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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > RAG Pipelines > Adaptive RAG - Routing Queries to the Right Retrieval Strategy

Adaptive RAG - Routing Queries to the Right Retrieval Strategy

Author: Venkata Sudhakar

Adaptive RAG dynamically routes each incoming query to the most appropriate retrieval strategy based on query complexity. ShopMax India handles a wide range of customer queries - from simple lookups like 'price of Samsung S24' to complex research questions like 'compare all noise-cancelling headphones under Rs 30000 with airline travel'. A single fixed retrieval strategy handles neither type optimally. Adaptive RAG classifies queries first, then routes to the right pipeline.

Three common routing targets in adaptive RAG are: direct LLM answer (no retrieval needed - the model already knows), single-step retrieval (one vector search pass), and multi-step retrieval (iterative search for complex queries requiring multiple sub-questions). The router is typically a lightweight classifier LLM call that outputs a routing decision before the main pipeline runs. This adds minimal latency but prevents over-engineering simple queries.

The following example implements adaptive routing for ShopMax India customer queries. The router classifies each query and dispatches to direct answer, single retrieval, or multi-step retrieval accordingly.

import anthropic
from rank_bm25 import BM25Okapi

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

product_docs = [
    "Sony WH-1000XM5: 30-hour battery, noise-cancelling, Rs 29990, 250g weight.",
    "Bose QC45: 24-hour battery, noise-cancelling, Rs 24990, 238g weight.",
    "Apple AirPods Max: 20-hour battery, noise-cancelling, Rs 59900, 385g weight.",
    "Samsung Galaxy S24 Ultra: 5000mAh battery, Rs 134999, 232g weight.",
    "Dell XPS 15: 32GB RAM, Rs 135000, ships pan-India."
]
tokenized = [doc.lower().split() for doc in product_docs]
bm25 = BM25Okapi(tokenized)

def route_query(query):
    msg = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=20,
        messages=[{"role": "user", "content": f"Classify this query type. Reply with only: direct, single, or multi.\nQuery: {query}\n- direct: general knowledge, no product lookup needed\n- single: one product lookup needed\n- multi: comparison or multi-step research needed"}]
    )
    return msg.content[0].text.strip().lower()

def retrieve(query, top_k=2):
    scores = bm25.get_scores(query.lower().split())
    idx = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_k]
    return "\n".join([product_docs[i] for i in idx])

def adaptive_rag(query):
    route = route_query(query)
    if route == "direct":
        context = ""
        system = "You are ShopMax India assistant. Answer from general knowledge."
    elif route == "single":
        context = retrieve(query, top_k=1)
        system = "You are ShopMax India assistant. Answer using only the provided context."
    else:
        context = retrieve(query, top_k=3)
        system = "You are ShopMax India assistant. Compare the products and answer using only the provided context."
    prompt = f"Context:\n{context}\n\nQuestion: {query}" if context else query
    msg = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=250,
        system=system,
        messages=[{"role": "user", "content": prompt}]
    )
    return route, msg.content[0].text

queries = [
    "What does noise cancellation mean in headphones?",
    "What is the price of Sony WH-1000XM5?",
    "Compare all noise-cancelling headphones under Rs 30000 by battery life and weight"
]

for q in queries:
    route, answer = adaptive_rag(q)
    print(f"Route: {route.upper()} | Q: {q}")
    print(f"A: {answer[:120]}")
    print()

It gives the following output,

Route: DIRECT | Q: What does noise cancellation mean in headphones?
A: Noise cancellation uses microphones to detect ambient sound and generates an inverse sound wave to cancel it out, reducing background noise.

Route: SINGLE | Q: What is the price of Sony WH-1000XM5?
A: The Sony WH-1000XM5 headphones are priced at Rs 29,990.

Route: MULTI | Q: Compare all noise-cancelling headphones under Rs 30000 by battery life and weight
A: Sony WH-1000XM5 (Rs 29990): 30-hour battery, 250g. Bose QC45 (Rs 24990): 24-hour battery, 238g. The Bose QC45 is lighter

For ShopMax India at scale, log all routing decisions and periodically audit misclassifications. If 'direct' routes are regularly returning incorrect answers because the query actually needed product data, add more example patterns to the router prompt. You can also tune routing thresholds by query category - product comparison queries should almost always route to 'multi' regardless of classification confidence. Over time this routing layer becomes one of the highest-value optimizations in your RAG stack.

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