Advanced Cortex Functions

By now, most Snowflake developers have run a COMPLETE() function. But Cortex is evolving into a comprehensive AI suite. Let’s look at the advanced capabilities that separate a “demo” from a production AI app.

1. Hybrid Search (Cortex Search)#

Vector search (semantic similarity) is great, but sometimes you need exact keyword matching (Lexical search). Cortex Search now supports Hybrid Retrieval.

• Use Case: Searching for a specific product SKU (Lexical) while asking “which product is best for hiking” (Semantic).
• Implementation: The SEARCH() capability now accepts a hybrid_fusion parameter to weight the results from both algorithms.

2. Cortex Analyst (Text-to-SQL)#

Unlike generic LLMs which struggle with schema accuracy, Cortex Analyst relies on a Semantic Model (YAML file defining metrics and dimensions).

• Best Practice: Don’t just point it at raw tables. Build a simplified view layer and define your synonyms in the semantic model. “Revenue” might mean TOTAL_SALES - DISCOUNTS—the model captures this business logic so the LLM doesn’t have to guess.

3. Structured Output#

Getting an LLM to output valid JSON is notoriously hard. Cortex now supports “constrained generation”. You can pass a JSON schema to the completion function, ensuring the output matches your downstream application’s requirements strictly.

SELECT snowflake.cortex.complete(
    'llama-4-70b',
    'Extract the date and amount from this invoice...',
    options => {'response_format': {'type': 'json_object'}}
);

Conclusion#

Cortex is moving up the stack. It’s not just a wrapper around an LLM; it’s a platform for reliable, governed AI applications.