Dspy

DSPy reframes LLM development as programming rather than prompting — you define a pipeline as composable Python modules (Predict, ChainOfThought, ReAct), specify a metric, and DSPy's compiler automatically optimizes the prompts and few-shot examples. This is particularly powerful for teams doing systematic evaluation-driven improvement rather than manual prompt tweaking. Steeper learning curve than simpler libraries but yields measurably better pipelines when properly used. Confidence is docs-derived.

Python SDK: pip install dspy. Define modules by subclassing dspy.Module with forward() methods. Signatures define input/output fields: class MyTask(dspy.Signature): question -> answer. Use dspy.ChainOfThought(MyTask) for CoT reasoning. Optimizers (BootstrapFewShot, MIPRO, etc.) tune the pipeline against a dev set. dspy.configure(lm=dspy.LM("openai/gpt-4o")) sets the LM. Compiled programs are serializable.

No DSPy API key — open-source library. LLM provider keys configured via dspy.configure(lm=...) or environment variables. Open-source access model (MIT license). No hosted auth complexity for library usage. DSPy Optimizer runs use the configured LLM provider API.

Optimization runs (compilers) make many LLM API calls — cost management matters. dspy.settings.inspect_history() for debugging. Module forward() failures surface as Python exceptions. Compiled programs are deterministic at inference time. Active GitHub (Stanford NLP) with regular releases. Research-grade quality with increasing production tooling.

dspy.ai covers quickstart, module definitions, signatures, optimizers, and advanced patterns. Getting started: pip install dspy, configure LM, write first module in under 15 minutes. Steeper conceptual onboarding than CrewAI/AutoGen but well worth it for quality-focused teams. GitHub: stanfordnlp/dspy. Active community via DSPy Discord.