dspy

Automate prompt engineering by composing declarative modules and optimization loops. You define task signatures, supply a labeled trainset and evaluation metric, then run a teleprompter to automatically generate and refine prompts for deployment.

Declarative LM programming works by defining task signatures and composing modular reasoning steps like ChainOfThought or ReAct. This structure replaces manual prompt tweaking with composable modules that execute multi-step reasoning and tool use.

Yes, you can optimize RAG pipelines and few-shot prompts across models. By configuring an LM provider and applying optimizers like BootstrapFewShot or MIPRO, the system automatically improves module performance based on your defined metric.

Yes, you need a labeled trainset and a custom metric to run language model optimizations. The optimizers require labeled examples to bootstrap few-shot prompts and a metric to evaluate and guide the automatic improvement of your modules.

The best way to reduce manual prompt overhead is using automated teleprompters. Optimizers like COPRO and BootstrapFinetune systematically tune prompts and weights, replacing manual adjustments with versioned, self-improving modules for production workflows.

Declarative signatures provide clearer task definitions and better maintainability than manual prompt engineering. By defining inputs and outputs structurally, you enable optimizers to automatically tune the underlying prompts, ensuring scalable and consistent performance.