Agents and Tools for Generative AI¶
Large Language Models (LLMs) are powerful but have inherent limitations in areas like real-time data access, mathematical computations, and specialized domain knowledge. To overcome these constraints, you can extend LLM capabilities with external functions, structured workflows, and integrations with other systems.
This section explores key approaches for enhancing LLM performance, from basic tool integration to complex agentic behaviors.
Concepts & examples
You can find code samples on this subject in the Developer Guide: Agents.
Agents¶
Visual Agent and custom tool¶
To build a custom tool, follow Creating a custom tool. Once you have built a custom tool, or if you want to use tools provided by Dataiku, you can follow Leveraging a custom tool in a Visual Agent.
Code Agent¶
To learn how you can use Code Agents in Dataiku, see Creating and using a Code Agent. If you want to package your agent as a plugin, you should follow Creating a custom agent. You may also need to define your own LLM connection, as described in Creating a custom LLM Connection.
Custom Python Tool¶
To build a tool with inline code, in a simplified experience compared to a custom tool packaged in a plugin, see Creating a Custom Python Tool.
Multi-agent systems¶
To make several agents work together, which you will need when working on complex tasks, see Multi agents: sequential workflow.
Integrating an agent framework¶
To use an agent framework in a Code Agent, see Integrating an agent framework.
LLM Mesh agentic applications¶
The LLM Mesh agentic applications series demonstrates how to build agentic applications using the LLM Mesh in Dataiku, through Defining and using tools with the LLM Mesh, Creating an LLM-based agent that uses multiple tools, and Building a Web Application with the agent.
Integrate with external Vector Stores¶
To connect to an external Vector Store using the Custom Python Tool, see Connecting to an external Vector Store.
Model Context Protocol (MCP)¶
The Model Context Protocol (MCP) series demonstrates how to build custom and third-party MCP servers using Code Studios and webapps in Dataiku, starting with Building your MCP Server in Dataiku.
Langchain agents¶
In addition, you could also build agents in Dataiku using the Langchain framework. Langchain enhances LLM capabilities by integrating planning, memory, and tools modules. This allows LLMs to perform more complex tasks like accessing databases or interfacing with other software.
You can find the tutorial in Building and using an agent with Dataiku’s LLM Mesh and Langchain.
Using a Dataiku Agent in an external application¶
To use a Dataiku Agent in your application outside of Dataiku, see Using a Dataiku Agent in an external application.
Processing¶
Using JSON outputs¶
To process and get structured outputs via the LLM Mesh, see Using the LLM Mesh to parse and output JSON objects.
Prompt¶
Auto Prompt Strategies with DSPy¶
To use an auto-prompting library, see Building Auto Prompt Strategies with DSPy in Dataiku.
Logging and Monitoring¶
Adding traces to your Agent¶
To use traces to help you understand your agent’s behavior, see Adding traces to your Agent.
Agent logging¶
To set up and use agent logging, see Agent logging.