Jump Ahead:
Generative AI & Agents
Explore all Generative AI capabilities in Dataiku
Real-Time Chat in Prompt Studios
Prompt Studios now includes a powerful chat interface, allowing you to test your fine-tuned and augmented LLMs and agents in a conversational setup as you’re building it.
The new chat mode enables you to:
- Test in a conversational environment: Check the behavior of your LLM or Agent in a chat setup that mirrors how it will function when deployed in Answers or Agent Connect.
- See responses in real-time: The streaming capability shows responses as they’re generated, providing immediate insights on output quality, with full visibility on sources for better control.
- Edit and iterate quickly: Modify questions, re-run completions, and save conversations for later, making tests more efficient.
Real-time chat helps streamline the Gen AI development process by allowing you to check the behavior of your AI application as you’re building it.
Check out the knowledge base to learn more about Prompt Studios:
Multimodal Evaluation
Multimodal AI is evolving quickly, but until recently, there wasn’t a good way to evaluate systems that work with both text and images. Dataiku 13.5 solves this problem by expanding our LLM Evaluation toolkit to include comprehensive multimodal assessment capabilities.
With documents like PDFs containing embedded tables, diagrams, and visual elements, traditional text-only evaluation methods fall short. Our new multimodal evaluation framework enables you to:
- Ensure accuracy and relevance across modalities: The built-in Multimodal Faithfulness metric verifies that AI responses accurately reflect visual content in source documents, while Multimodal Relevance confirms answers directly address prompts in the context of images.
- Visualize evaluation results: The row-by-row analysis now includes an image view, allowing you to see exactly what visual content the model was responding to when evaluating performance.
- Build for your specific use cases: Leverage provided code snippets to create custom metrics tailored to your unique multimodal scenarios, from image generation to complex document intelligence.
Multimodal Evaluation enables more trust and higher quality outputs across your multimodal use cases.
Check out the knowledge base to learn more:
AI Governance
Explore all Governance capabilities in Dataiku
Automatic Project Governance and Conditional Step Display
As AI governance requirements continue to evolve globally, Dataiku remains at the forefront of equipping data teams with the tools they need to maintain compliance without sacrificing productivity. Dataiku 13.5 takes another significant step forward with automatic project governance and more comprehensive custom governance workflow capabilities.
With 13.5, you can now decide to automatically govern all projects created on your synchronized instances, making governance and rigorous sign-off a critical part of your AI portfolio management process. No more gaps in governance coverage or delayed compliance processes—governance becomes a seamless part of how teams work. This capability is available in all versions of Dataiku Govern.
As regulatory requirements and new AI techniques emerge, so does the need to manage more tailored governance processes, adjusted to specific types of risks, used capabilities, or audit needs. To address this, we have added:
- Conditional workflow step display: Tailor governance processes to different project types by showing relevant workflow steps based on specific conditions, improving compliance without unnecessary complexity.
- Enhanced filtering for Custom Pages: Create more focused governance dashboards that surface exactly what oversight teams need to see, streamlining compliance reviews and audits.
Check out the knowledge base to learn more:
And a Few More Things
Dataiku 13.5 includes several additional enhancements to improve your data workflows:
- Upsert Recipe: Efficiently handle differential data with a new recipe that updates existing rows or inserts new ones in a single operation
- Data Quality Improvements: Export rows that fail quality checks to a separate dataset for easier investigation and remediation
- Brand Colors and Theming: Customize Dataiku’s color palette with your company’s brand colors and easily deliver dashboards aligned to your brand guidelines
- Enhanced Filters: Show only selected facets in visualization filters, making it easier to track context when working with complex data
- Time Series Forecasting: Leverage customizable prediction lengths and ARIMA with fixed orders for more precise time-series models
- Hugging Face Integration: Monitor kernel status with visual feedback and access detailed logs directly from the interface
Check out the documentation and knowledge base to learn more about these and other improvements in Dataiku 13.5.