Skip to content

Learning path · Programmatic Prompting · 33

MiPROv2

DSPy teleprompter using model-informed proposal and search to refine instructions and few-shot sets efficiently.

Why it matters

  • Often faster than brute-force prompt grids on structured pipelines.
  • Leverages LM feedback to propose edits.
  • Popular starting point for DSPy optimization workflows.

Key ideas

  • Guided proposal
  • Instruction refinement
  • Demonstration selection

MiPROv2 iteratively proposes instruction and demo changes informed by model behaviour on errors. It works well on classification, extraction, and multi-hop RAG modules with crisp metrics. Provide diverse failure examples in the train set so proposals target real edges. Compare against a strong manual baseline—optimization should beat human defaults, not merely beat garbage. Seed MiPROv2 with human-written baselines that already pass minimum bars; optimizers should refine strong starting points, not salvage broken pipelines. Ship only after eval gates pass on representative production failures.

Updated 2026-08-09 · Full learning path