SnowPro Specialty: Gen AI - Complete Exam Guide

SnowPro Specialty: Generative AI (SGA-C01) is Snowflake's first AI-focused certification, launched in 2025. It validates your ability to build generative AI on Snowflake using Cortex AI, LLM Functions, Vector Search, and the RAG pattern. This article breaks down each exam domain and the key points to study for every technology covered.

Exam Overview and Specs

The Gen AI Specialty exam is a Specialty-level certification that validates your ability to apply generative AI and LLMs on the Snowflake platform. It tests your knowledge of Snowflake's rapidly expanding Cortex AI ecosystem.

Exam Domains and Weighting

Cortex LLM Functions in Detail

Cortex LLM Functions are a set of SQL functions for calling LLMs directly from inside Snowflake. Their biggest advantage is that no external API keys or extra infrastructure are needed - you can use LLMs entirely under Snowflake's governance.

SNOWFLAKE.CORTEX.COMPLETE

The most basic LLM call function. It takes a model name and a prompt as arguments and returns the generated text.

-- Basic Cortex COMPLETE call
SELECT SNOWFLAKE.CORTEX.COMPLETE(
  'mistral-large2',
  'List three benefits of Snowflake's micro-partitions'
) AS llm_response;

-- Batch process table data
SELECT
  product_name,
  SNOWFLAKE.CORTEX.COMPLETE(
    'mistral-large2',
    CONCAT('Summarize the following product review: ', review_text)
  ) AS review_summary
FROM product_reviews
LIMIT 100;

Specialized Functions (SUMMARIZE / SENTIMENT / TRANSLATE)

Beyond COMPLETE, there are functions tuned for specific tasks. The exam asks you to decide when to use them versus COMPLETE.

• SNOWFLAKE.CORTEX.SUMMARIZE: Summarizes long text. Ideal for auto-summarizing reviews and meeting notes.
• SNOWFLAKE.CORTEX.SENTIMENT: Sentiment analysis on text. Returns a score from -1 (negative) to +1 (positive).
• SNOWFLAKE.CORTEX.TRANSLATE: Multilingual translation. Also supports automatic source-language detection.
• SNOWFLAKE.CORTEX.EXTRACT_ANSWER: Extracts the answer to a specific question from a body of text.

Vector Search and Embeddings

Cortex Search brings semantic search (meaning-based search) over text data to Snowflake. It surfaces semantic similarities that traditional keyword search (LIKE / CONTAINS) cannot capture.

SNOWFLAKE.CORTEX.EMBED_TEXT_768

Converts text into a vector (a 768-dimensional numeric array). You store these vectors in VECTOR-typed columns and search them via cosine similarity.

-- Embed text and store vectors
CREATE TABLE knowledge_base (
  doc_id INT,
  content VARCHAR,
  content_vector VECTOR(FLOAT, 768)
);

INSERT INTO knowledge_base
SELECT
  doc_id,
  content,
  SNOWFLAKE.CORTEX.EMBED_TEXT_768('e5-base-v2', content) AS content_vector
FROM raw_documents;

Implementing Similarity Search

-- Semantic search via cosine similarity
SELECT
  doc_id,
  content,
  VECTOR_COSINE_SIMILARITY(
    content_vector,
    SNOWFLAKE.CORTEX.EMBED_TEXT_768('e5-base-v2', 'Snowflake pricing model')
  ) AS similarity_score
FROM knowledge_base
ORDER BY similarity_score DESC
LIMIT 5;

RAG Pattern Design and Implementation

RAG (Retrieval-Augmented Generation) is an architecture pattern that gives an LLM domain-specific knowledge. The Gen AI exam puts heavy emphasis on implementing RAG on Snowflake.

Building Blocks of a RAG Pipeline

1. Document ingestion: load files from a stage or extract text from a table.
2. Chunking: split text into appropriately sized pieces. Balancing token limits and context quality is critical.
3. Vectorization: convert chunks to vectors with EMBED_TEXT_768 and store them in VECTOR columns.
4. Retrieval: embed the user query and pull the top-K matches by cosine similarity.
5. Generation: inject the retrieved context into the prompt and produce an answer with COMPLETE.

Using Cortex Search Service

With Cortex Search Service, you get a managed semantic search endpoint instead of having to hand-build the embed → vector store → similarity search pipeline yourself. It also handles automatic index refreshes and hybrid search (keyword + semantic).

-- Create a Cortex Search Service
CREATE CORTEX SEARCH SERVICE product_search
  ON content
  WAREHOUSE = compute_wh
  TARGET_LAG = '1 hour'
  AS (
    SELECT product_id, content, category
    FROM product_docs
  );

Cortex Analyst (Natural Language to SQL)

Cortex Analyst converts natural-language questions into SQL and generates answers against Snowflake data. The exam covers how to define a Semantic Model and the constraints Cortex Analyst operates under (for example, it does not generate DDL).

• Semantic Model (YAML definition): describes the meaning of tables, columns, metrics, and dimensions.
• Verified Queries: a mechanism for validating the SQL Cortex Analyst generates and feeding the result back in.
• Streamlit in Snowflake integration: build apps that call Cortex Analyst from a chat UI.

Building Apps with Streamlit in Snowflake

Streamlit in Snowflake (SiS) lets you run Python Streamlit apps inside Snowflake. The Gen AI exam covers patterns for building chatbots and RAG apps with SiS.

• Build chat UIs with st.chat_input / st.chat_message.
• Call Cortex LLM Functions through the Snowpark session.
• App access control: set the scope of access by role via RBAC.
• SiS constraint: external network access requires an External Access Integration.

Governance for AI Usage

Governance is non-negotiable when generative AI moves into production use. The exam tests Snowflake-specific AI governance features.

• Trust Center: verify Cortex AI's data handling policies and privacy guarantees.
• Region restrictions: Cortex LLM models are only available in specific regions, with constraints on cross-region calls.
• Token cost management: Cortex LLM Functions are billed in credits. Use Resource Monitors to cap usage.
• Audit logs: Cortex function call history is traceable through the QUERY_HISTORY view.

Study Strategy for Passing

• Use the free Snowflake trial to call Cortex LLM Functions and see the responses for yourself.
• Read the Cortex Search Service docs end to end and memorize the CREATE CORTEX SEARCH SERVICE syntax.
• Do a hands-on lab implementing the RAG pattern with Streamlit in Snowflake.
• Summarize the characteristics of embedding models (e5-base-v2 and others) and the criteria for choosing one.
• Read every documentation link listed in Snowflake's official Gen AI Study Guide.

Check Your Understanding

Frequently Asked Questions