Snowflake Launches Cortex to Centralize Generative AI on Its Data

Snowflake brings together generative AI components in Cortex, accessible where the data already resides. From SQL queries to document searches, and orchestrated agents, the entire suite aims to limit the dispersion of components and address operational needs such as customer support.

Customer Support: Classifying, Summarizing, and Accelerating Responses

Handling support requests exposes companies to large volumes of messages, complaints, and inquiries, the manual management of which mobilizes teams and complicates the rapid detection of recurring issues. Cortex's AI Functions are designed to classify these requests, identify their sentiment, and produce summaries. The search service can complement this setup to enhance processing efficiency. The overall functionality aims to reduce the time spent on repetitive tasks and facilitate the utilization of existing data.

Architecture: Reducing Fragmentation of AI Projects

Organizations were exporting their data to external services to use language models, then building RAG pipelines or automations, resulting in a proliferation of components and governance challenges. A traditional project stacks data warehouses, inference services, vector databases, orchestrators, and application layers, each with its own security parameters, availability, and billing. Snowflake proposes to integrate AI functions within the same environment as the data to limit dispersion. In Cortex, developers can utilize models provided by Anthropic, Mistral, Meta, or OpenAI depending on regions and models, bringing AI capabilities closer to the information to be leveraged.

Direct Access from SQL and Python for Data Teams

Cortex makes AI operations available in the tools already used by teams, with model calls from SQL and Python. Cortex AI Functions execute generation and analysis tasks directly within queries, without an intermediary application to process corpuses. AI_COMPLETE generates a text response from an instruction, AI_SUMMARIZE_AGG produces multi-line summaries, and AI_FILTER returns True or False as a safeguard in a WHERE clause. Other functions cover translation, sentiment analysis, and classification. By 2026, the suite will consolidate these components within the same scope, with managed services to work on existing data.

Cortex Analyst Translates Natural Language Questions into SQL Queries

For teams that do not write queries, Cortex Analyst translates questions posed in natural language into SQL tailored to the schema. Users can request quarterly revenue, compare subsidiaries, or search for the most profitable products. This process requires an understanding of the relationships between tables and relevant measures; a semantic model guides the service on the structure. Dimensions, measures, and relationships can be described in YAML or managed through Semantic Views, which frame the meaning of fields and calculation rules. In the absence of a semantic layer, the model must infer the revenue column, the region field, the customer-order link, and the period to consider.

Cortex Search Facilitates the Utilization of PDF Files and Technical Reports

Beyond tables, many informational assets reside in PDF contracts, technical reports, internal documentation, meeting notes, or Markdown files. Cortex Search allows these to be utilized without fully converting them into relational tables, by combining lexical and semantic search to locate relevant passages. Lexical search relies on term matching, useful for references or identifiers, while semantic search is based on vector embeddings that encode the meaning of texts. The setup begins with selecting a textual source; documents can be stored in a Snowflake stage, processed by AI_PARSE_DOCUMENT, and then prepared for indexing. Depending on the structure, chunking is used, a key practice in RAG architectures to avoid sending complete documents to models.

Agents: Orchestrating Analyst and Search in a Single Response

AI agents, offered through Cortex Agents, break down a goal, mobilize multiple tools, and combine components from the ecosystem, including Analyst and Search. They handle requests requiring both structured data and documents, chaining SQL queries, document searches, and generating a contextualized response. Each agent relies on instructions that define its role, objectives, and limits, knows the authorized tools, and triggers operations based on the request. The integration avoids the need to separately develop a document engine and a SQL generator while allowing for the possibility of connecting an external orchestrator. The stated goal of the integrated AI layer is to leverage LLMs without building a separate infrastructure from the warehouse.