DSPy

• You are building a multi-stage AI pipeline (e.g. retrieve → reason → answer) and want maintainable, modular code instead of a wall of prompt strings.
• You have training examples and a metric, and want to systematically improve model outputs without manual prompt iteration.
• You need to swap LM providers (e.g. from GPT-4 to Claude) without rewriting your pipeline logic.
• You are building RAG systems or tool-using agents and want a structured, testable approach.

• You need a one-off, single-turn prompt with no optimization loop — plain API calls are simpler.
• You have no training data or metric and are purely prototyping; DSPy's value compounds with data.
• Your team is unfamiliar with Python and cannot run a pip-based install environment.

Define task inputs and outputs as typed Python classes. DSPy constructs and manages prompts from these signatures automatically, keeping your code readable and independent of prompt wording.

Predict, ChainOfThought, ReAct, and ProgramOfThought are drop-in building blocks. Combine them into custom dspy.Module subclasses to express multi-step reasoning pipelines cleanly.

Optimizers like BootstrapFewShot and MIPRO use your labeled examples and a metric function to iteratively improve prompts and few-shot demonstrations — replacing manual trial-and-error with a data-driven loop.

Configure OpenAI, Anthropic Claude, or local Ollama models with a single dspy.settings.configure(lm=...) call. You can even use different models for different stages of the same pipeline via context managers.

Built-in dspy.Retrieve integrates with vector stores like ChromaDB. ReAct enables tool-using agents. Both patterns compose naturally with optimizers for end-to-end improvement.

TypedPredictor accepts Pydantic models as output fields, giving you schema-validated, strongly-typed responses from any LM — no post-processing parsing required.

Wrap a ChainOfThought module in BootstrapFewShot, supply a small labeled dataset, and get a prompt-optimized QA system that measurably outperforms its untuned baseline.

Combine dspy.Retrieve with ChromaDB and a ChainOfThought answer generator. Optimize the full RAG pipeline end-to-end using training examples and an F1 or exact-match metric.

Use ReAct to build agents that call external tools (search, calculators, APIs) in a structured reasoning loop, with DSPy managing the think-act-observe cycle.

Define a Pydantic model for the fields you need (names, dates, amounts) and use TypedPredictor to extract validated structured data from unstructured text reliably.