One Week with Dataiku Cobuild
I've spent the past week experimenting with Dataiku Cobuild, the new AI building agent in Dataiku 14.7. If you’ve been living under a rock or your Dataiku Marketing emails go directly into your spam folder, Cobuild allows you to build and optimize analytic solutions in Dataiku via natural language. Why use Dataiku Cobuild rather than one of the many other options? The primary reason is because the result of your conversation is a visual flow, not 1000s of lines of AI generated code, that you can easily inspect, edit, and approve.
My first impression has been an eye-opening experience to see how it bridges the gap between ideation and execution. In this short blog I will share some of the capabilities that I have been exploring and hope to build on going forward.
Generate Instant Insights
Dataiku Cobuild transforms natural language into queries and visual charts but the true differentiator is in its ability to create reusable assets. As you’re building your projects with Cobuild, you can ask questions about the assets you are creating along the way. In this scenario, Cobuild acts as your readily available expert.
In my simple example shown in the video below, Cobuild queried the data to show me a chart of weekly sales for the Downtown store. The chart was shown directly in the chat window and while this interaction is cool, it’s more valuable when the results are easily repeatable and shareable.
Cobuild is uniquely integrated into the core Dataiku platform, allowing for bi-directional interactions. Cobuild is aware of your active screen (within Dataiku) and is also able to make immediate modifications to your project. We can save this chart to add it as a visualization on the dataset in our flow. Each time the data is updated, the chart will also. We could take this a step further and now publish this chart to a dashboard within our project for sharing more broadly.
Streamlining Development-to-Production
Promoting projects to production often involves a complex setup that can deter users from promoting their work, but Cobuild alleviates any complexity and follows best practices for pushing to production. Failing to properly productionalize your work can have a huge negative impact on your organization by facilitating decision making on stale or inaccurate data and analysis.
In the video below, I wanted to rerun the store forecasting model I trained (with the help of Cobuild of course) each week so that I can more efficiently staff my store. On days with higher predicted store sales, I need additional staff to handle the volume.
I follow the instructions outlined by Cobuild to take this forecasting project and move it from my development environment to production. In most cases, I’m simply approving the recommended actions and Cobuild builds the necessary assets for me! For example, it starts out by recommending a scenario to automate the execution of the forecast. This will rerun the forecast for next week. Cobuild was recommending a few additional tasks to prepare the project for production like failure notifications and metrics and checks on data but I decided to move the project to my production node at this point.
Cobuild is aware of your Dataiku environment and is able to create a project bundle that is used to move the project to the deployer node via a scenario step. If you’re unfamiliar with these terms, this acts as the “plumbing” that handles the movement of your projects from development to test and production environments. Following this process is a highly recommended best practice because it allows you to update the production project more easily in the future when you’re ready to push a v2!
Customize Your Cobuild Experience
You can tailor Cobuild’s behavior by appending custom instructions to its system prompt, making it adapt to your unique preferences or organizational standards. This may be my personal favorite highlight given the infinite possibilities it enables. Every organization, user, and even project can have a different Cobuild experience that’s customized to their preferences. Here are a few examples that I experimented with last week:
- Accelerated Onboarding and Tool Bridging: Cobuild can help map terminology and functionality from legacy environments (such as Alteryx or SAS) directly to Dataiku. This minimizes the learning curve for transitioning users, allowing them to remain productive while they acclimate to the platform.
- Scalable Governance and Quality Standards: By embedding corporate best practices into the system prompt, Cobuild becomes a way to improve adoption and enforcement. It can provide customized recommendations that ensure every project meets organizational quality benchmarks before promoting to production, reducing the "glue code" and manual review time that typically slows down development cycles.
- Persona Specific Efficiency: Users can tailor Cobuild outputs to match their unique working styles, whether they prefer a code-first approach or a GUI-centric visual flow. Adjusting Cobuild's communication style through its tone, verbosity, and depth of technical jargon ensures that the interface remains engaging and efficient for both seasoned data scientists and newer business analysts.
You can see an example of this in the short video below. In this video I asked Cobuild to give a basic explanation of the current project. Although not shown in the video, I enriched the system prompt for my project with the guidelines here to help a new user transitioning from Alteryx to Dataiku. One of the biggest deterrents of adopting a new tool is the temporary drop in efficiency as you learn the user interface, naming conventions, and best practices. Augmenting Cobuild with specifics about the legacy tool you’re familiar with can drastically improve productivity during this adoption period and allow you to unlock the additional value of Dataiku!