liveBook · Manning

QueryFusionRetriever

As explained in Retrieval Augmented Generation, The Foundational Ideas

QueryFusionRetriever is a LlamaIndex component that generates multiple queries, retrieves results, and fuses the resulting rankings.

QueryFusionRetriever improves retrieval reliability by generating query variants, searching with multiple retrievers, and fusing the resulting rankings before answer synthesis. section 8.4

A typical configuration:

  • Retains the original query and generates four additional queries. section 8.4
  • Runs all five queries through vector and BM25 retrievers. section 8.4
  • Combines the ten result sets with reciprocal rank fusion, then passes reranked documents to the language model. section 8.4

Laptop search

Query variants such as “durable laptops for students,” “lightweight laptops for carrying to class,” and “best budget laptops for college” produce overlapping document lists. Reciprocal rank fusion promotes documents appearing consistently across lists.

Figure 8.3 RAG-Fusion complete data flow.

Use the complete data-flow diagram to trace query expansion, parallel retrieval, rank fusion, and answer synthesis. section 8.4

How Retrieval Augmented Generation, The Foundational Ideas covers it

How deep it goes
Explains query generation, parallel retrieval, reciprocal rank fusion, and LlamaIndex wiring, but not fusion internals beyond ranking output.
Examples
Demonstrates laptop query variants, fused document rankings, and Chelsea AI Ventures web retrieval.
What it assumes
Assumes familiarity with ordinary RAG, vector indexes, BM25 retrieval, language models, and document loading.
What you can do after
Apply it via LlamaIndex to build a multi-query retriever combining vector and BM25 results.
How it connects
Connects multi-query retrieval with hybrid search, vector retrieval, BM25, reciprocal rank fusion, and RAG.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Retrieval Augmented Generation, The Foundational Ideas. The section links point to the parts of the book it draws on.