Snowflake Cortex AI

Snowflake Cortex AI

This skill covers Snowflake Cortex — the SQL-callable AI layer of Snowflake, including Cortex AI Functions for LLM/ML tasks and Cortex Search for managed hybrid (vector + keyword) search — all running inside Snowflake with no data leaving the platform.

Workflow for Building a Cortex-Powered Feature

1. Pick the narrowest function for the task — Use AI_CLASSIFY for categorization, AI_FILTER for natural-language row filtering, AI_EXTRACT for structured field pulls, and reserve AI_COMPLETE for open-ended generation — narrower functions are cheaper and more reliable than routing everything through AI_COMPLETE.
2. Check token cost before batch jobs — Run AI_COUNT_TOKENS(model, text) on a sample before running a function over a large table.
3. Prototype in SQL — Call the function on a small LIMIT-ed sample and inspect results before running it over a full table.
4. Structure prompts explicitly — Use PROMPT('template {0}', arg) for parameterized prompts and cast JSON output to VARIANT when you need structured fields back.
5. For RAG, stand up Cortex Search — Create a CORTEX SEARCH SERVICE over the source table, query it via the Python or REST API to retrieve context, then pass that context into AI_COMPLETE.
6. Use dedicated compute — Size Cortex Search's backing warehouse no larger than MEDIUM, and separate it from other pipeline warehouses.
7. Guard against failures — Use TRY_COMPLETE in place of AI_COMPLETE for batch jobs where a single failure shouldn't fail the whole run; it returns NULL instead of raising an error.

Cortex AI Functions

Available functions (use these current names — do not use deprecated names like COMPLETE or CLASSIFY_TEXT):

Helper functions:

• TO_FILE('@stage', 'filename') — build a file reference for document/image processing.
• AI_COUNT_TOKENS(model, text) — check token count before calling a model, especially before large batch jobs.
• PROMPT('template {0}', arg) — build parameterized prompt objects for AI_COMPLETE.
• TRY_COMPLETE — like AI_COMPLETE but returns NULL on failure instead of raising an error.

AI_COMPLETE — the primary function

Supported models include claude-4-opus, claude-4-sonnet, claude-sonnet-4-5, claude-opus-4-5, claude-haiku-4-5, gemini-3-pro, llama3.1-70b, llama3.1-8b, llama3.3-70b, mistral-large2, mistral-small2, and deepseek-r1. Model availability varies by region — do not hardcode a model name without checking regional availability.

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

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

-- Structured JSON output
SELECT AI_COMPLETE(MODEL => 'claude-4-sonnet',
  PROMPT => 'Extract name, email, company as JSON: ' || raw_text)::VARIANT AS extracted
FROM contacts;

Classification, filtering, and aggregation

-- AI_CLASSIFY: cheaper than AI_COMPLETE for categorization tasks
SELECT AI_CLASSIFY(ticket_text, ['billing', 'technical', 'account', 'other']) AS category FROM tickets;
-- Multi-label: AI_CLASSIFY(input, categories, {'output_mode': 'multi'})

-- AI_FILTER: natural-language predicate directly in WHERE
SELECT * FROM reviews WHERE AI_FILTER(review_text, 'mentions product quality issues');

-- AI_AGG: cross-row aggregation with no context-window limit
SELECT AI_AGG(feedback_text, 'What are the top 3 themes?') FROM customer_feedback;

-- AI_EXTRACT: pull named fields out of free text
SELECT AI_EXTRACT(email_body, 'meeting date', 'attendees', 'action items') FROM emails;

Other functions

SELECT review_text, AI_SENTIMENT(review_text) AS sentiment FROM product_reviews;
SELECT AI_EMBED(description) AS embedding FROM products;
SELECT AI_PARSE_DOCUMENT(TO_FILE('@docs', 'contract.pdf'), MODE => 'LAYOUT');
SELECT AI_TRANSCRIBE(TO_FILE('@media', 'recording.mp3')) AS transcript;
SELECT AI_REDACT(customer_notes) AS redacted FROM support_cases;

Privileges

Cortex AI functions require the USE AI FUNCTIONS account privilege plus the SNOWFLAKE.CORTEX_USER database role — both are granted to PUBLIC by default, so most workloads work without extra grants.

Cortex Search — Hybrid Vector + Keyword Search

Cortex Search is fully managed search combining vector (semantic) and keyword (lexical) retrieval. Typical use cases are RAG for LLM chatbots, enterprise search, and AI-powered Q&A.

Single-index service

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);

Multi-index service (text + vector across columns)

CREATE OR REPLACE CORTEX SEARCH SERVICE my_multi_search
  TEXT INDEXES transcript_text, summary
  VECTOR INDEXES transcript_text (model='snowflake-arctic-embed-l-v2.0')
  ATTRIBUTES region
  WAREHOUSE = my_wh
  TARGET_LAG = '1 hour'
  AS (SELECT transcript_text, summary, region FROM support_transcripts);

Key parameters: ON (single-index column), TEXT INDEXES, VECTOR INDEXES, ATTRIBUTES (filterable columns), TARGET_LAG (freshness), EMBEDDING_MODEL, and PRIMARY KEY (enables optimized incremental refresh — set it whenever the source table has a stable key).

Querying — Python API (recommended for applications)

from snowflake.core import Root

root = Root(session)
service = root.databases["db"].schemas["schema"].cortex_search_services["my_search"]
resp = service.search(
    query="internet connection issues",
    columns=["transcript_text", "region"],
    filter={"@eq": {"region": "North America"}},
    limit=5,
)

Querying — REST API

POST /api/v2/databases/<db>/schemas/<schema>/cortex-search-services/<service>:query
Body: {"query": "...", "columns": [...], "filter": {...}, "limit": N}

Filter syntax

{"@eq": {"region": "NA"}}
{"@contains": {"tags": "urgent"}}
{"@gte": {"score": 0.8}}
{"@and": [f1, f2]}
{"@or": [f1, f2]}
{"@not": f}

Tuning relevance

Adjust the weighting between text match, vector similarity, and reranker score:

resp = service.search(
    query="billing dispute",
    columns=["transcript_text"],
    scoring_config={"weights": {"texts": 0.3, "vectors": 0.5, "reranker": 0.2}},
    limit=10,
)

RAG pattern

1. Retrieve context: results = service.search(query=question, columns=["content"], limit=5)
2. Pass the retrieved context into AI_COMPLETE:

SELECT AI_COMPLETE(
  MODEL => 'claude-4-sonnet',
  PROMPT => 'Answer using only this context: ' || context || ' Question: ' || question
);

Best Practices

• Use AI_CLASSIFY instead of AI_COMPLETE for categorization — it is purpose-built and cheaper.
• Run AI_COUNT_TOKENS before large batch jobs to estimate cost and avoid truncation surprises.
• Set PRIMARY KEY on a Cortex Search service so refreshes are incremental rather than full rebuilds.
• Use ATTRIBUTES for any column you need to filter on at query time.
• Use SEARCH_PREVIEW for interactive testing during development; use the Python or REST API for production integrations.
• Size the warehouse backing a search service no larger than MEDIUM, and dedicate it to that service.

Anti-Patterns

• Do not use deprecated function names (COMPLETE, CLASSIFY_TEXT, etc.) — use the current AI_* versions.
• Do not pass an entire table through AI_COMPLETE row-by-row without first estimating cost with AI_COUNT_TOKENS.
• Do not hardcode model names without checking regional availability.
• Do not use AI_COMPLETE for tasks a narrower function (AI_CLASSIFY, AI_FILTER, AI_EXTRACT) already covers.