GenAI - Snowflake GenAI Overview - Part II
1. GenAI - Snowflake GenAI Overview - Part II
Welcome back. In Part I we mapped the simpler end of Snowflake's AI capabilities. Now, we're going to move along the spectrum toward capabilities that take a bit more setup, but give you significantly more power and customization. Let's start with **Cortex Search**. Cortex Search is Snowflake's semantic search service. Traditional SQL search is exact-match — you're looking for records where a column equals a specific value or contains a specific string. But what if you want to search the way a human would search? You type a question in English and you want back the most semantically relevant results even if they don't contain the exact words you used. That's what Cortex Search does. You point it at a column of text in one of your Snowflake tables — like customer reviews, support tickets, product descriptions, or whatever you have — and Cortex Search builds a search service on top of it. Then you or your application can query that service with natural language and get back the most relevant results. This is the foundation of what people in the AI world call **RAG**, or Retrieval-Augmented Generation. The idea is that instead of relying purely on what an LLM knows from its training data, you give it access to your specific data as context. Cortex Search is the retrieval part of that pipeline, and it pairs naturally with the `AI_COMPLETE` function to build question-answering systems grounded on your data. Setting up a Cortex Search service looks like this. You're specifying which column to search on, giving it a warehouse for the indexing work, a refresh lag, and the underlying query that defines the data the service searches over. Once that's set up, you or your application can query it using a REST API or from within a Streamlit app. One small note: if you're setting up a search service on a trial account, you'll get an error. The search service uses embeddings that are only available on non-trial accounts, so you'll need to upgrade to a paid account to access this feature. Now, let's talk about **Cortex Analyst**. Cortex Analyst enables business users to ask questions in natural language and receive direct answers without writing SQL. It is available as a REST API, so it can be easily integrated into any application. The key ingredient that makes this work is a **semantic model** — a definition layer you build that describes what your tables and columns mean in business terms. Once Cortex Analyst understands your data model, it can reliably translate a business user's plain-English question into accurate SQL against your specific data, and do so reliably enough to power a production application. Now, let's cover **Cortex Fine-tuning**. The Cortex LLM functions we covered in Part I use general-purpose models. For most use cases, those general-purpose models work great. But sometimes you have a specific domain — like medical records, legal documents, or highly technical product data — where a general model just doesn't perform as well as you'd like. Cortex Fine-tuning lets you take a supported base model and fine-tune it with your own data, all within Snowflake. You provide training examples via pairs of inputs and ideal outputs, and Snowflake handles the training infrastructure. The result is a custom model that's tuned to your specific domain, and you can then call it just like you would call any other Cortex function. No infrastructure to manage, no GPUs to provision — it all happens inside Snowflake's compute environment. This falls into the heavier end of our spectrum because it takes more care and more data to do well. But for use cases where the general models aren't cutting it, fine-tuning is a powerful option that Snowflake makes significantly more accessible than the traditional approach. Finally, let's talk about what sits at the far right of our spectrum: **Snowpark Container Services**. Snowpark Container Services are Snowflake's most flexible option for AI workloads. If you want to bring your own model, use a partner LLM from the Snowflake Marketplace, run a completely custom AI workflow, or do something that none of the Cortex functions support out of the box, Snowpark Container Services is where you want to go. You package your code and its dependencies into a container, and Snowflake runs that container inside its own managed infrastructure, including GPU-enabled hardware for model training and inference. The main thing I want you to take away is where it sits on the landscape. It's the escape hatch for when you need complete customization and the ready-made Cortex features don't cover your use case. Let's bring this all together. Snowflake's AI platform spans a full spectrum: - At the accessible end, you've got **AI and ML Studio**, **Document AI**, and **Universal Search**. These are things you can use in seconds from the Snowsight UI. - Sitting underneath those and powering your SQL workflows, you've got the Cortex LLM functions, including `AI_COMPLETE`. - Moving along the spectrum, you've got **Cortex Search** for semantic retrieval, **Cortex Analyst** for embedding natural-language SQL into your apps, and **Cortex Fine-tuning** for domain-specific models. - At the fully custom end, **Snowpark Container Services** gives you the most flexibility to do almost anything. All of it runs on your data, all of it respects your governance policies, and all of it is available without leaving Snowflake. To recap: AI and ML Studio gives you no-code access to AI and ML features. Document AI extracts structured data from unstructured documents. Cortex Search enables semantic search and powers RAG workflows. Cortex Analyst translates natural language into SQL for embedding in your applications. Cortex Fine-tuning lets you customize a base model on your own domain data. Finally, Snowpark Container Services is the fully-custom option for AI workloads that need complete flexibility. In the next video, we're going to get hands-on with one of the Cortex LLM functions directly by using `AI_COMPLETE`, and we'll see just how easy it is to start working with LLMs inside your SQL workflows. Let's go.2. Let's practice!
Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.