dspy

You can automate prompt optimization by applying teleprompter optimizers like BootstrapFewShot or MIPRO to your training data. This declarative programming approach automatically generates and refines prompts, eliminating manual trial-and-error when building scalable AI pipelines.

Building multi-step reasoning tasks is best achieved using declarative modules like ChainOfThought, ReAct, and ProgramOfThought. These modules enable safer composition and strong typing, allowing you to structure complex reasoning and generation workflows with built-in guardrails.

You can structure retrieval-augmented generation and agents by defining declarative LM pipelines that combine retrieval, reasoning, and generation. The framework automatically handles prompt optimization from your training data, removing the need for manual prompt engineering in multi-stage workflows.

Yes, you can use declarative programming to optimize prompts for RAG workflows. The framework supports combining retrieval, reasoning, and generation stages, while teleprompter optimizers automatically refine the underlying prompts using your provided training data.

Yes, you need training data to use automatic prompt optimizers like BootstrapFewShot, MIPRO, and COPRO. These optimizers analyze your provided examples to automatically improve and refine prompts, reducing manual effort in complex AI systems.

You should avoid using declarative LM pipelines if your project lacks sufficient training data for the teleprompter optimizers or if your tasks do not require structured composition, multi-step reasoning, or automated prompt refinement.