New Dataiku RAG Capabilities to Expand Your GenAI Powers

Dataiku Product, Scaling AI, Featured Nanette George

RAG chatbots are one of the most highly adopted GenAI use cases in enterprises around the world. However, organizations are discovering that building truly effective GenAI applications requires more than connecting an LLM to a document store. Indeed, critical blockers persist. For example, when you try to add a PDF report with embedded graphs and tables into a traditional RAG pipeline, the LLM can fail to retrieve the information accurately.

These challenges can significantly impact the accuracy and reliability of retrieval augmented generation (RAG) implementations. They aren't just technical hurdles — they are barriers that can make the difference between a GenAI application that transforms your business and one that falls short of expectations.

New RAG Capabilities for Better Outputs

At Dataiku, we've continuously evolved our RAG capabilities based on real-world feedback from enterprises implementing GenAI at scale. Building on our foundation of streamlined RAG workflows, we're excited to introduce new capabilities that directly address our customers' most pressing challenges when building production-ready RAG applications.

1. Hybrid Search for Enhanced Retrieval Precision

Most RAG implementations rely on semantic search to find related concepts, but this approach is limited when dealing with industry-specific jargon. For example, searching thousands of clinical documents to find a specific medical procedure using traditional semantic search can be a challenge. While the search might understand the general concept, it could miss crucial protocol numbers or specific technical terms.

This challenge isn't unique to healthcare — from financial regulations to engineering specifications, organizations across industries struggle to balance semantic understanding with precise terminology matching.

What's New: We've engineered a hybrid search capability for Azure AI Search and ElasticSearch/OpenSearch vector stores that intelligently blends semantic understanding with keyword precision. It is like having a domain expert and a semantic search engine working in tandem — one catches the technical specifics while the other grasps the broader context.

This sophisticated fusion of search methodologies automatically adapts to your domain's unique requirements, ensuring conceptual relevance and terminological accuracy in your retrieval results. The result is more accurate and reliable information retrieval, especially in technically complex domains.

Screenshot of user’s test in the Dataiku RAG Augmented LLM configuration pane. Chosen large language model is GPT-4o mini.Dataiku RAG users now can choose Hybrid Search to combine a similarity search and a keyword search to retrieve more relevant documents.

2. Embed Documents Recipe for Seamless Multimodal Processing

Organizational knowledge doesn't live in text alone. Critical insights often hide in document-based tables, technical diagrams, and embedded images. Traditional RAG approaches often ignore these rich sources of information, which is unacceptable for teams that are serious about comprehensive knowledge capture.

What's New: Our embed documents recipe represents a breakthrough in multimodal document processing. This zero-code solution seamlessly handles document content, including text, tables, and images, in a unified workflow, much like having a team of specialists to analyze your documents.

By leveraging state-of-the-art vision language models for images and specialized LLMs for text and tables, the recipe ensures that all document information is appropriately captured and embedded for retrieval. No more broken tables, missed image insights, or complex coding.

Screenshot of a user’s work in the Dataiku Prompt Studio. The prompt and two test cases are shown.The RAG embed documents recipe allows Dataiku users to turn text and image information contained in PDF documents into numeric vectors that can be searched.

3. RAG Quality Guardrails for Reliable AI Responses

When organizations deploy a mission-critical RAG application, they want the app to consistently provide accurate responses based on the organization's knowledge base. Traditional RAG implementations might generate plausible-sounding but potentially inaccurate responses or drift away from the original query intent. This uncertainty can pose significant risks for regulated industries like healthcare or financial services and erode trust in AI systems.

What's New: We've introduced comprehensive RAG quality guardrails that act as an intelligent quality control system for AI responses. This innovative capability performs a live evaluation of RAG outputs using two sophisticated LLM-as-a-judge metrics: faithfulness and response relevancy. 

It’s like having an expert reviewer who instantly validates every response — checking both how accurately it reflects the source documents and how well it addresses the user's query. GenAI builders can set specific quality thresholds at the model level in their knowledge base. When responses don't meet these standards, they can configure explicit error messages or custom fallback responses, ensuring that only high-quality, trustworthy information reaches users.

Screenshots of a user’s work in the Dataiku Augmented LLM configuration pane. The user’s chosen LLM is GPT 4.The RAG quality guardrails allow Dataiku users to choose guardrails, including “faithfulness” and “relevancy,” and set a scoring threshold for each, as well as an action to take when scores fall below each threshold. Users also can create a custom answer to provide end users of the LLM when thresholds are not met.

Why It Matters

These three new capabilities represent a significant leap forward in making RAG implementations more powerful and accessible:

  1. Hybrid search ensures that your RAG applications can navigate the nuanced territory of specialized domain knowledge with unprecedented precision.
  2. The embed documents recipe breaks down the barriers between different content formats, ensuring that valuable information does not remain locked away in tables or images.
  3. RAG quality guardrails ensure GenAI applications maintain high accuracy and relevance standards, a particularly crucial capability for enterprise deployments.
Together, these capabilities solve technical challenges and fundamentally transform what's possible with RAG applications. Organizations can now build more comprehensive, accurate, and reliable GenAI solutions while dramatically reducing the technical complexity and development time required. More broadly, better RAG pipelines can improve a wide array of GenAI applications, from chatbots to AI agents. Increasing quality inputs results in better, more valuable outputs from GenAI.

Catch the GenAI Wave With Dataiku

We listen to our customers and the broader marketplace to add powerful capabilities to the Dataiku platform. Whether you're in healthcare, financial services, manufacturing, or any other domain with complex knowledge requirements, Dataiku can provide the foundation for a new wave of more sophisticated and practical GenAI applications for your organization.

You May Also Like

How Aviva Uses Dataiku to Unlock Deeper Customer Insights

Read More

How Aviva Uses Dataiku for Enhanced Scalability & Efficiency

Read More

The Evolution of NLP (& How Dataiku Can Help)

Read More

A Dizzying Year for Language Models: 2024 in Review

Read More