dspy

Automated prompt optimization uses optimizer-driven prompt search to iteratively improve language model programs. Teleprompters like BootstrapFewShot and MIPRO automatically generate and validate few-shot examples to enhance multi-stage reasoning performance.

Declarative language model programming defines inputs and outputs using typed signatures, abstracting prompt engineering details. You need it for modular pipelines, retrieval-augmented generation, and multi-stage reasoning workflows to ensure maintainable, repeatable performance.

Building multi-stage AI agents for retrieval-augmented generation involves composing declarative modules for document retrieval, reasoning, and classification. This approach satisfies typed signatures and integrates deterministically with external calculation tools.

Automated prompt optimization works directly with existing evaluation metrics by applying optimizer-driven prompt search. Teleprompters evaluate candidate prompts against your defined metrics to iteratively bootstrap effective few-shot examples.

The best way to structure classification and structured prediction tasks is using typed signatures to define expected inputs and outputs. This declarative approach allows optimizers to handle prompting details and evaluate accuracy automatically.

Hand-tuned prompt engineering fails to scale for complex reasoning pipelines because manual adjustments are difficult to maintain and port across models. Declarative LM programming solves this by treating tasks as reusable, testable programs with automated optimization.