BUILD 2024: Agents & multi-modal apps are the tip of the “Iceberg”

Developers came together virtually and around the globe from Nov 12th through the 15th for BUILD, The Dev Conference for Apps & AI . We collected the best across product announcements, luminary talks, technical sessions, hands-on labs, and Gen AI bootcamps from this week, directly for you.

Highlights

✨ What’s New ✨ on Snowflake 

• 🤖 Snowflake Intelligence (PrPr soon) enables your business users to ask questions across structured and unstructured enterprise data through data agents you create. 👉 Check out Sri and Dash's live demo here 👈
• 😎 Unistore, powered by Hybrid Tables, is now generally available  
• 🧊⛰️Snowflake Open Catalog uses the Apache Iceberg™ REST protocol to integrate external data engines and apply your organization’s governance controls across them. Snowflake Open Catalog is a managed version of the open source Apache Polaris (Incubating) project you might remember from Summit.  
• 🥳 Multi-modal 3.1 and 3.2 models from Llama, Anthropic’s Claude 3.5 models , + more languages, context window choices and embedding model sizes , support for container runtime for your notebook based AI workflows 
• 🛒 Develop, distribute, and monetize AI models to third-parties on Snowflake Marketplace with Secure Model Sharing

💡 Luminary Talks 💡

Catch up with Dr. Andrew Ng , founder and Executive Chairman of LandingAI, on the rise of AI, agentic reasoning and the role of unstructured data such as text, images, video and audio on automation and application building across industries.

Also, for the first time this year, we brought you some lessons directly from the data trenches. Three data leaders joined us on stage to talk about how to:

• work with the thorniest data and even generate missing data to feed your AI models
• best practices for building AI augmented BI workflows; and
• what needs to change in your warehouse / lakehouse to get you and your team ready for AI workloads

Watch the replay featuring Katie Bauer (Head of Data at GlossGenius), Kasia Sitkiewicz (Staff Product Manager at GitHub) and Holden Karau (Open Source Engineer at Netflix).

💙 Thank you to our partners 💙

Our partners showcased their integrations with Snowflake through a series of labs and technical sessions covering trust & safety tools, leveraging the GPU-powered Container Runtime in Snowflake Notebooks, leveraging Cortex AI for data interpretation, using LLMs to tackle text data processing, and more. 

BUILD it yourself

Structuring with Schemas

Start the journey to build a reliable text-to-SQL product with joins across multiple tables. Snowflake Cortex Analyst now supports joins for star and snowflake schemas in public preview. Learn more . 

What are star and snowflake schemas? 

• Star schema: central “fact table” with “dimension tables” that provide context for those facts
• Snowflake schemas: an extension of star schemas allowing for breaking down dimension tables into more granular tables 

Runtime to Shine

The recently launched public preview of Container Runtime that sits on top of Snowpark Container Services (SPCS) provides a fully-managed container environment that distributes processing on CPUs and GPUs for popular Python libraries and frameworks. 

If you want to train a complex ML model in Python, you can leverage a superpower of helpers in the form of CPUs (data processing tasks) and GPUs (machine-learning, mathematical tasks)  to train your model 3-7x faster than if you did it yourself. Snowflake Notebooks is the interface where you code and train your model. 

To access Container Runtime, simply create a Snowflake Notebook and choose “Run on the container” as the backend option and explore the blogpost below for more.

 Search and you shall receive

Cortex Search (now in public preview) is a fully managed service for retrieval from unstructured data sources (e.g., documents, text files, etc) in your enterprise RAG stack. 

How does it work? 

#1 Build efficient indexed data structures 

• Construct a vector index: text from documents is converted to vector form using Snowflake’s Arctic-embed model to allow for efficient semantic search 
• Term index: tokenize the text to create that links terms to the documents they appear in, to allow for efficient keyword search

#2 Index-Serving 

• Purpose-retrieval backend: after indexing, documents are served via a backend that supports both vector and keyword retrieval,
• Query processing: when a query is received, it's understood and expanded by adding synonyms, stemming, and embedding it into a vector
• Multistage storing: progressively refining search results for relevance and accuracy

We’re Not Done Yet! 

In case you missed it, you can still access the sessions on-demand or view them from your existing registration. Watch now !  

Join a BUILD Meetup in-person event at a local user group near you. Find your city and register today at the link below to complete a workshop on Cortex Analyst for a chance to earn a badge, and meet peers in your community!