MMOOZZGG.com

Glossary

What is RAG-Fusion

In short

RAG-Fusion is a document search method in which a model invents several versions of the question, searches with each and merges the results into one list.

Search and RAG · RAG-Fusion

MMOOZZGG.com editorsChecked: 1 min read

For a child

You lost your ball in the yard and you ask your friends. To one you say: “Where is my ball?” To another: “Did you see a red ball near the swings?” To a third: “Who played football yesterday?” Each one remembers something different, and together you find the ball faster. RAG-Fusion asks the computer one question in several ways at once and collects all the hints.

In plain words

People often ask a question badly or too briefly, and a search for one phrase misses what they need. RAG-Fusion works differently. First, the AI model writes several versions of the original question from different sides. The system searches for each version separately and then merges the lists: a passage found by several versions moves higher. The model answers from this shared result list. For example, the question “why does the power supply get hot” turns into queries about overheating, load, ventilation and faults, and the answer draws on sections from different manuals.

For business

The original paper tested the method on a task at the chipmaker Infineon: engineers, account managers and customers need product information fast. The approach fits when questions are short and the documentation is large and written in other words than buyers use. The cost grows in a predictable way: an extra model call to write the variants and several searches instead of one, so the answer comes more slowly. The same paper names the risk: if the generated variants drift from the point, the answer drifts too. So the number of variants is limited and tested on real customer questions. Search across a catalog and product specs is often the first job of an AI sales agent.

For a specialist

RAG-Fusion was described by Rackauckas (2024) on the example of a chatbot at Infineon. The method combines RAG with Reciprocal Rank Fusion (RRF). A language model writes several queries from the original question, a separate retrieval runs for each, and then the lists are merged. In the formula of Cormack, Clarke and Büttcher (2009), a document gets the sum of 1/(k + r) over all lists, where r is its position; the authors fixed k = 60 in a pilot study. The final passages go to the generator together with the original and generated queries. The author judged the answers by hand for accuracy, relevance and completeness. He links the gain to the fact that different wordings cover the question from several sides. The limit: if the generated queries are poorly linked to the original, the answer drifts off topic. The technique belongs to query expansion and is close to query rewriting and HyDE. But it builds a set of wordings instead of one and gains from merging them. The cost: N extra retrievals and a model call to write the queries.

Example

A buyer writes in a chat: “need a cable for outdoors for a pump”. The system searches for “outdoor cable”, “cable for a submersible pump” and “waterproof power cable”. It shows the manager three matching items from the catalog.

Related terms

Questions and answers

How is RAG-Fusion different from hybrid search? Hybrid search merges the results of different search methods for one query. RAG-Fusion merges the search results for several versions of the query.

In both cases the lists are often merged with the same RRF formula.

How many query variants should you generate? There is no fixed rule: more variants give more recall, but they raise the delay and the risk of drifting off topic. Choose the number on a set of real questions by comparing answer quality and time.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us