dspy-ruby_skill

Overview

This skill packages DSPy.rb capabilities for building type-safe, modular LLM applications in Ruby. It exposes signatures, modules, agents, tools, optimization, and evaluation primitives so you can implement predictable AI features with Sorbet-backed types and production-ready observability. Use it to replace brittle prompt tinkering with typed contracts and repeatable optimization.

How this skill works

The skill maps Ruby types (T::Struct, T::Enum) to LLM schemas and enforces typed inputs and outputs via DSPy signatures. It provides modular building blocks — Predict, ReAct, ChainOfThought, and tool/agent tooling — plus optimization engines (MIPROv2, GEPA) and an evaluation framework for systematic improvement. Configuration layers let you select provider adapters or a unified RubyLLM router, and fiber-local context and lifecycle hooks control runtime behavior.

When to use it

• Building reliable LLM endpoints that must return structured, type-safe data
• Composing agent workflows that call typed tools and need predictable tool arguments
• Optimizing prompts with data-driven search or genetic evolution rather than manual trial
• Integrating LLM logic into Rails with clear directory patterns, hooks, and observability
• Evaluating model performance with repeatable metrics and test datasets

Best practices

• Define clear DSPy::Signature types instead of relying on free-form text outputs
• Keep context as T::Struct types (Typed Context Pattern) so each field is explicit and testable
• Use provider adapters or RubyLLM routing to manage multi-provider deployments
• Instrument modules with DSPy events and lifecycle callbacks for observability and metrics
• Run optimization (MIPROv2/GEPA) against real train/val splits and export checkpoints for reproducibility

Example use cases

• Customer support triage: a typed EmailClassifier signature that returns category, priority, and confidence
• Data-processing agent: a Toolset that converts and validates formats with typed enums
• Feature-flagged model selection: fast model for selectors and powerful model for synthesis in Rails
• Prompt optimization: run GEPA to evolve instructions with reflection LMs and feedback maps
• Automated evaluation: batch-run DSPy::Evals with exact_match or composite metrics and export scores