dspy

Declarative LM programming structures language model systems into modular components that define specific inputs, outputs, and behaviors. It enables scalable RAG systems and ReAct-style agent workflows by composing reusable modules rather than writing hardcoded prompts.

You build a RAG system by defining declarative modules for retrieval and generation, then chaining them into agent workflows. An optimizer like BootstrapFewShot runs on a training set to automatically generate demonstrations and refine the pipeline's logic.

Yes, declarative LM programming supports building multi-stage reasoning pipelines across diverse LM providers. You define modules and metrics independently, allowing the teleprompter workflow to optimize prompts regardless of the underlying language model provider.

The best way to generate few-shot examples is using a metric-driven teleprompter workflow like BootstrapFewShot. It evaluates your declarative modules against a training set to automatically extract effective demonstrations and instruction tweaks for deployment.

Avoid automated prompt optimization if you lack a defined metric to evaluate multi-stage reasoning tasks, or if your training set is too small. The teleprompter requires metric-driven feedback to generate effective demonstrations and fine-tuning data.