Title: Enhancing Data Insights: Integrating Visualization Capabilities into Amazon’s Returns & ReCommerce Data Assistant

Subtitle: Transforming Natural Language Queries into Engaging Visual Analytics for Democratized Data Access

Overview:

In this second part of our series, we delve into the pivotal enhancements made to the Returns & ReCommerce Data Assist (RRDA), focusing on how we integrated Amazon Q into QuickSight, enabling a seamless transition from natural language inquiries to compelling data visualizations.

Key Highlights:

• Intent Classification for Visual Analytics: Discover how RRDA classifies user intents to deliver targeted insights.
• Q Topic Retrieval and Selection Workflow: Understand the sophisticated methods utilized for data topic selection to enhance visualization accuracy.
• Optimizing User Queries: Explore our methods for rephrasing questions to align with QuickSight’s capabilities for improved visualization outcomes.
• Embedding Visualizations in User Interfaces: Learn how we embed interactive visualizations within conversational interfaces for a cohesive user experience.
• Automated Metadata Management: Insights into maintaining an up-to-date knowledge base for effective data visualization sourcing.

Conclusion:

Join us as we examine best practices for implementing generative BI solutions and the ongoing journey towards a fully democratized data landscape at Amazon.

Empowering Data Insights: The Evolution of Amazon’s Returns & ReCommerce Data Assist

In the rapidly evolving world of data analytics, the need for accessible and actionable information has never been greater. In Part 1 of our series, we introduced how Amazon’s Worldwide Returns & ReCommerce (WWRR) organization successfully implemented the Returns & ReCommerce Data Assist (RRDA)—a revolutionary generative AI solution. This tool cleverly transforms natural language inquiries into validated SQL queries using Amazon Bedrock Agents, streamlining data retrieval for technical users. However, our journey towards truly democratizing data access doesn’t stop there.

Expanding the Vision: From SQL to Visualization

While the capabilities of RRDA enhance data accessibility for technical stakeholders, many users within WWRR yearn for visual insights rather than solely dealing with raw data or SQL results. The common demand is for quick, easily digestible trends and patterns that drive decision-making processes without the requisite technical know-how.

To address this need, beyond just SQL generation, we integrated visualization capabilities using Amazon Q in QuickSight. This enhancement allows users to express their inquiries in natural language, such as “Show me how many items were returned in the US over the past 6 months,” transforming them into engaging and informative visualizations.

The Architectural Framework

The architecture of the RRDA system comprises two principal pathways. Part 1 investigated the upper pathway responsible for generating SQL queries. In this installment, we explore the lower pathway, responsible for visual insights, situated within Amazon Q in QuickSight.

Key Features of RRDA Architecture:

• Intent and Domain Classification: User queries are classified to determine the necessary processing route—if a query aligns with visual analytics, it triggers a switch to leverage Amazon Q in QuickSight, with simultaneous domain identification to focus the search scope.

• Q Topic Retrieval and Selection: After classifying the intent as SHOW_METRIC, RRDA intelligently finds the most suitable Q topic from a vast catalog of over 50 configurations, identifying the best match for the user’s specification through vector search and an Amazon Bedrock foundation model.

Bridging Natural Language and Optimal Visualization

An inherent challenge in our system is translating user inquiries into formats that Amazon Q can optimally process. Users naturally express their needs in various ways, leading to potential ambiguity and missed insights.

To mitigate this:

• We harnessed the Amazon Bedrock Converse API to rephrase user questions into clearer, structured formats. This approach ensures that requests retain the user’s original intent while transitioning them into forms Amazon Q is designed to understand.

For instance, if a user asks for “How many items were returned lately?” our system reformulates the question into a more precise inquiry, improving response quality.

Delivering Insights in Real Time

When users request visual data through the RRDA chat interface, they experience seamless integration with Amazon Q in QuickSight. Utilizing secure embedding functionality, RRDA can display interactive visualizations directly within the conversation, providing a fluid transition from inquiry to insight.

Users can now view relevant Q topics presented as interactive suggestion cards, allowing them to select and generate visualizations promptly. This intuitive interaction streamlines the decision-making process, offering users an efficient pathway from inquiry to understanding.

Continuous Improvement: Automated Q Topic Metadata Management

Keeping our knowledge base current is paramount. We’ve implemented a robust, automated metadata management workflow using AWS Step Functions that refreshes daily. This workflow collects essential information on Q topics, ensuring our analytics landscape stays dynamic and responsive to user needs.

By automating feedback loops, we harvest validated user queries and responses, enriching our data, which in turn enhances the quality of our recommendations.

Best Practices for Implementing Generative BI Solutions

Based on our insights from extending RRDA’s visualization capabilities, we offer several best practices for enterprises embarking on similar generative BI initiatives:

1. Intelligent Suggestions: Utilize AI to anticipate user needs, surfacing relevant visualizations during metric discussions.
2. Automatic Translation Layer: Implement AI to convert natural language queries into the structured formats required by various systems.
3. Feedback Loops: Establish automated systems that gather user interactions to continuously improve the knowledge base.
4. Retrieval Before Generation: Employ vector search to identify relevant data sources prior to formulating optimal question content.
5. Maintain Domain Context: Ensure consistency in business context throughout user inquiries, whether they are receiving SQL outputs or visual insights.

Conclusion

In this two-part series, we have showcased how Amazon’s WWRR organization has extended the capabilities of the RRDA, creating a comprehensive solution that transforms natural language inquiries into both SQL queries and insightful data visualizations. This development marks a significant step toward democratizing data access, making analytics more intuitive and approachable for users across the enterprise.

As we continue to enhance RRDA based on user feedback and advancements in foundational models, we remain steadfast in our mission to connect natural language questions with actionable business insights.

About the Authors

• Dheer Toprani, System Development Engineer, specializes in large language models and scalable data systems.
• Nicolas Alvarez, Data Engineer, focuses on optimizing recommerce data systems and building advanced technical solutions.
• Lakshdeep Vatsa, Senior Data Engineer, is committed to designing and optimizing large-scale reporting solutions.
• Karam Muppidi, Senior Engineering Manager, leads teams in data engineering and analytics to drive AI adoption.
• Sreeja Das, Principal Engineer, focuses on system architecture transformations to enhance self-service experiences.

For more insights into building production-ready generative AI solutions, refer to the AWS Well-Architected Framework Generative AI Lens for best practices across critical operational areas.