Agent Skills: DSPy MIPROv2 Optimizer

DSPy MIPROv2 Optimizer

Goal

Jointly optimize instructions and few-shot demonstrations using Bayesian Optimization for maximum performance.

When to Use

• You have 200+ training examples
• You can afford longer optimization runs (40+ trials)
• You need state-of-the-art performance
• Both instructions and demos need tuning

Inputs

| Input | Type | Description | |-------|------|-------------| | program | dspy.Module | Program to optimize | | trainset | list[dspy.Example] | 200+ training examples | | metric | callable | Evaluation function | | auto | str | "light", "medium", or "heavy" | | num_trials | int | Optimization trials (40+) |

Outputs

| Output | Type | Description | |--------|------|-------------| | compiled_program | dspy.Module | Fully optimized program |

Workflow

Three-Stage Process

1. Bootstrap - Generate candidate demonstrations
2. Propose - Create grounded instruction candidates
3. Search - Bayesian optimization over combinations

Phase 1: Setup

import dspy

dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))

Phase 2: Define Program

class RAGAgent(dspy.Module):
    def __init__(self):
        self.retrieve = dspy.Retrieve(k=3)
        self.generate = dspy.ChainOfThought("context, question -> answer")
    
    def forward(self, question):
        context = self.retrieve(question).passages
        return self.generate(context=context, question=question)

Phase 3: Optimize

optimizer = dspy.MIPROv2(
    metric=dspy.evaluate.answer_exact_match,
    auto="medium",  # Balanced optimization
    num_threads=24
)

compiled = optimizer.compile(RAGAgent(), trainset=trainset)

Auto Presets

| Preset | Trials | Use Case | |--------|--------|----------| | "light" | ~10 | Quick iteration | | "medium" | ~40 | Production optimization | | "heavy" | ~100+ | Maximum performance |

Production Example

import dspy
from dspy.evaluate import Evaluate
import json
import logging

logger = logging.getLogger(__name__)

class ReActAgent(dspy.Module):
    def __init__(self, tools):
        self.react = dspy.ReAct("question -> answer", tools=tools)
    
    def forward(self, question):
        return self.react(question=question)

def search_tool(query: str) -> list[str]:
    """Search knowledge base."""
    results = dspy.ColBERTv2(url='http://20.102.90.50:2017/wiki17_abstracts')(query, k=3)
    return [r['text'] for r in results]

def optimize_agent(trainset, devset):
    """Full MIPROv2 optimization pipeline."""
    
    agent = ReActAgent(tools=[search_tool])
    
    # Baseline evaluation
    evaluator = Evaluate(
        devset=devset,
        metric=dspy.evaluate.answer_exact_match,
        num_threads=8
    )
    baseline = evaluator(agent)
    logger.info(f"Baseline: {baseline:.2%}")
    
    # MIPROv2 optimization
    optimizer = dspy.MIPROv2(
        metric=dspy.evaluate.answer_exact_match,
        auto="medium",
        num_threads=24,
        # Custom settings
        num_candidates=15,
        max_bootstrapped_demos=4,
        max_labeled_demos=8
    )
    
    compiled = optimizer.compile(agent, trainset=trainset)
    optimized = evaluator(compiled)
    logger.info(f"Optimized: {optimized:.2%}")
    
    # Save with metadata
    compiled.save("agent_mipro.json")
    
    metadata = {
        "baseline_score": baseline,
        "optimized_score": optimized,
        "improvement": optimized - baseline,
        "num_train": len(trainset),
        "num_dev": len(devset)
    }
    
    with open("optimization_metadata.json", "w") as f:
        json.dump(metadata, f, indent=2)
    
    return compiled, metadata

Instruction-Only Mode

# Disable demos for pure instruction optimization
optimizer = dspy.MIPROv2(
    metric=metric,
    auto="medium",
    max_bootstrapped_demos=0,
    max_labeled_demos=0
)

Best Practices

1. Data quantity matters - 200+ examples for best results
2. Use auto presets - Start with "medium", adjust based on results
3. Parallel threads - Use num_threads=24 or higher if available
4. Monitor costs - Track API usage during optimization
5. Save intermediate - Bayesian search saves progress

Limitations

• High computational cost (many LLM calls)
• Requires substantial training data
• Optimization time: hours for "heavy" preset
• Memory intensive for large candidate sets