dspy

Declarative LM programming builds self-improving pipelines by defining tasks as signatures with field-based inputs and outputs, then composing modular components like ChainOfThought or ReAct into cohesive workflows. This replaces ad-hoc prompting with structured, optimizable modules.

You can automatically optimize RAG and agent prompts using teleprompter optimizers like BootstrapFewShot, MIPRO, and BootstrapFinetune. These optimizers leverage labeled training data to automatically generate and refine few-shot examples or fine-tune models for your pipelines.

Yes, using prompt optimizers like BootstrapFewShot and MIPRO assumes you have labeled training data available. Providing a trainset allows the optimizers to automatically compile and improve your declarative LM modules for tasks such as classification and retrieval-augmented generation.

The best way to build modular RAG pipelines is by composing declarative building blocks such as Predict and ChainOfThought. This modular composition allows you to create complex data-to-decision workflows that can be systematically improved using teleprompter optimizers.

Yes, you can fine-tune language models using the BootstrapFinetune optimizer. Alongside prompt optimization techniques like BootstrapFewShot and MIPRO, BootstrapFinetune enables you to automatically improve and adapt models for RAG, agents, and classifiers.

You should avoid declarative self-improving prompts if you lack access to LM providers or labeled training data. This approach relies on available training data to run teleprompter optimizers, making it unsuitable for zero-shot workflows without evaluation sets.