Snowflake Cortex AI Rules

tl;dr: pull snowflake-cortex-ai-cursorrules-prompt-file.mdc from awesome-cursorrules before writing Cortex queries in a Cursor project. Models reach for deprecated names like COMPLETE and CLASSIFY_TEXT, which 404 in current Snowflake.

What it does

This rule set configures Cursor to use Snowflake Cortex AI the way it ships today: SQL-callable AI Functions with the AI_* prefix, Cortex Search for managed hybrid (vector + keyword) retrieval, and the standard RAG pattern (search → context → AI_COMPLETE). Without it, models default to deprecated function names, hardcode model IDs that aren’t deployed in the user’s region, and row-by-row feed tables into prompts.

Sourced from PatrickJS/awesome-cursorrules — the community registry for .mdc Project Rules.

The rules

// Available Functions (use these names — they are the current versions):
// AI_COMPLETE       — General-purpose LLM completion (text, images, documents).
// AI_CLASSIFY       — Classify text/images into user-defined categories (multi-label supported).
// AI_FILTER         — Returns TRUE/FALSE for text/image input. Use in WHERE clauses.
// AI_AGG            — Aggregate insights across rows of text (no context window limit).
// AI_EMBED          — Generate embedding vectors (similarity search, clustering).
// AI_EXTRACT        — Extract structured info from text, images, or documents.
// AI_SENTIMENT      — Sentiment score from text (-1 to 1).
// AI_SUMMARIZE_AGG  — Summarize across rows (no context window limit).
// AI_SIMILARITY     — Embedding similarity between two inputs.
// AI_TRANSCRIBE     — Transcribe audio/video from stages.
// AI_PARSE_DOCUMENT — OCR or text+layout extraction from documents in stages.
// AI_REDACT         — Redact PII from text.
// AI_TRANSLATE      — Translate between supported languages.

// Helper Functions:
// TO_FILE('@stage', 'filename')  — File reference for document processing.
// AI_COUNT_TOKENS(model, text)   — Check token count before calling a model.
// PROMPT('template {0}', arg)    — Build prompt objects for AI_COMPLETE.
// TRY_COMPLETE                   — Returns NULL on failure instead of error.

// Privileges: USE AI FUNCTIONS account privilege + SNOWFLAKE.CORTEX_USER database role.

// Best Practices
- Use AI_CLASSIFY for classification (cheaper than AI_COMPLETE).
- Check token counts with AI_COUNT_TOKENS before large batch jobs.
- Set PRIMARY KEY on Cortex Search for optimized incremental refresh.
- Use ATTRIBUTES for filterable columns. Use SEARCH_PREVIEW for testing, Python/REST for production.
- Use dedicated warehouse (no larger than MEDIUM) per search service.

// Anti-Patterns
- Do NOT use old function names (COMPLETE, CLASSIFY_TEXT, etc.) — use AI_* versions.
- Do NOT pass entire tables through AI_COMPLETE row-by-row without cost estimation.
- Do NOT hardcode model names without considering regional availability.

Installation

Drop the .mdc into .cursor/rules/ of your project:

mkdir -p .cursor/rules
curl -sL \
  https://raw.githubusercontent.com/PatrickJS/awesome-cursorrules/main/rules/snowflake-cortex-ai-cursorrules-prompt-file.mdc \
  -o .cursor/rules/snowflake-cortex-ai.mdc

Cursor picks up files in .cursor/rules/ automatically.

AI_COMPLETE — text and documents

-- Text completion
SELECT AI_COMPLETE(
  MODEL => 'claude-4-sonnet',
  PROMPT => 'Summarize: ' || review_text
) AS summary
FROM reviews;

-- Document processing (PDF in stage)
SELECT AI_COMPLETE(
  MODEL => 'claude-4-sonnet',
  PROMPT => PROMPT('Extract the invoice total from {0}', TO_FILE('@docs', 'invoice.pdf'))
) AS total
FROM @docs;

Cortex Search — single index

CREATE OR REPLACE CORTEX SEARCH SERVICE my_search
  ON transcript_text
  ATTRIBUTES region, agent_id
  WAREHOUSE = my_wh
  TARGET_LAG = '1 day'
  EMBEDDING_MODEL = 'snowflake-arctic-embed-l-v2.0'
  AS (SELECT transcript_text, region, agent_id FROM support_transcripts);

RAG pattern (Python)

from snowflake.core import Root

root = Root(session)
service = root.databases["db"].schemas["schema"].cortex_search_services["my_search"]

results = service.search(
    query="internet connection issues",
    columns=["transcript_text", "region"],
    filter={"@eq": {"region": "North America"}},
    limit=5,
)
context = "\n\n".join(r["transcript_text"] for r in results)

answer = session.sql(
    "SELECT AI_COMPLETE(MODEL => 'claude-4-sonnet', "
    "PROMPT => 'Answer from context: ' || :1 || ' Q: ' || :2)"
).bind((context, question)).collect()

Why it helps

Without Cortex-aware rules, models will:

• Call deprecated COMPLETE/CLASSIFY_TEXT — these no longer exist in current Snowflake
• Skip AI_COUNT_TOKENS and run a 50k-row batch that blows the warehouse budget
• Hardcode claude-4-opus in a region where only claude-sonnet-4-5 is deployed
• Run AI_COMPLETE row-by-row without AI_AGG/AI_SUMMARIZE_AGG for cross-row aggregations
• Skip PRIMARY KEY on Cortex Search, slowing incremental refresh

These rules keep Cortex queries on the supported function surface and budget-aware.

Gotchas

• These are Cursor .mdc rules, not a Claude Code CLAUDE.md file. For Claude Code, paraphrase into your project’s CLAUDE.md.
• The rule list quotes the current model roster from the fetched .mdc — verify model availability per region before shipping.
• Cortex Search Python API needs the snowflake-core-python package; REST hits /api/v2/databases/<db>/schemas/<schema>/cortex-search-services/<service>:query.

Related

• TanStack Query v5 Rules — client-side caching layer for Cortex RAG answers in a React UI
• React Router v7 Rules — route loaders can call the Cortex REST API to prefetch RAG context