Cobuild

Cobuild is Dataiku’s AI building agent that builds data pipelines, analytics, machine learning, multi-agent workflows, applications, and automation pipelines inside Dataiku from trusted data, with governance built-in.

You can find Cobuild within a project’s top nav bar, on the right.

Cobuild is an AI Service, and your administrator needs to setup AI Services first (see Overview and setup).

Note

Cobuild replaces Flow Assistant starting in version 14.7.

What can Cobuild do

The following list is not exhaustive but covers the main use-cases that Cobuild supports.

Cobuild can identify the right datasets for a use case and generate Flows, including:

  • Visual and code recipes

  • ML models

  • Knowledge Banks

  • Agents

Cobuild can also build Charts and Dashboards, Jupyter notebooks, Webapps and Wiki.

Cobuild can build scenarios to bring all of these together.

Cobuild can ask for explicit confirmation before key actions. Every output lands as an inspectable, editable Dataiku Flow.

Cobuild can also inspect and describe your flow, datasets, recipes (including code recipes), web apps, notebooks, etc.

Cobuild can modify and refactor existing flows, webapps, notebooks, charts, dashboards, …

Cobuild can run recipes, inspect job logs, and provide guidance on errors.

Giving context to Cobuild

Custom instructions

Cobuild is most useful when it understands you, your preferences and your business. To achieve this, Dataiku offers you different levels of customization for the system prompt that will guide Cobuild’s actions and behavior:

  • Global instructions (added to Cobuild instructions for all projects by default). They can be configured from the AI Services admin panel.

  • Project-level instructions (added to global instructions or replacing them for a specific project). They can be configured from the project settings page, in the Cobuild tab.

  • User-level instructions (added to global and project instructions or replacing them for a specific user). They can be configured from the user settings page, in the Cobuild tab. Available in DSS 14 since 14.7.4 and DSS 15 since 15.0.1.

Instance administrators can disable project-level instructions if needed, and can also prevent project-level and user-level instructions from overriding global instructions.

Prompt-level context

When you send a message to Cobuild, it will know which page you are currently looking at, and which objects in the Flow are currently selected, so you can write prompts such as “filter these datasets by year > 2025”. You can also explicitly mention DSS objects by name, by typing @ and selecting the desired object from the autocomplete suggestions.

Supported recipe types

Core Visual recipes

  • Prepare

  • Filter/Sample

  • Sync

  • Group

  • Sort

  • Join (including fuzzy join and geo join)

  • Distinct

  • Top N

  • Pivot

  • Stack

  • Window

Code recipes

  • Python

  • SQL Query

Plugin recipes

ML recipes

  • Training

  • Prediction

GenAI recipes

  • Embed datasets

  • Embed documents

Other

  • Extract failed rows

Cobuild-specific settings

You can find Cobuild admin settings in Administration > Dataiku AI & Cobuild. Your first decision will be the way Dataiku AI connects to your LLM, either via Dataiku AI Services (managed by Dataiku) or with your own LLM going through the LLM Mesh. For general AI assistants configuration, see Overview and setup.

In particular, it is possible for you to configure custom Markdown instructions which will be included in all Cobuild sessions. Keep it tight and informative, to tell Cobuild about your domain, global conventions, etc. You can choose whether this custom prompt can be overridden at project level or not. You will be able to define a custom prompt for each project by going to the project page and then Settings > Cobuild. It can either override global instructions (if allowed by instance-level admin settings), or be appended to it.