dspy

To automatically optimize prompts for RAG pipelines, define declarative signatures and modules, then use teleprompters like BootstrapFewShot with a user-defined correctness metric. This replaces manual prompt tweaking by systematically compiling and tuning language-model programs using evaluation-driven feedback loops.

Declarative programming for language-model programs defines inputs and outputs for tasks like classification and extraction without hand-written prompt chains. Use it for dependable research workflows, multi-stage reasoning, or agentic systems that require systematic improvement rather than manual prompt adjustments.

Build multi-stage agents with structured outputs by configuring an OpenAI or Anthropic LM provider and composing modular pipelines. Define task signatures, implement retrieval-augmented modules with ChainOfThought, and apply teleprompters to compile and optimize the agentic system using metric-based feedback.

Yes, you can build ReAct agents and multi-hop reasoning systems using composable modules. Define declarative signatures for each reasoning stage, configure your LM provider, and apply automatic prompt optimization with teleprompters to systematically improve agent performance using evaluation metrics.

DSPy supports evaluation-driven prompt tuning for classification and extraction by applying teleprompters to optimize declarative signatures. You define a metric-based feedback loop, then compile the language-model program using BootstrapFewShot or MIPRO to automatically improve structured generation accuracy.

Automatic prompt optimization for complex AI pipelines requires defining explicit signatures, configuring LM providers, and providing a user-defined evaluation metric. It depends on metric-based feedback loops and teleprompters, meaning poor metrics or insufficient examples will yield suboptimal compilation results for multi-stage reasoning systems.