dspy

Declarative LM programming replaces manual prompt writing by defining tasks as reusable signatures and modules. You structure multi-stage AI workflows into composable components, enabling easier maintenance and optimization across RAG pipelines, agents, and classifiers without hand-crafting prompts.

Teleprompting optimizes language model pipelines using automated optimizers like BootstrapFewShot, MIPRO, and COPRO. It applies metric-driven evaluation and few-shot demonstration generation to tune module behavior, improving reasoning and output quality without manual prompt adjustments.

Yes, you can build RAG systems and multi-step agents using structured signatures. Signatures define retrieval and reasoning steps declaratively, allowing you to compose multi-stage AI workflows with reusable components for assistants, extraction, and summarization tasks.

Yes, metric-driven evaluation is required to optimize a declarative language model program. You must supply an evaluation metric and available examples so teleprompters can automatically generate few-shot demonstrations and instruction optimizations for your modules.

The best way to structure multi-stage AI workflows for summarization and extraction is using declarative module composition. Define each stage as a signature, chain the modules together, and apply instruction or few-shot optimization to refine the structured outputs.

You should avoid declarative programming for language model pipelines when you lack defined evaluation metrics or available examples for optimization. Without metric-driven evaluation, teleprompters cannot automatically tune modules, making the declarative approach less effective than direct prompting.