view back into Snowflake through the Open Semantic Interchange standard - so the context an agent depends on comes from one governed source of truth instead of being rebuilt, inconsistently, by hand. Put simply: **Cortex runs the agent; a governed catalog is where the agent gets its context.** We cover the full workflow in [Govern Snowflake Cortex Agents and Semantic Views](https://www.dawiso.com/blog-post/snowflake-cortex-agents-semantic-views-dawiso), and the deeper mechanics in [semantic views and Cortex Analyst](https://www.dawiso.com/glossary/snowflake-semantic-views). ## Conclusion Snowflake Cortex brings AI to the data instead of moving data to the AI - and in doing so inherits the governance that makes enterprise AI safe. Its functions, Analyst, Search, and Agents turn the warehouse into an AI platform reachable from plain SQL. But the lesson that holds across every AI system holds here too: the model is rarely the limiting factor. The accuracy of a Cortex answer is decided upstream, in the quality of the semantic views, definitions, and governed context that tell the AI what your data means. [See it in action](https://www.dawiso.com/product/ai-powered-features) AI-Powered Features AI-assisted documentation, intelligent search, and automated categorization help you work smarter. ### Table of contents Snowflake Cortex DefinedThe Components of CortexCortex LLM Functions (AISQL)Analyst, Search & AgentsWhy Governance Matters ## Related Terms ### [What Is the Snowflake MCP Server? Tools, Setup and Limits](https://www.dawiso.com/glossary/snowflake-mcp-server) The Snowflake-managed MCP server gives AI agents Cortex Agents, Analyst, Search and SQL as MCP tools. Setup, OAuth, access control, limits and where it fits. Read more → ### [Context Layer for Snowflake & Horizon Catalog](https://www.dawiso.com/glossary/context-layer-for-snowflake) A context layer gives Snowflake data the governed business meaning AI needs. Learn how it works with Snowflake Horizon Catalog and Cortex, and what spans beyond Snowflake. Read more → ### [What Are Snowflake Semantic Views & Cortex Analyst?](https://www.dawiso.com/glossary/snowflake-semantic-views) Snowflake semantic views are native objects defining metrics, relationships, and synonyms so Cortex Analyst turns plain-English questions into accurate, governed SQL. Read more → ### [What Is a Large Language Model (LLM)?](https://www.dawiso.com/glossary/large-language-model) A large language model (LLM) is a neural network trained on massive text corpora to understand and generate language. Learn how LLMs work, their limits, and why data governance matters for enterprise AI. Read more → ### [What Is Prompt Engineering?](https://www.dawiso.com/glossary/prompt-engineering) Prompt engineering is the practice of crafting inputs to language models to get reliable, accurate, and useful outputs. Learn the key techniques, why context quality matters more than prompt tricks, and how governed data makes prompts work. Read more → ### [What Is Context Engineering?](https://www.dawiso.com/glossary/context-engineering) Context engineering is the practice of designing and managing the information delivered to an AI model's context window to maximize output quality. Learn why it matters more than prompt crafting and how governed data makes it tractable. Read more →