The paper introduces PALIMPZEST, a system designed to optimize AI-powered analytical queries through a declarative language. The system aims to simplify the process of executing complex AI tasks by allowing users to define queries in a high-level, declarative manner, while automatically optimizing the underlying execution plan. PALIMPZEST addresses the challenges of managing large, data-intensive AI workloads, such as legal discovery, real estate search, and medical schema matching, by providing a cost-effective, efficient, and high-quality solution. The system uses a cost optimization framework to generate and select the best execution plans based on runtime, financial cost, and output data quality. Key features include logical and physical optimizations, such as filter reordering, convert reordering, model selection, code synthesis, multi-data prompt marshaling, input token reduction, and output token reduction. The paper evaluates PALIMPZEST on various tasks, demonstrating significant improvements in speed, cost, and quality compared to baseline methods. The system is designed to be extensible and can handle parallelism, further enhancing its performance.The paper introduces PALIMPZEST, a system designed to optimize AI-powered analytical queries through a declarative language. The system aims to simplify the process of executing complex AI tasks by allowing users to define queries in a high-level, declarative manner, while automatically optimizing the underlying execution plan. PALIMPZEST addresses the challenges of managing large, data-intensive AI workloads, such as legal discovery, real estate search, and medical schema matching, by providing a cost-effective, efficient, and high-quality solution. The system uses a cost optimization framework to generate and select the best execution plans based on runtime, financial cost, and output data quality. Key features include logical and physical optimizations, such as filter reordering, convert reordering, model selection, code synthesis, multi-data prompt marshaling, input token reduction, and output token reduction. The paper evaluates PALIMPZEST on various tasks, demonstrating significant improvements in speed, cost, and quality compared to baseline methods. The system is designed to be extensible and can handle parallelism, further enhancing its performance.