Skip to main content

Reciprocal Rank Fusion Retriever

Overview​

The Reciprocal Rank Fusion (RRF) Retriever is an advanced retrieval method that enhances search results by generating multiple queries and re-ranking the results. This feature improves the relevance and diversity of retrieved documents, especially for complex or ambiguous queries.

Key Benefits​

  • Improved search accuracy by considering multiple query variations
  • Enhanced result diversity through re-ranking of documents
  • Better handling of complex or ambiguous user queries

How to Use​

  1. Add the "Reciprocal Rank Fusion Retriever" node to your AnswerAI workflow canvas.

  2. Connect the required inputs:

    • Vector Store Retriever: Connect a vector store retriever node.
    • Language Model: Connect a language model node (e.g., GPT-3.5, GPT-4).

  3. Configure the node parameters:

    • Query (optional): Specify a custom query or leave blank to use the user's question.
    • Query Count: Set the number of synthetic queries to generate (default is 4).
    • Top K: Specify the number of top results to fetch (default is the base retriever's Top K).
    • Constant: Set the constant value for the RRF algorithm (default is 60).

  4. Connect the output to subsequent nodes in your workflow.

RRF Retreiver Node & Drop UI

Tips and Best Practices​

  1. Experiment with different Query Count values to find the optimal balance between result quality and processing time.

  2. Adjust the Top K value based on your specific use case. A higher value may provide more comprehensive results but might increase processing time.

  3. Fine-tune the Constant value to control the balance between high-ranked and lower-ranked items in the results.

  4. Use this retriever for queries that might benefit from multiple perspectives or interpretations.

Troubleshooting​

  1. If results seem less relevant than expected:

    • Increase the Query Count to generate more variations of the original query.
    • Adjust the Top K value to retrieve more initial results before re-ranking.

  2. If processing time is too long:

    • Reduce the Query Count or Top K values.
    • Ensure your vector store and language model are optimized for performance.

  3. If you receive an error about missing inputs:

    • Double-check that both the Vector Store Retriever and Language Model inputs are properly connected.