dspy-advanced-module-composition

Files (2)

• SKILL.md
• example.py

SKILL.md (first lines)

# DSPy Advanced Module Composition

## Goal

Compose complex DSPy programs using the Ensemble optimizer, MultiChainComparison for reasoning synthesis, and sequential module patterns.

## When to Use

- Need consensus from multiple approaches
- Comparing different reasoning strategies
- Building robust pipelines with fallbacks
- Complex multi-step workflows with branching
- Ensemble methods for improved accuracy

## Related Skills

- Design modules: [dspy-custom-module-design](../dspy-custom-module-design/SKILL.md)
- Define signatures: [dspy-signature-designer](../dspy-signature-designer/SKILL.md)
- Evaluate performance: [dspy-evaluation-suite](../dspy-evaluation-suite/SKILL.md)

## Inputs

| Input | Type | Description |
|-------|------|-------------|
| `modules` | `list[dspy.Module]` | Modules to compose |
| `composition_type` | `str` | "ensemble", "sequential", "comparison" |

## Outputs

| Output | Type | Description |
|--------|------|-------------|
| `composed_program` | `dspy.Module` | Composed multi-module program |

## Workflow

### Phase 1: Ensemble Voting

Combine multiple programs using the Ensemble optimizer:

```python
import dspy
from dspy.teleprompt import Ensemble

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

# Define a signature for the task
class BasicQA(dspy.Signature):
    """Answer questions with short factoid answers."""
    question = dspy.InputField()
    answer = dspy.OutputField()

# Create multiple program instances (should be optimized/compiled programs)
# For simple demonstration, we'll use different predictors
program1 = dspy.Predict(BasicQA)
program2 = dspy.ChainOfThought(BasicQA)
program3 = dspy.Predict(BasicQA)

# Ensemble is an optimizer that compiles programs together
ensemble = Ensemble(reduce_fn=dspy.majority)
ensembled_program = ensemble.compile([program1, program2, program3])

# Use the ensembled program
result = ensembled_program(question="What is 2 + 2?")
print(result.answer)  # Voted answer
```

### Phase 2: MultiChainComparison

Compare multiple reasoning attempts:

```python
import dspy

class BasicQA(dspy.Signature):
    """Answer questions with short factoid answers."""
    question = dspy.InputField()
    answer = dspy.OutputField(desc="often between 1 and 5 words")

class ComparisonPipeline(dspy.Module):
    def __init__(self):
        # Generate multiple reasoning attempts
        self.cot = dspy.ChainOfThought(BasicQA)

        # Compare M attempts and select best
        # Must pass a Signature class, not a string
        self.compare = dspy.MultiChainComparison(
            BasicQA,
            M=3,  # Number of attempts to compare
            temperature=0.7
        )

    def forward(self, question):
        # Generate multiple completions to compare
        # Each completion must have rationale/reasoning field
        completions = [
            self.cot(question=question)
            for _ in range(3)
        ]

        # MultiChainComparison synthesizes them into best answer
        # Pass completions as positional arg, not keyword arg
        return self.compare(completions, question=question)

# Usage
dspy.configure(lm=dspy.LM("openai/gpt-4o-mini"))
pipeline = ComparisonPipeline()
result = pipeline(question="Explain quantum computing")
print(f"Best answer: {result.answer}")
print(f"Rationale: {result.rationale}")
```

### Phase 3: Sequential Composition

Chain modules for multi-step workflows:

```python
import dspy

# Define signatures for each step

From OmidZamani/dspy-skills, MIT.

More from OmidZamani/dspy-skills