description: Reciprocal Rank Fusion to re-rank search results by multiple query generation.

Reciprocal Rank Fusion Retriever

The Reciprocal Rank Fusion (RRF) Retriever is a specialized retriever that uses the Reciprocal Rank Fusion algorithm to re-rank search results obtained from multiple query generations. This node enhances the retrieval process by generating synthetic queries and combining their results for improved relevance.

Input Parameters

1. Vector Store Retriever (required)

   • Type: VectorStoreRetriever

   • Description: The base retriever used for initial document retrieval.

2. Language Model (required)

   • Type: BaseLanguageModel

   • Description: The language model used for generating synthetic queries.

3. Query (optional)

   • Type: string

   • Description: Custom query to retrieve documents. If not specified, the user’s question will be used.

4. Query Count (optional)

   • Type: number

   • Default: 4

   • Description: Number of synthetic queries to generate.

5. Top K (optional)

   • Type: number

   • Description: Number of top results to fetch. Defaults to the TopK of the Base Retriever.

6. Constant (optional)

   • Type: number

   • Default: 60

   • Description: A constant added to the rank, controlling the balance between high-ranked and lower-ranked items.

Outputs

1. Reciprocal Rank Fusion Retriever

• Type: RRFRetriever, BaseRetriever

• Description: The initialized RRF retriever object.

2. Document

• Type: Document, json

• Description: Array of document objects containing metadata and pageContent.

3. Text

• Type: string, json

• Description: Concatenated string from pageContent of retrieved documents.

How It Works

1. The node initializes a ReciprocalRankFusion object using the provided language model, base retriever, and configuration parameters.

2. It then wraps this object in a ContextualCompressionRetriever for additional processing.

3. Depending on the selected output, the node can return:

   • The retriever object itself

   • An array of relevant documents

   • A concatenated text of the retrieved documents' content

Use Cases

• Improving search relevance in document retrieval systems

• Enhancing question-answering systems with more diverse and relevant context

• Boosting the performance of RAG (Retrieval-Augmented Generation) applications

Notes

• The node uses escape character handling for text output to ensure proper formatting.

• The synthetic query generation and re-ranking process happens internally within the ReciprocalRankFusion class.

{% hint style="info" %} This section is a work in progress. We appreciate any help you can provide in completing this section. Please check our Contribution Guide to get started. {% endhint %}