This package provides an implementation of the Reciprocal Rank Fusion (RRF) algorithm in C#. The RRF algorithm is used to combine search results from multiple sources, taking into account the ranks of documents to produce a fused ranking.

• Introduction
• Usage
  • Installation
  • Basic Usage
• Complete Example
• Choosing the k Parameter
  • For 1 to 10 Search Results
  • For Up to 100 Search Results with Highly Relevant Top 10
• License
• Acknowledgments

Reciprocal Rank Fusion (RRF) is a simple yet effective method for merging ranked lists of documents. This algorithm assigns scores to documents based on their ranks across multiple queries, and then combines these scores to produce a final ranking.

You can install the drittich.ReciprocalRankFusion NuGet package using the NuGet Package Manager or the .NET CLI.

1. Open your project in Visual Studio.
2. Right-click on your project in the Solution Explorer and select Manage NuGet Packages.
3. In the Browse tab, search for drittich.ReciprocalRankFusion.
4. Select the package from the search results and click Install.

Run the following command in your terminal:

dotnet add package drittich.ReciprocalRankFusion

After installing the package, you can use the ReciprocalRankFusion class to fuse your search results.

var searchResultsDict = ...

var fusedResults = ReciprocalRankFusion.FuseSearchResults(searchResultsDict);
foreach (var result in fusedResults)
{
    Console.WriteLine($"Document: {result.Key}, Score: {result.Value}");
}

Here is an example of how to use the FuseSearchResults method:

using drittich.ReciprocalRankFusion;

var searchResultsDict = new Dictionary<string, Dictionary<string, double>>
{
    { "query1", new Dictionary<string, double> { 
        { "doc1", 3.0 }, 
        { "doc2", 1.5 }, 
        { "doc3", 2.0 } 
    }},
    { "query2", new Dictionary<string, double> { 
        { "doc4", 2.5 }, 
        { "doc5", 3.5 }, 
        { "doc6", 1.0 } 
    }}
};

var fusedResults = ReciprocalRankFusion.FuseSearchResults(searchResultsDict);
var top100Results = fusedResults
    .Take(100)
    .ToDictionary(x => x.Key, x => x.Value);
foreach (var result in fusedResults)
{
    Console.WriteLine($"Document: {result.Key}, Score: {result.Value}");
}

/*
Output:

Document: doc1, Score: 0.016666666666666666
Document: doc5, Score: 0.016666666666666666
Document: doc3, Score: 0.01639344262295082
Document: doc4, Score: 0.01639344262295082
Document: doc2, Score: 0.016129032258064516
Document: doc6, Score: 0.016129032258064516
*/

The k parameter in the Reciprocal Rank Fusion (RRF) algorithm plays a pivotal role in balancing the influence of document ranks on the final fused score. Selecting an appropriate value for k depends on the number of search results you aim to retrieve and the desired relevance distribution among them.

When your goal is to retrieve a small set of highly relevant results (1 to 10), the choice of k should emphasize the top ranks more strongly to ensure that the most relevant documents are prioritized.

Recommended k Values:

• k = 5 to 15

  • k = 5: Provides strong emphasis on the top ranks, making the algorithm highly sensitive to the highest-ranked documents.
  • k = 10: Offers a balanced approach, still prioritizing top ranks while allowing some flexibility.
  • k = 15: Slightly less aggressive in emphasizing top ranks but still suitable for small result sets.

Rationale:

• Lower k values increase the weight difference between higher and lower ranks, ensuring that the top 1-10 results are highly relevant.
• Higher k values within this range provide a balance, preventing the top results from being too narrowly focused.

Example Usage:

var fusedResults = ReciprocalRankFusion.FuseSearchResults(searchResultsDict, k: 10);

// Extract the top 10 results
var top10Results = fusedResults.Take(10).ToDictionary(x => x.Key, x => x.Value);

When you aim to retrieve a larger set of results (up to 100) while ensuring that the first 10 are highly relevant, a different range of k values is appropriate to balance overall coverage with top-tier relevance.

Recommended k Values:

• k = 30 to 50

  • k = 30: Strong emphasis on top ranks while allowing meaningful contributions from ranks beyond 10.
  • k = 40: Provides a balanced approach, maintaining high relevance in the top 10 and comprehensive coverage up to 100 results.
  • k = 50: Slightly less emphasis on the very top ranks but ensures a diverse and comprehensive set of results.

Rationale:

• Moderate k values ensure that while the top 10 results are prioritized for high relevance, documents ranked up to 100 still contribute significantly to the final ranking.
• This range prevents the dilution of the top results' relevance while maintaining the algorithm's ability to consider a broader set of documents.

Example Usage:

var fusedResults = ReciprocalRankFusion.FuseSearchResults(searchResultsDict, k: 40);

// Extract the top 100 results
var top100Results = fusedResults.Take(100).ToDictionary(x => x.Key, x => x.Value);

• Start with Recommended Values: Use the suggested ranges as a baseline.
• Empirical Testing: Evaluate different k values using your specific datasets and relevance metrics to find the optimal balance.
• Adjust Based on Results: Depending on the performance, you might fine-tune k upwards or downwards within the recommended ranges.
• Consider Application Needs: Align the choice of k with the specific requirements of your application, such as the importance of top results versus overall coverage.

This project is licensed under the MIT License. See the LICENSE file for details.

This project contains code ported from the python source code in rag-fusion.