DSPy tutorial: local models and GEPA prompt optimization

This English summary is based on the complete English caption track of Fraunhofer IEM’s DSPy workshop, whose title and description are in German.

Tommy Falkowski introduces DSPy through Python examples that replace manually written prompts with signatures describing inputs and outputs. The walkthrough uses a local LLM configured with an API URL and model name. It starts with a basic request, then builds joke generators with topic, language and boolean inputs before defining signatures as Python classes.

The tutorial inspects the prompts DSPy generates behind the scenes and explains response caching, typed fields and validation. Falkowski describes how validation can trigger another request when an answer fails to match the expected format. He also demonstrates Chain of Thought output and a ReAct module that calls a mock population lookup and a calculation function to answer a question.

The later examples focus on GEPA prompt optimization using labeled data, with results prepared before recording. These examples optimize instructions rather than model weights. A support-ticket classifier uses optimized prompts for an urgency scheme where C is most urgent and A is least urgent. A headline-rating example uses the speaker's personal annotations on a scale he corrects to 0 through 5. He reports accuracy rising from 56% to 83.3% in that example; these figures describe his demonstration rather than general performance guarantees.

Falkowski argues that DSPy can help when developers run models locally, particularly with smaller models or applications that need traceable behavior. Optimization generates multiple prompt candidates, takes time and consumes tokens. He explains that developers can repeat it after changing models, and briefly discusses schema-based extraction.