Master The Cortex AI in Snowflake

Cortex AI in Snowflake is a comprehensive training curriculum that teaches ML engineers and data scientists to build, deploy, and operationalize AI applications entirely within the Snowflake platform. From LLM functions and vector search to fine-tuning, feature stores, and production RAG systems — every capability runs where the data lives, eliminating data movement and simplifying governance.

The curriculum follows a progressive learning path across four tiers: Foundation (Modules 01–02), Intermediate (03–05), Advanced (06–08), and Expert (09–10). Each module includes a professionally designed PPTX slide deck with architecture diagrams and code blocks, plus a hands-on lab project with real SQL and Python examples.

Topics Covered

• What is Cortex AI — LLM functions, ML models, and AI services within Snowflake
• Three-layer architecture — API Layer, Model Hosting, Vector Index & Search
• Model catalog — Mistral Large, Llama, Snowflake Arctic, Gemma 7B
• Cortex AI vs External APIs — data governance, latency, cost comparison
• Data governance advantage — “AI to Data”: data never leaves Snowflake
• Pricing — Token-based credit consumption per model tier
• RBAC for LLM access — GRANT USE_LLM, role-based model permissions
• • Region availability — AWS, Azure, GCP (all GA)

Hands-on Lab: Run Your First COMPLETE() Call

Execute CORTEX.COMPLETE with different models, compare response quality and latency, batch-process product descriptions, and configure role-based LLM access.

Topics Covered

• COMPLETE() — Free-form generation with temperature, max_tokens, top_p parameters
• Model selection strategy — Quality vs latency vs cost decision framework
• EXTRACT_ANSWER() — Structured extraction from support tickets and contracts
• SENTIMENT() — Scoring text from -1 (negative) to +1 (positive) at scale
• SUMMARIZE() — Configurable short/medium/long summarization
• TRANSLATE() — 30+ language pairs with auto source detection
• CLASSIFY_TEXT() — Zero-shot and few-shot categorization
• CTE chaining — Multi-step AI pipelines: classify → extract → summarize in one query
• Batch processing — Three tiers: direct (<1M), warehouse (1–10M), task-based (10M+)

Hands-on Lab: Customer Feedback Analysis Pipeline

Build a complete feedback analysis pipeline: ingest raw reviews, score sentiment, classify into categories (praise/complaint/suggestion), extract key entities, and aggregate insights — all in SQL.

Topics Covered

• Search architecture — INDEX (chunk + embed + store) → EMBED (query) → RETRIEVE
• Vector embeddings — EMBED_TEXT with voyage-3-large (1024-dim) and snowflake-arctic-embed-m
• CREATE CORTEX SEARCH SERVICE — Configuration, columns, warehouse, target lag
• Hybrid search — Vector similarity (semantic) + BM25 (keyword/lexical) combined
• Chunking strategies — Fixed, Sentence-based, Recursive, With Overlap
• RAG pipeline — Retrieve top-K chunks → Augment prompt with context → Generate answer
• Metadata filtering — AND/OR conditions on category, date, source columns
• Search quality tuning — similarity_threshold, num_results, reranking_enabled
• Incremental index updates — Timestamp-based, event-driven, periodic refresh

Hands-on Lab: Build a Q&A Chatbot with Cortex Search

Create a knowledge base from documents, build a Cortex Search Service, write RAG queries that retrieve relevant chunks and generate grounded answers with source citations.

Topics Covered

• Cortex Analyst architecture — NL Question → Semantic Model → LLM → SQL → Validation
• Semantic model YAML — Tables, columns, measures (SUM/COUNT/AVG), dimensions, time grains
• Measures and dimensions — Aggregation expressions and categorical groupings
• NL-to-SQL generation — 5-step flow with automatic validation
• Verified queries — Pre-approved SQL for common KPIs ensuring accuracy
• Multi-turn conversations — Context retention across follow-up questions
• Streamlit integration — Building self-service analytics dashboards
• Security integration — RBAC and Dynamic Data Masking automatically honored
• Accuracy tuning — Clear column names, business synonyms, sample queries

Hands-on Lab: Build a Self-Service Analytics Dashboard

Create an ecommerce semantic model YAML, configure Cortex Analyst, build a Streamlit app where business users ask questions like “What was revenue last quarter by region?” and get instant SQL-backed answers.

Topics Covered

• Document AI capabilities — OCR, field extraction, document classification
• Supported file types — PDF (50 MB), PNG/JPEG (20 MB), TIFF (100 MB multi-page)
• Creating models in Snowsight — 6-step UI workflow with sample documents
• Defining extraction fields — Invoice, receipt, contract field templates
• PREDICT() function — Extract structured JSON from PDF/image files
• Confidence scores — Routing: >90 auto-process, 70–89 review, <70 reject
• Batch processing — DIRECTORY(@stage) to process all documents in a stage
• Automated pipelines — Streams + Tasks for continuous document ingestion
• Fine-tuning accuracy — Iterative improvement with diverse samples and edge cases

Hands-on Lab: Automated Invoice Processing

Upload sample invoices to a stage, create an extraction model, use PREDICT() to extract fields (invoice number, date, vendor, total), validate with confidence scores, and set up a daily processing task.

Topics Covered

• Fine-tuning decision tree — Prompt Engineering vs RAG vs Fine-Tuning
• Supported models — Llama 3.1 70B (16 credits), Mistral Large (24 credits), Llama 3.1 8B (4 credits)
• Training data format — Prompt-completion pairs in a Snowflake table (100–1,000 examples)
• FINETUNE() function — Creating and managing fine-tuning jobs
• Hyperparameters — Epochs (2–5), learning_rate (1e-5 to 1e-3), batch_size, warmup_steps
• Evaluation metrics — Training Loss, Validation Loss, Perplexity, Token Accuracy
• Deploying fine-tuned models — Drop-in replacement via COMPLETE(‘ft-job-id’, prompt)
• A/B testing — Side-by-side comparison of base vs fine-tuned on test queries
• Cost optimization — Model size tradeoffs: 8B (fast, cheap) vs 70B (balanced) vs 123B (best quality)

Hands-on Lab: Fine-Tune for Domain-Specific Q&A

Create a training dataset of domain-specific prompt-completion pairs, run a FINETUNE() job on Llama 3.1, evaluate training metrics, deploy the fine-tuned model, and compare responses against the base model.

Topics Covered

• Snowpark ML overview — Python-native ML on Snowflake warehouse compute
• DataFrame API — pandas-like feature engineering with server-side execution
• Preprocessing module — StandardScaler, OneHotEncoder, LabelEncoder (built-in)
• Model training — XGBoost, LightGBM, scikit-learn running in-warehouse
• Feature Store — Entities, Feature Views, Pipelines for reusable features
• Point-in-time correct retrieval — Prevents future data leakage in training sets
• Model Registry — Versioning, lineage, metrics, deployment status
• UDF deployment — Scalar UDFs (real-time, 1 row/call) vs Vectorized UDFs (batch, 10K+/batch)
• Batch inference — Score millions of rows using vectorized UDFs

Hands-on Lab: End-to-End Churn Prediction Pipeline

Engineer features from customer data using Snowpark DataFrames, preprocess with built-in scalers, train an XGBoost churn model, register it, deploy as a UDF, and score the full customer base in batch.

Topics Covered

• Notebooks architecture — Container runtime with Python 3.10+, three cell types
• Multi-language cells — Python, SQL, and Markdown in a single notebook
• Package management — Anaconda channel, pip, Git-based installs without restart
• Snowpark integration — Pre-initialized session, DataFrame API, SQL interop
• Data visualization — Matplotlib, Plotly (3D/interactive), Altair (declarative)
• GPU compute — L40S (48 GB), A100/H100 (80 GB) for deep learning workloads
• Experiment tracking — MLflow integration: log params, metrics, models, compare runs
• Sharing and scheduling — Team collaboration, recurring notebook execution
• Git integration — 7-step workflow: connect repo → branch → develop → commit → PR → merge

Hands-on Lab: EDA to Model Deployment in a Notebook

Perform exploratory data analysis with Plotly visualizations, train a model with experiment tracking (MLflow), compare runs, deploy the best model as a UDF, and schedule the notebook for weekly retraining.

Topics Covered

• RAG maturity — Naive (simple retrieve+inject) → Advanced (hybrid, reranking) → Agentic (multi-hop, tools)
• AI agents — 5-step loop: query → analyze → call tools → reason → respond
• Guardrails — PII detection (regex + LLM scoring), content filtering, output validation
• Streaming responses — Real-time token streaming for chat interfaces
• Full-stack Streamlit apps — Three-layer architecture: UI → Logic → Data
• Evaluation metrics — Precision (retrieval relevance), Recall (coverage), Human Feedback
• Cost management — Response caching (30–50% savings), model routing (40–60%), token limits
• Monitoring — Latency alerts (>2s), error rate (>5%), credit burn (>100/hr)
• Hybrid architectures — Cortex AI + External APIs decision matrix

Hands-on Lab: Build a Production RAG Chatbot

Build a production-grade RAG chatbot with guardrails (PII blocking, toxicity filtering), streaming responses, source citations, user feedback collection, cost tracking, and a monitoring dashboard.

Topics Covered

• Capstone project — Multi-model AI app using 2+ Cortex services with Streamlit
• MLOps lifecycle — Develop → CI/CD → Monitor → Evaluate → Retrain
• Model governance — Lineage tracking, audit trails, access control (RBAC for models)
• Feature drift detection — LAG() + standard deviation thresholds for auto-retrain triggers
• A/B testing — Canary deployments: 5% → 25% → 100% traffic with metric gates
• Cost governance — Resource monitors, daily credit tracking, per-app attribution
• Security best practices — Dynamic masking, role separation, Private Link, prompt sanitization
• Cortex AI roadmap — Image understanding, fine-tuning APIs, multi-modal RAG, agent orchestration
• Career path — 4 levels: Foundational → Intermediate → Advanced → Expert with skill mapping

Hands-on Lab: Capstone Presentation

Present your capstone project covering: business problem, architecture, Cortex AI services used, guardrails, Streamlit demo, monitoring setup, cost analysis, and lessons learned.