Subtitle: Learn how an AI decomposes complex questions, searches for each intent separately, and merges the results to deliver more relevant answers without making the customer repeat their request.
1. The Problem: One Question Can Have Multiple Intents
Imagine a customer asks Varastra, an Indian ethnic-wear store:
“Show me something festive for a wedding and something light for warm weather.”
This request contains two intentions:
- Find something festive for a wedding.
- Find something lightweight for warm weather.
A single search may focus on one part and miss the other.
Multi-query retrieval helps by splitting a multi-intent question into smaller searches, retrieving results for each, and combining them.
2. Understand the Architecture
Follow the arrows in the diagram:
Customer query → LLM decomposer → One-intent or two-intent path
- One intent: Use the standard hybrid retrieval and reranking path, returning the top three results.
- Two intents: Create two paraphrased sub-queries, retrieve results for each, then combine them using Reciprocal Rank Fusion (RRF).
The final merged results contain the top three fused candidates for both intents.
3. Step 1: The Customer Sends a Query
The customer asks a question through Varastra's shopping assistant.
For example:
“I need something festive and something comfortable for hot weather.”
The system must determine whether this question expresses one search intent or two distinct intentions.
That decision determines the path shown in the diagram.
4. Step 2: The LLM Decomposer Identifies the Intent
An LLM decomposer examines the customer's question and determines whether it should be handled as one intent or split into two.
An intent is the goal behind a request.
For example:
- “Find a festive kurta” expresses one main intent.
- “Find a festive kurta and a lightweight summer outfit” can express two intents.
The diagram shows two possible outcomes.
Why it matters: Decomposition helps the system avoid overlooking a meaningful part of a complex request.
The diagram does not specify the exact prompt or decision rules used by the decomposer.
5. Step 3A: One Intent Follows the Standard Path
If the decomposer identifies one intent, the query goes directly to:
Hybrid retrieval + reranking → Top 3
Hybrid retrieval combines keyword search and semantic search to find candidate products. Reranking then orders those candidates by relevance.
For example, a customer asks:
“Find a festive cotton kurta.”
The system searches for suitable products and returns the top three candidates from the standard path.
Why keep this path simple? A straightforward question does not need to be split into multiple searches unnecessarily.
6. Step 3B: Two Intents Become Two Sub-Queries
If the decomposer identifies two intents, it creates two paraphrased sub-queries.
The diagram gives these example labels:
- Sub-query A: “festive”
- Sub-query B: “warm”
These are simplified examples showing the two search directions.
In a real shopping request, the sub-queries should preserve the customer's actual meaning. For example:
- Sub-query A: “Festive outfits suitable for a wedding.”
- Sub-query B: “Lightweight, breathable outfits for hot weather.”
Paraphrasing means expressing the same underlying intent in different words that are useful for search.
Why split the query? Each search can focus on one requirement instead of competing with several requirements at once.
7. Step 4: Retrieve and Rerank Each Sub-Query Separately
Both sub-queries pass through their own hybrid retrieval + reranking path.
Each path finds candidate products relevant to its particular intent.
For example:
| Sub-query | Possible candidates |
|---|---|
| Festive outfits | Embroidered kurta, festive Anarkali, occasion-wear set |
| Lightweight outfits | Cotton kurta, linen co-ord set, breathable cotton dress |
These examples illustrate the process; they are not actual Varastra search results.
The two result lists may overlap. A lightweight festive kurta, for example, might appear in both.
Why search separately? Separate retrieval helps preserve both intentions and gives the merging stage evidence from each search.
8. Step 5: RRF Merges the Results
The two ranked lists flow into RRF merge, which stands for Reciprocal Rank Fusion.
RRF combines rankings from multiple lists using each candidate's position rather than simply adding raw retrieval scores.
A common RRF formula is:
RRF score(d) = Σ 1 / (k + rank(d))
Here:
dis a candidate product or document.rank(d)is its position in a result list.kis a ranking constant.
The diagram does not specify the value of k for this particular multi-query merge.
Simple Example
Suppose a cotton kurta ranks:
- 2nd in the festive search.
- 1st in the lightweight search.
Because it ranks highly in both lists, RRF can give it a strong combined score.
A product that ranks highly in only one list can still remain competitive. The final ranking depends on the positions of all candidates across the lists.
The system then selects the top three fused candidates, as shown in the diagram.
Important: RRF combines rankings; it does not independently verify that a product satisfies every customer requirement.
9. Step 6: The Customer Receives One Seamless Experience
The diagram's final note says:
“Single intent passes through. Two intents split, retrieve, merge—the customer never sees the seam.”
This means the internal search process may branch into multiple paths, but the customer does not need to manage those searches.
The assistant can use the retrieved results to construct its response.
If the request requires two distinct products, the application must preserve that distinction when assembling the final answer. A merged top-three list alone does not guarantee that both intentions are adequately represented.
10. What Can Go Wrong?
| Problem | Why it matters |
|---|---|
| The decomposer misses an intent | Part of the customer's request may be ignored. |
| A query is split unnecessarily | Extra retrieval calls can increase latency and cost. |
| A sub-query changes the original meaning | Search may return unsuitable products. |
| One branch returns poor results | That intent may be underrepresented in the merge. |
| Both branches return the same candidates | The final list may lack variety. |
| RRF ranks a product highly that misses a requirement | Ranking does not replace business-rule validation. |
The team should evaluate intent detection, sub-query quality, retrieval relevance, merged rankings, and end-to-end response quality.
Prices and stock availability should still be verified against authoritative, current product data.
11. Key Takeaways
- One customer question can contain multiple search intentions.
- An LLM decomposer decides whether to use one path or split the query.
- Single-intent questions follow the standard hybrid retrieval and reranking path.
- Multi-intent questions generate separate sub-queries.
- Each sub-query is retrieved and reranked independently.
- RRF combines the ranked results into a fused list.
- The diagram specifies a top-three result list after merging.
- Good decomposition and retrieval are essential; merging cannot fix every upstream failure.
The big lesson: A useful AI search system must understand not just what words a customer used, but how many different things the customer is asking it to find.