dspy

Automatic prompt optimization uses optimizers like BootstrapFewShot and MIPRO to generate better few-shot demonstrations and instructions from labeled training examples. This replaces manual trial-and-error by systematically improving pipeline quality from defined metrics and datasets.

Declarative LM programming defines input/output signatures and composes modules like ChainOfThought and ProgramOfThought to build AI agents. This approach assembles modular RAG pipelines and multi-stage reasoning workflows into maintainable systems without hardcoded prompts.

Build a multi-stage RAG pipeline by assembling retrieval, reasoning, and tool-using modules like ReAct into a composable system. This modular approach improves reliability for multi-hop RAG and question answering where measurable quality and consistent structure matter.

Signature programming supports information extraction and classification by defining declarative input/output constraints. You compose modules into pipelines and apply teleprompter optimization using defined evaluation metrics and validation datasets to ensure measurable quality.

DSPy supports different language model providers for multi-stage reasoning workflows. You define composable modules and signatures, then run teleprompter optimizers across providers to maintain consistent structure and measurable quality for your target task.

Automatic prompt optimization requires labeled training and validation datasets along with defined evaluation metrics to function. Without sufficient labeled examples for the teleprompter, the optimizer cannot generate effective few-shot demonstrations or improve pipeline reliability.