Pipeline Multi Hop

Pipeline Multi-Hop QA

🎯 Trigger Conditions

Use when asked about multi-hop question answering, complex reasoning chains, or iterative retrieval-augmented generation.

📚 Prerequisites

• dspy package installed
• Retriever configured
• Multi-hop dataset available

🛠️ Multi-Hop Patterns

1. Basic Multi-Hop

class MultiHopQA(dspy.Module):
    def __init__(self, n_hops=2):
        self.n_hops = n_hops
        self.retrieve = dspy.Retrieve(k=3)
        self.answer = dspy.Predict("context, question -> answer")
    
    def forward(self, question):
        context = []
        current_question = question
        
        for _ in range(self.n_hops):
            # Retrieve relevant passages
            passages = self.retrieve(current_question).passages
            context.extend(passages)
            
            # Generate next question if needed
            if _ < self.n_hops - 1:
                next_q = dspy.Predict("context, question -> next_question")(
                    context=context, question=current_question
                ).next_question
                current_question = next_q
        
        return self.answer(context=context, question=question)

2. Adaptive Multi-Hop

class AdaptiveMultiHopQA(dspy.Module):
    def __init__(self, max_hops=3):
        self.max_hops = max_hops
        self.retrieve = dspy.Retrieve(k=3)
        self.answer = dspy.Predict("context, question -> answer")
        self.decide = dspy.Predict("context, question -> continue_search")
    
    def forward(self, question):
        context = []
        for hop in range(self.max_hops):
            # Retrieve passages
            passages = self.retrieve(question).passages
            context.extend(passages)
            
            # Decide if more search is needed
            should_continue = self.decide(
                context=context, question=question
            ).continue_search
            
            if not should_continue:
                break
        
        return self.answer(context=context, question=question)

3. Multi-Hop with Evidence

class MultiHopWithEvidence(dspy.Module):
    def __init__(self):
        self.retrieve = dspy.Retrieve(k=3)
        self.answer = dspy.Predict("context, question -> answer, evidence")
    
    def forward(self, question):
        context = []
        evidence = []
        
        # First hop
        passages1 = self.retrieve(question).passages
        context.extend(passages1)
        
        # Second hop with refined question
        passages2 = self.retrieve(question + " from " + passages1[0]).passages
        context.extend(passages2)
        evidence.extend(passages1)
        
        return self.answer(context=context, question=question)

⚠️ Pitfalls

• Error propagation: Errors compound across hops
• Context length: Manage context window carefully
• Latency: Multi-hop increases response time
• Quality: Later hops may degrade in quality

📖 References

• Multi-Hop QA Paper
• DSPy Multi-Hop