	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > RAG Pipelines > RAG Fusion - Reciprocal Rank Fusion for Multi-Source Retrieval

RAG Fusion - Reciprocal Rank Fusion for Multi-Source Retrieval

Author: Venkata Sudhakar

RAG Fusion improves retrieval quality by generating multiple query variants for a single user question, retrieving documents for each variant separately, then merging the ranked result lists using Reciprocal Rank Fusion (RRF). ShopMax India benefits from this when customers use imprecise language - a query like 'good wireless earbuds for gym' may not retrieve the same documents as 'sweat-resistant in-ear headphones for exercise', but RAG Fusion captures both perspectives in one pass.

Reciprocal Rank Fusion scores each document by summing 1/(rank + k) across all result lists, where k is a constant (typically 60) that dampens the influence of high ranks. Documents that appear in the top results of multiple retrieval lists score highest. This approach is robust to any single retrieval failing and systematically surfaces documents that are broadly relevant across query phrasings.

The following example implements RAG Fusion for ShopMax India product search. The system generates three query variants using an LLM, retrieves documents for each, applies RRF to merge the rankings, and sends the top-fused documents to the answer generator.

import anthropic
from rank_bm25 import BM25Okapi

client = anthropic.Anthropic(api_key="sk-ant-...")

product_docs = [
    "Sony WH-1000XM5: noise-cancelling, 30-hour battery, 250g, Rs 29990, IP rating none.",
    "Bose QC45: noise-cancelling, 24-hour battery, 238g, Rs 24990, no IP rating.",
    "Jabra Elite 7 Active: sweat-proof IP57, 9-hour battery, 5.5g, Rs 12990, gym-ready.",
    "Samsung Galaxy Buds2 Pro: IPX7 sweat-resistant, 8-hour battery, 5.5g, Rs 17990.",
    "Apple AirPods Pro 2: IPX4 sweat-resistant, 6-hour battery, 5.3g, Rs 26900."
]
tokenized = [doc.lower().split() for doc in product_docs]
bm25 = BM25Okapi(tokenized)

def generate_query_variants(query, n=3):
    msg = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=150,
        messages=[{"role": "user", "content": f"Generate {n} alternative phrasings of this product search query. Return one per line, no numbering.\nQuery: {query}"}]
    )
    lines = [l.strip() for l in msg.content[0].text.strip().split("\n") if l.strip()]
    return lines[:n]

def retrieve_scores(query, top_k=5):
    scores = bm25.get_scores(query.lower().split())
    ranked = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_k]
    return ranked

def rrf_merge(all_rankings, k=60):
    scores = {}
    for ranking in all_rankings:
        for rank, doc_idx in enumerate(ranking):
            scores[doc_idx] = scores.get(doc_idx, 0) + 1.0 / (rank + k)
    return sorted(scores.keys(), key=lambda d: scores[d], reverse=True)

def rag_fusion_query(query, top_k=3):
    variants = generate_query_variants(query) + [query]
    all_rankings = [retrieve_scores(v) for v in variants]
    fused = rrf_merge(all_rankings)[:top_k]
    context = "\n".join([product_docs[i] for i in fused])
    msg = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=200,
        system="You are ShopMax India assistant. Answer using only the provided context.",
        messages=[{"role": "user", "content": f"Context:\n{context}\n\nQuestion: {query}"}]
    )
    return msg.content[0].text

print(rag_fusion_query("best wireless earbuds for gym workouts under Rs 20000"))

It gives the following output,

For gym workouts under Rs 20,000, the Jabra Elite 7 Active is the top recommendation at Rs 12,990. It features IP57 sweat-proof protection, making it ideal for intense exercise. The Samsung Galaxy Buds2 Pro at Rs 17,990 is also a strong option with IPX7 water resistance and a secure fit for workouts.

For ShopMax India at scale, cache the query variant generation step using a hash of the original query so repeated queries do not regenerate variants. You can also skip LLM-based variant generation and use a synonym dictionary for common product terms - replacing 'earbuds' with ['in-ear headphones', 'true wireless earphones'] programmatically. This hybrid approach delivers RAG Fusion benefits at lower latency and cost for high-traffic product categories like headphones and smartphones.

Send your comments, suggestions or queries regarding this site to [email protected].