Building effective generative AI workflows, especially agentic ones, involves navigating a vast space of possible configurations, which syftr aims to streamline through Pareto optimization.
Syftr, an open-source framework, automates the identification of Pareto-optimal workflows that balance accuracy, cost, and latency constraints in generative AI setups.
By using multi-objective Bayesian Optimization, syftr efficiently explores workflow spaces that manual testing cannot cover, leading to optimal configurations.
Syftr's approach involves evaluating around 500 workflows per run, with a Pareto Pruner mechanism to halt evaluations unlikely to improve the Pareto frontier, reducing computational costs.
Traditional model benchmarks are insufficient for understanding complex AI systems, where syftr excels by evaluating entire workflows to capture nuanced trade-offs in larger pipelines.
Syftr combines with tools like Trace for prompt optimization, offering a two-stage approach for workflow refinement and improved accuracy, particularly in cost-sensitive scenarios.
The framework's modular design allows easy extension and customization of workflow configurations by users, supporting various optimization strategies.
Syftr leverages open source libraries like Ray, Optuna, and Hugging Face Datasets, making it flexible for diverse tooling preferences and adaptable to different modeling stacks.
In a case study on CRAG Sports, syftr outperformed default pipelines, demonstrating significant improvements in accuracy and cost efficiency, showcasing the effectiveness of Pareto optimization.
Syftr's ongoing research includes meta-learning, multi-agent workflow evaluation, and composability with prompt optimization frameworks to enhance its capabilities further in generative AI design.