dspy

Declarative language-model programming assembles modular AI pipelines by defining typed Signatures and structured outputs, enabling you to build multi-stage workflows without manual prompt engineering. This approach ensures reliable composition for research and production use.

Automatic prompt optimization uses data-driven teleprompter optimizers to improve prompts for RAG systems and agents. By treating prompts as optimizable parameters within modular pipelines, the framework systematically enhances output reliability without manual tuning.

Declarative LM programming works with OpenAI and Anthropic models by routing typed Signature definitions and structured outputs through modular components like Predict and ChainOfThought. This abstracts model-specific prompt formatting into a unified programming interface.

Yes, you can use modular components like ReAct, ChainOfThought, and ProgramOfThought to construct multi-stage RAG systems and agents. These modules compose together declaratively to handle complex reasoning and classification tasks across your pipeline.

Typed Signatures enforce strong type safety and structured outputs in language-model workflows, defining exact input and output fields for each module. This prevents unpredictable model responses and enables automatic prompt optimization across composed pipelines.

Automatic prompt optimization may not suit AI pipelines with highly dynamic or undocumented tasks where defining stable typed Signatures is impossible. If your workflow requires zero-latency execution without optimization overhead, manual prompt engineering offers a lighter alternative.