DSPy — Skill Router

Program — don't prompt — foundation models. DSPy lets you define LLM pipelines as typed Python modules with declarative Signatures, then compile (optimize) them against a metric using optimizers like BootstrapFewShot and MIPROv2.

Source: dspy.ai | Python: v2.5.x | License: MIT | GitHub: stanfordnlp/dspy

Reference Files

Reference	File	Read When
Overview & Quickstart	`references/00-overview.md`	First install, core concepts, `dspy.configure`, the compile loop
Signatures	`references/01-signatures.md`	Inline `"q -> a"` strings, class-based signatures, `InputField`/`OutputField`, typed fields
Modules (Predict, CoT, ReAct, PoT)	`references/02-modules.md`	`dspy.Predict`, `ChainOfThought`, `ReAct`, `ProgramOfThought`, composing `dspy.Module`
LM Configuration	`references/03-lm-configuration.md`	`dspy.LM`, providers (OpenAI, Anthropic, Ollama, vLLM), caching, temperature, context settings
Optimizers (Compilers)	`references/04-optimizers.md`	`BootstrapFewShot`, `BootstrapFewShotWithRandomSearch`, `MIPROv2`, `COPRO`, `BootstrapFinetune`
Metrics & Evaluation	`references/05-metrics-evaluation.md`	Writing metric functions, `dspy.Evaluate`, answer-exact-match, LLM-as-judge metrics
Retrieval & RAG	`references/06-rag-retrieval.md`	`dspy.Retrieve`, ColBERTv2, Qdrant/Chroma/Weaviate integrations, multi-hop RAG patterns
Assertions & Suggestions	`references/07-assertions.md`	`dspy.Assert`, `dspy.Suggest`, self-refinement loops, constraint-driven retries
Deployment & Production	`references/08-deployment.md`	Saving/loading compiled programs, streaming, async, FastAPI serving, observability