Building Agents With Snowflake Cortex AI at Every Level

Use Cases & Projects, Scaling AI, Featured, Partner Patrick Phrayme

Organizations care about tech optionality — not just to avoid vendor lock-in, but also to make sure they’re using the best tool for the job. This particularly applies to AI agents. The more tools and systems at an agent’s disposal, the better the agent. Dataiku has always aimed to put these tools in the hands of a diverse range of users, which is why we’re excited to provide no-code UIs to build AI agents leveraging the entire Snowflake Cortex AI stack. 

Along with improvements to Dataiku’s existing connector to Cortex AI LLMs, we are introducing new integrations with Snowflake Cortex Search, Snowflake Cortex Analyst, and Snowflake Cortex Agents. Many organizations trust the Cortex AI suite for LLM inferencing and information retrieval. With Dataiku, they can expand investments in Cortex to new use cases and new builders of agentic applications. 

Dataiku continues to be one of Snowflake’s leading partners on AI and ML integrations. We’re excited to continue this momentum and work with them to build impactful, enterprise-grade AI agents with our joint customers.

- Baris Gultekin, Head of AI, Snowflake

Snowflake Logo Blue

The LLM Level: Cortex LLMs

The Dataiku LLM Mesh is a secure API gateway to break down hard-coded dependencies, managing and routing requests between GenAI applications and LLMs hosted locally or through a provider, like Snowflake. Behind the scenes, we’ve changed supported Cortex LLMs from JDBC SQL queries to REST API calls.

This unlocks token streaming and tool use. With token streaming, your Dataiku Answers chatbot powered by Deepseek R1 on Cortex AI will now return responses more naturally as they come in, instead of the entire response all at once. With tool use, your visual and code-based Dataiku agents backed by Claude 3.5 Sonnet on Cortex AI (or any other Cortex LLM that supports tool calling) can, well, use tools. 

Snowflake Cortex LLMs

The Tool Level: Cortex Search and Cortex Analyst

Snowflake offers two managed services for information retrieval that can be helpful tools to add to an agent:

  • Cortex Search performs keyword and vector search plus semantic reranking over unstructured and text data in Snowflake.
  • Cortex Analyst is a generative text-to-SQL chat API over structured tables in Snowflake.

We’re releasing two new no-code agent tools that cover these services. This means that organizations can easily take their already-built Cortex Search services and Cortex Analyst semantic models and add them as tools to Dataiku agents.

new agent tool

For example, here we’ve built a visual agent to assist with consumer lending decisions. The agent is of course powered by Claude 3.5 Sonnet on Cortex AI, and has three tools:

  • Dataset lookup: To look up information about an application from a table in Snowflake
  • ML model prediction: A random forest model trained in Dataiku to predict whether an applicant is likely to default on the loan
  • Cortex Search: To search internal policy documents and make sure an applicant or loan offer doesn’t violate internal policies

loan prediction agentWe can add this agent to Agent Connect, then ask it to offer some informed lending advice on an application.

loan advisory agent

The Agent Level: Cortex Agents 

Snowflake offers a higher-level agentic workflow API called Cortex Agents that includes planning, tool use (Cortex Search and Cortex Analyst), reflection, and monitoring. 

We’re releasing a no-code agent plugin to “create” Cortex Agents on the fly. Simply choose a Cortex Agent-supported LLM (like Llama 3.3 70B here) and enter the paths to your Cortex Search and Cortex Analyst services.

new agent in Dataikuversion 1

Then, you can use this agent across your Dataiku projects, like as an input to a prompt recipe for batch agentic operations on an entire dataset, or as one agent behind a multi-agent Agent Connect chatbot.

Dataiku thrives at structured agent evaluation. We can assemble some trusted sample queries and answers, then evaluate the performance of this Cortex Agent with metrics like answer correctness, BERT score, BLEU, and ROUGE. Some of these metrics require an LLM-as-a-judge — for that we can choose Deepseek R1 on Cortex AI.

evaluate data verification

When Should You Use These New Integrations?

Cortex LLMs: Use these in any Dataiku LLM recipe (prompt, classify text, summarize text, embed dataset) or when powering any Dataiku agent or chatbot.

Cortex Search and Cortex Analyst Tools: Use these tools to equip any Dataiku visual or code agent with an existing Cortex Search service or Cortex Analyst semantic model for RAG or text-to-SQL use cases. You can customize the agent as you’d like, including adding any of Dataiku’s additional managed tools.

Cortex Agents: Use this to create an Agent that uses only Cortex Search and Cortex Analyst tools. Let Snowflake's proprietary agentic workflow handle the routing to these tools.

You May Also Like

The Architecture Behind a High-Impact CoE for AI Agents & Self-Service

Read More

Dataiku Named a Gartner Magic Quadrant Leader for 4th Consecutive Year

Read More

5 AI Agent Use Cases to Kickstart Your Team's Transformation

Read More

Create and Control AI Agents at Scale With Dataiku

Read More