DSPy 3.4 introduces Jev integration via the TypeSafe client, adding three experimental decision types—Noul, Choice, and Score—that provide probability evidence within DSPy signatures. This release also brings native LM engines using bundled lm15 types, a local CPython interpreter for trusted code, async ReActV2, and versioned documentation.

  • Jev integration allows users to access Noul (boolean), Choice (selection), and Score (rubric) decision types with associated probability and confidence metrics through the standard lm interface.
  • The new ReAnchor optimizer calibrates these decisions against the program's metric by fitting thresholds and weights without rewriting instructions or demonstrations.
  • Native LM engines use bundled lm15 request and response types, with engine="auto" preferring native execution and falling back to LiteLLM before inference.
  • Custom HTTP providers can be registered via dspy.lm15.register_provider to declare endpoints and capabilities for OpenAI- or Anthropic-compatible services.
  • DSPy 3.4 is designated as the LM transition release, with 3.5 serving as the migration deadline for legacy custom-LM integrations.

The update marks a significant shift in how DSPy handles language model interactions by standardizing decision outputs and native engine support.