multi-hop-rag

name	multi-hop-rag
description	Use when user needs multi-step reasoning with iterative retrieval for complex questions. Triggers on: multi-hop, multi-step RAG, complex questions, chain of retrieval, iterative retrieval, complex reasoning, cross-document reasoning, question decomposition, research questions, fact synthesis, connecting information across documents.

Multi-Hop RAG

Answer complex questions that require connecting information from multiple documents through iterative retrieval — each retrieval step informs the next.

When to Activate

This skill should be activated when the user:

• Asks questions requiring information from multiple sources
• Needs to connect facts across documents ("A founded B, B acquired C, what's A's connection to C?")
• Has research questions requiring iterative discovery
• Asks "compare", "trace", "investigate", or "how did X lead to Y" questions
• Needs cross-document fact checking or verification

Interactive Flow

Multi-hop adds complexity and latency. Validate the need and choose the right strategy.

Step 1: Validate Multi-Hop Need

"Multi-hop retrieval is for questions that need information from multiple places.

Can you give me an example question users will ask?
___

I'll analyze if it truly needs multi-hop or if single retrieval suffices."

Analysis criteria:

Question Pattern	Needs Multi-Hop?
"What is X?"	No - single retrieval
"Compare X and Y"	Yes - need both X and Y
"How did X lead to Y?"	Yes - need chain
"What are all the..."	Maybe - depends on distribution

"Your example '[question]' [does/doesn't] need multi-hop because [reason].

Proceed with multi-hop? (yes / try basic RAG first)"

Step 2: Choose Decomposition Strategy

"How should I break down complex questions?

A) **Upfront decomposition** - Split into sub-questions first, retrieve each
   Best for: Questions with clear parts ("Compare A and B")

B) **Iterative discovery** - Retrieve, evaluate, follow leads
   Best for: Questions where you don't know what you need ("Investigate X")

C) **Hybrid** - Decompose first, then iterate if needed
   Best for: Complex research questions

Which strategy? (A/B/C)"

Strategy	Example Flow
A (Upfront)	Q: "Compare pricing of A and B" → Sub-Q1: "A pricing" → Sub-Q2: "B pricing" → Combine
B (Iterative)	Q: "Who owns company X?" → Find X → X owned by Y → Find Y → Y owned by Z → Answer chain
C (Hybrid)	Decompose into parts, then iterate within each part

Step 3: Set Hop Limits

"How deep should the search go?

A) 2 hops (fast, simple chains)
B) 3 hops (balanced, most cases)
C) 4-5 hops (deep research, complex chains)

More hops = more complete but slower. Recommend B.

Which? (A/B/C)"

Step 4: Sufficiency Criteria

"When should I stop searching and answer?

A) **Strict** - Only stop when ALL parts are fully answered
   Use for: Legal, compliance, fact-checking

B) **Lenient** - Stop when main information is found
   Use for: General research, exploratory questions

Which? (A/B)"

Step 5: Confirm Configuration

"Multi-hop configuration:

- **Strategy**: [from Step 2]
- **Max hops**: [from Step 3]
- **Sufficiency**: [from Step 4]
- **Per-hop retrieval**: Top-5

Example flow for your question:
  Hop 1: [query] → find [what]
  Hop 2: [follow-up] → find [what]
  → Synthesize answer

Does this approach make sense? (yes / adjust [what])"

Checkpoints During Execution

Checkpoint	Question to User
After decomposition	"I've split into these sub-questions: [list]. Look right?"
After each hop	"Hop [N] found: [summary]. Continue searching or enough info?"
Before synthesis	"I have info from [N] sources. Ready to synthesize?"
If stuck	"Hop [N] found nothing relevant. Try different terms or answer with what we have?"

Red Flags - When to Simplify

"I notice:
- Your question only needs info from one source
- OR: Each hop returns the same documents
- OR: The 'chain' is actually just one lookup

Basic RAG would be faster and simpler. Switch? (yes / keep multi-hop)"

Strategy Selection Helper

If user is unsure, use this decision tree:

"Let me help choose the right approach:

Is the answer in ONE document?
  → Yes: Use basic `rag`
  → No: Continue...

Do you know what pieces you need upfront?
  → Yes: Upfront decomposition (Strategy A)
  → No: Continue...

Is it a chain (A→B→C) or parallel (A, B, C)?
  → Chain: Iterative discovery (Strategy B)
  → Parallel: Upfront decomposition (Strategy A)
  → Both: Hybrid (Strategy C)

Based on this: I recommend Strategy [X]. Agree?"

Core Concepts

The Multi-Hop Problem

Some questions can't be answered with a single retrieval:

Question: "Who is the CEO of the company that acquired our main competitor?"

Single-hop fails:
  Search: "CEO company acquired competitor" → No direct match

Multi-hop succeeds:
  Hop 1: "main competitor" → "TechCorp is our main competitor"
  Hop 2: "TechCorp acquisition" → "TechCorp was acquired by MegaInc in 2023"
  Hop 3: "MegaInc CEO" → "Jane Smith is CEO of MegaInc"
  Answer: "Jane Smith"

Retrieval Chain

┌─────────────────────────────────────────────────────────────────┐
│                    Multi-Hop Retrieval                          │
├─────────────────────────────────────────────────────────────────┤
│                                                                 │
│  Question ──▶ Decompose ──▶ [Sub-Q 1] ──▶ [Sub-Q 2] ──▶ ...    │
│                                 │             │                 │
│                                 ▼             ▼                 │
│                            Retrieve 1    Retrieve 2             │
│                                 │             │                 │
│                                 ▼             ▼                 │
│                            Facts 1   ───▶ Facts 2               │
│                            (inform next query)                  │
│                                                                 │
│                                     │                           │
│                                     ▼                           │
│                               Synthesize                        │
│                                     │                           │
│                                     ▼                           │
│                                 Answer                          │
│                                                                 │
└─────────────────────────────────────────────────────────────────┘

Mental Model: Detective Investigation

Think of multi-hop RAG like a detective following leads:

• Clue 1 leads to Person A
• Person A mentions Location B
• Location B reveals Evidence C
• Evidence C solves the case

Each finding opens new avenues of investigation.

Why Multi-Hop Over Alternatives

Approach	Single Query	Multi-Step	Information Linking	Use Case
Basic RAG	Yes	No	No	Simple factual Q&A
Agentic RAG	Yes	Yes (dynamic)	Limited	Conversational
Multi-Hop RAG	No	Yes (structured)	Yes	Complex research

Choose Multi-Hop when:

• Answer requires connecting 2+ pieces of information
• Question involves relationships, comparisons, or chains
• Single retrieval returns partial information
• Research tasks need systematic exploration

Skip Multi-Hop when:

• Question is simple and factual
• All information is in one document
• Latency is critical (each hop adds ~200ms)

Implementation

from pymilvus import MilvusClient, DataType
from openai import OpenAI
from langchain.text_splitter import RecursiveCharacterTextSplitter

class MultiHopRAG:
    def __init__(self, uri: str = "./milvus.db"):
        self.client = MilvusClient(uri=uri)
        self.openai = OpenAI()
        self.splitter = RecursiveCharacterTextSplitter(chunk_size=512, chunk_overlap=50)
        self.collection_name = "multi_hop_rag"
        self._init_collection()

    def _embed(self, texts):
        if isinstance(texts, str):
            texts = [texts]
        response = self.openai.embeddings.create(model="text-embedding-3-small", input=texts)
        return [item.embedding for item in response.data]

    def _init_collection(self):
        if self.client.has_collection(self.collection_name):
            return
        schema = self.client.create_schema()
        schema.add_field("id", DataType.INT64, is_primary=True, auto_id=True)
        schema.add_field(, DataType.VARCHAR, max_length=)
        schema.add_field(, DataType.VARCHAR, max_length=)
        schema.add_field(, DataType.FLOAT_VECTOR, dim=)

        index_params = .client.prepare_index_params()
        index_params.add_index(, index_type=, metric_type=)
        .client.create_collection(.collection_name, schema=schema, index_params=index_params)

     ():
        chunks = .splitter.split_text(text)
        embeddings = ._embed(chunks)
        data = [{: c, : source, : e}  c, e  (chunks, embeddings)]
        .client.insert(.collection_name, data)

     () -> []:
        embedding = ._embed(query)[]
        results = .client.search(.collection_name, [embedding], limit=top_k,
                                     output_fields=[, ])
         [{: h[][], : h[][]}  h  results[]]

     () -> []:
        
        response = .openai.chat.completions.create(
            model=,
            messages=[{
                : ,
                : 
            }],
            temperature=
        )
        subs = response.choices[].message.content.strip().split()
         [s.strip()  s  subs  s.strip()]

     () -> :
        
        all_contexts = []
        queries = [question]
        hop_details = []

         hop  (max_hops):
            current_query = queries[-]
            results = .retrieve(current_query, top_k=)
            all_contexts.extend(results)

            hop_details.append({
                : hop + ,
                : current_query,
                : (results)
            })

            
            context_text = .join([c[]  c  all_contexts])
            check = .openai.chat.completions.create(
                model=,
                messages=[{
                    : ,
                    : 
                }],
                temperature=
            )

            response_text = check.choices[].message.content
               response_text.upper():
                

            
               response_text:
                missing = response_text.split()[-].strip()
                followup = .openai.chat.completions.create(
                    model=,
                    messages=[{
                        : ,
                        : 
                    }],
                    temperature=
                )
                queries.append(followup.choices[].message.content.strip())

         {
            : all_contexts,
            : queries,
            : hop_details
        }

     () -> :
        
        
        retrieval = .multi_hop_retrieve(question, max_hops)

        
        seen = ()
        unique_contexts = []
         c  retrieval[]:
             c[]   seen:
                seen.add(c[])
                unique_contexts.append(c)

        
        context_text = .join([  c  unique_contexts])
        response = .openai.chat.completions.create(
            model=,
            messages=[{
                : ,
                : 
            }],
            temperature=
        )

         {
            : response.choices[].message.content,
            : (retrieval[]),
            : retrieval[],
            : ((c[]  c  unique_contexts)),
            : retrieval[]
        }

Usage:

rag = MultiHopRAG()
rag.add_document(open("people.md").read(), source="people.md")
rag.add_document(open("companies.md").read(), source="companies.md")
rag.add_document(open("deals.md").read(), source="deals.md")

result = rag.query("Who founded the company that acquired TechCorp?")
print(f"Answer: {result['answer']}")
print(f"Hops taken: {result['hops']}")
print(f"Queries: {result['queries']}")

Configuration Guide

Max Hops

Complexity	max_hops	Example Question
Simple chain	2	"Who is X's manager?"
Medium	3	"What company did X's former employer acquire?"
Complex research	4-5	"Trace the ownership history of product Y"

Rule: Start with 3, increase only if answers are incomplete.

Hop Strategy

Strategy	When to Use	Implementation
Question Decomposition	Clear sub-questions	Split upfront, retrieve each
Iterative Discovery	Unknown structure	Retrieve → evaluate → follow leads
Hybrid	Complex research	Decompose first, then iterate

Sufficiency Checking

# Strict (for high-precision needs)
STRICT_CHECK = """Answer SUFFICIENT only if ALL parts of the question can be fully answered.
Any uncertainty = NEED: <what's uncertain>"""

# Lenient (for exploratory research)
LENIENT_CHECK = """Answer SUFFICIENT if we have the main information.
Only say NEED if critical information is completely missing."""

Common Pitfalls

1. Too Many Hops

Symptom: Slow responses, redundant information Fix: Lower max_hops, improve sufficiency check

2. Circular Retrieval

Symptom: Same documents retrieved repeatedly Fix: Track seen documents, exclude from subsequent searches

def retrieve_excluding(self, query: str, exclude_ids: set, top_k: int = 5):
    results = self.retrieve(query, top_k=top_k * 2)
    return [r for r in results if r["id"] not in exclude_ids][:top_k]

3. Query Drift

Symptom: Follow-up queries stray from original question Fix: Always include original question in follow-up generation

4. Information Overload

Symptom: Too much context, LLM gets confused Fix: Limit total context tokens, summarize intermediate results

def summarize_context(self, contexts: list, question: str) -> str:
    """Summarize accumulated context to manage token budget"""
    if len(contexts) <= 5:
        return "\n".join([c["text"] for c in contexts])

    response = self.openai.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{
            "role": "user",
            "content": f"""Summarize these {len(contexts)} passages into key facts relevant to: {question}

{chr(10).join([c['text'] for c in contexts])}

Key facts (bullet points):"""
        }]
    )
    return response.choices[0].message.content

5. Missing Connections

Symptom: Has all facts but doesn't connect them Fix: Explicitly prompt for reasoning chain in generation

Advanced Patterns

Decompose-Then-Retrieve

def query_decomposed(self, question: str) -> dict:
    """Decompose first, then retrieve for each sub-question"""
    # Step 1: Decompose
    sub_questions = self.decompose(question)

    # Step 2: Retrieve for each sub-question
    all_results = []
    for sub_q in sub_questions:
        results = self.retrieve(sub_q, top_k=3)
        all_results.append({
            "sub_question": sub_q,
            "contexts": results
        })

    # Step 3: Synthesize
    # ... combine all results and generate answer

Parallel Multi-Hop

import asyncio

async def parallel_retrieve(self, queries: list[str]) -> list[list[dict]]:
    """Retrieve for multiple queries in parallel"""
    tasks = [self._async_retrieve(q) for q in queries]
    return await asyncio.gather(*tasks)

With Reranking

def multi_hop_with_rerank(self, question: str) -> dict:
    """Add reranking at each hop for higher precision"""
    # ... retrieve as usual
    # At each hop, rerank before adding to context
    reranked = self.reranker.rerank(query, results, top_k=3)
    all_contexts.extend(reranked)

When to Level Up

Symptom	Solution	Skill
Need dynamic conversation	Agent-based approach	agentic-rag
Need higher precision per hop	Add reranking	rag-with-rerank
External data sources needed	Full agent framework	LangChain/LlamaIndex

References

Internal:

• references/decomposition-strategies.md - Question decomposition techniques

Related skills:

• rag - Basic RAG (simpler questions)
• agentic-rag - Dynamic, conversational retrieval
• rag-with-rerank - Higher precision per hop

Verticals:

• verticals/troubleshooting.md - Multi-step diagnosis