Parcel Perform’s AI: Text-to-SQL for E-commerce

Parcel Perform, a leading AI Delivery Experience Platform, has revolutionized its data access with a generative AI-powered text-to-SQL solution. This innovative system empowers customer-facing business teams to directly query parcel delivery data using natural language, eliminating the need for manual SQL queries by the data team. The solution leverages AWS services including Amazon RDS, S3, Athena, MSK, EKS, Glue, Bedrock, and OpenSearch Serverless. Data ingestion begins with Amazon RDS, then flows through Amazon MSK to Apache Flink on Amazon EKS, which writes data to S3 in Apache Iceberg format. Amazon Athena provides the query engine for this data lake. The AI agent itself uses LangGraph for workflow orchestration, Anthropic's Claude models in Amazon Bedrock for SQL generation, and Amazon Bedrock Knowledge Bases for retrieving business context. This RAG (Retrieval Augmented Generation) approach enhances accuracy by incorporating business rules and industry-specific terminology not inherent in general LLM training. The UI is built with FastAPI, hosted on Amazon EKS, and fronted by an Application Load Balancer. The system addresses challenges such as ambiguous merchant names (solved via prompt engineering and planned user input clarification) and potentially expensive queries (mitigated by enforced query limits and planned EXPLAIN query analysis). The impact is a 99% reduction in time-to-insight, from an average of 2.3 days to 10 minutes, saving roughly 3850 hours monthly and freeing the data team for more strategic work. While the system offers significant improvements in data accessibility and efficiency, potential drawbacks include the need for careful prompt engineering and ongoing monitoring for query cost optimization. The solution is a compelling example of how generative AI can streamline data access, improve decision-making, and enhance customer experience within e-commerce businesses.

(Source: https://aws.amazon.com/blogs/machine-learning/democratize-data-for-timely-decisions-with-text-to-sql-at-parcel-perform/)