InstructionProposer

Generates instruction candidates for each optimizable module.

MIPRO Step 2: uses a meta-LLM to propose diverse instruction variants by providing context about the dataset, program structure, and few-shot demos.

Works with both strategy-level (system prompt) and per-subgraph modules.

Types

suspend fun proposeInstructionsForProgram(demoCandidates: Map<String, List<List<Demonstration>>>?, numCandidates: Int, previousInstructions: Map<String, List<Pair<String, Double>>> = emptyMap(), parallelism: Int = 1): Map<String, List<String>>

Generates instruction candidates for all modules.