How do DSPy "Teleprompters" work for few-shot learning?

All answers to this question.

Teleprompters are just the names for optimizers in DSPy. BootstrapFewShot is one of the most popular because it uses your existing program (the "teacher") to generate predictions for your 10 examples. If those predictions pass your validation metric, they are saved as "demonstrations" and automatically injected into the prompt. It’s essentially automated few-shot prompting. You don't have to manually pick the best examples; DSPy finds which ones actually help the model perform better based on the math, not just your intuition.

   Answered 2025-07-05 by Heather Miller

Can we use DSPy to optimize for cost and latency, or is it just for accuracy?

   Answered 2025-07-15 by Brandon Taylor

• You can optimize for whatever you can measure, Brandon. If your metric function penalizes long responses or high token counts, DSPy will try to find the shortest prompt that still meets your accuracy threshold. This is one of the most powerful features for production deployments where API costs are a major concern.

     Commented 2025-07-22 by Justin Fisher

The best part is that the DSPy teleprompter is model-agnostic. It works the same whether you're using OpenAI, Anthropic, or a local Llama.

   Answered 2025-08-10 by Ryan Garcia

• Exactly, Ryan. This portability is why everyone is moving away from hard-coded prompt templates in favor of the DSPy approach.

     Commented 2025-08-12 by Katherine Rivera