DSPy From First Principles: Programs, Metrics, Optimizers, Agents, and Reasoning Search

What if building applications with language models stopped being an exercise in hand-writing prompts and became an engineering discipline?

DSPy From First Principles develops that idea from the ground up.

Starting with the simplest language-model call, the book gradually turns prompts into typed programs, examples into experimental data, metrics into objectives, and optimization into a controlled engineering process.

Along the way, you will build and examine:

DSPy Signatures and Modules

structured language-model programs

few-shot optimization

instruction optimization

reflective optimization with GEPA

metrics and adversarial metric testing

agents and tool use

retrieval, memory, and long-context systems

reasoning search and Monte Carlo Tree Search

leakage controls and evidence boundaries

evaluation, promotion, and production deployment

a complete self-improving engineering program

The emphasis throughout is not simply on making a model produce better answers. It is on understanding why a system improved, what evidence supports that conclusion, and what information an optimizer or agent was allowed to see.

The experiments also include failures and null results. A metric that appears reasonable can reward the wrong behaviour. An optimizer can exploit a weak objective. A tree-search implementation can draw a tree without genuinely exploring one. A stronger model can make expensive reasoning search unnecessary.

By the end, DSPy becomes more than a prompt framework. It becomes a way to treat language-model behaviour as an experimental variable inside a larger software system.

This book is for developers, AI engineers, researchers, and technically curious readers who want to understand DSPy from first principles and build language-model systems that can be measured, tested, optimized, and trusted.