Home/Insights/Data Engineering
Watch the video with chapters & transcript →
AI IN FINANCE
In short: One YAML file, the semantic view, decides whether Snowflake Cortex Analyst works because it explains to the LLM what every column in your tables means. It translates cryptic column names into business meaning and synonyms, so the model picks the right column. Most of the work is editing the generated YAML, not building the agent.

Snowflake Cortex Analyst lets business users query SQL tables in plain language, and the hard part is not building the agent. The make-or-break component is the semantic view: a YAML file that explains to the LLM what every column in your tables actually means, demonstrated here on a compliance database with three tables: breaches, holdings, and rules.
Snowflake Intelligence sits on top as a chatbot, like ChatGPT but grounded in your Snowflake data, and underneath it the Cortex Agent plans, reasons, and orchestrates. The agent routes each question to Cortex Analyst for structured SQL tables or to Cortex Search for documents, and the semantic view sits between the chat layer and the tables, so the LLM checks it first to find the right columns.
A semantic view translates cryptic column names into business meaning, including synonyms the LLM can match against user phrasing. A column called weight_pct is described as "percentage of total portfolio" with synonyms like weight and allocation, and value_million_sek becomes "market value in million SEK" with synonyms value, worth, and size. When a user asks "How much Volvo do we own?", the LLM reads those descriptions to pick the right column instead of guessing between weight_pct, value_million_sek, and current_pct.
| Component | Role |
|---|---|
| Snowflake Intelligence | Chat interface grounded in your Snowflake data |
| Cortex Agent | Plans, reasons, and routes questions to the right tool |
| Cortex Analyst | Generates SQL against structured tables |
| Cortex Search | Retrieves answers from documents |
| Semantic view (YAML) | Explains column meanings and synonyms to the LLM |
In Snowsight, go to AI and ML, open AI Studio, choose Cortex Analyst, click Try, then Create New. Snowflake offers three options: create a new semantic view (the recommended path, saved directly under your schema), create a semantic model (a legacy YAML file you upload to a stage), or upload your own YAML. From there you pick the database and schema, select the tables the agent may use, select the columns, and click Create and Save; generation takes a couple of seconds.
Most of the work is editing the generated YAML: adding descriptions, synonyms, named filters, metrics such as sum, average, and count, and verified questions your team actually asks. Snowflake auto-classifies columns into dimensions (descriptive), time dimensions (date types), and facts (numeric), but that file is only a starting point. Ani Björkström, a Stockholm-based tech consultant, calls this editing stage the part that makes or breaks the agent.
No. A free Snowflake account includes $400 in credits valid for 1 month, which is enough to build and test a Cortex Analyst on your own tables.
The LLMs behind Cortex are only available in certain regions, so your Snowflake account must be hosted where they run. The demo account used West Europe.
Verified questions are expected user questions stored with a reviewed SQL query. You can run the SQL, confirm the result is correct, and save it so future identical questions get validated answers.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?