Reranker

Rerankers used in hybrid search to combine results from multiple vector searches.

public interface IReranker
{
    IReadOnlyList<KeyValuePair<string, string>> ToRankParams();
    FunctionScore? ToFunctionScore();
}

METHODS:

  • ToRankParams() -

    Gets the rank parameters describing this reranker’s strategy, used for rank_params-based strategies such as RRF and weighted.

  • ToFunctionScore() -

    Gets a structured FunctionScore for this reranker, or null when the reranker uses the rank_params channel instead. See FunctionScore.

FunctionReranker

Base class for function-score (boost/decay/model) rerankers; converts to a FunctionScore for the proto schema.FunctionScore wire member.

public abstract class FunctionReranker<T> : IReranker where T : FunctionReranker<T>
{
    public FunctionType Type { get; }
    public string RerankerName { get; }
    public IDictionary<string, string> Params { get; }
    public IList<string> InputFieldNames { get; }
}

PARAMETERS:

  • Type (FunctionType) -

    The function type (always FunctionType.Rerank for rerankers). See FunctionType.

  • RerankerName (string) -

    The rerank strategy name ("boost" / "decay" / "model") the server identifies the function by.

  • Params (IDictionary) -

    The rerank function’s parameters (boost/decay/model settings).

  • InputFieldNames (IList) -

    The input field names the rerank function operates on; decay/model require one.

RrfReranker

A Reciprocal Rank Fusion (RRF) reranker that combines results based on their reciprocal ranks, scoring each document as the sum of 1 / (k + rank) across all input searches.

public sealed class RrfReranker : IReranker
{
    public RrfReranker(float k) { ... }
    public float K { get; }
}

PARAMETERS:

  • K (float) -

    The constant used in the RRF formula; higher values give more weight to lower-ranked results. Defaults to 60.

METHODS:

  • RrfReranker(float k) (constructor) -

    Creates a new RRF reranker with a specified k value, which must be greater than or equal to 1.

WeightedReranker

A weighted reranker that combines results using specified weights for each input search; the final score of each document is a weighted sum of its scores from each input search.

public sealed class WeightedReranker : IReranker
{
    public WeightedReranker(params float[] weights) { ... }
    public IReadOnlyList<float> Weights { get; }
}

PARAMETERS:

  • Weights (IReadOnlyList) -

    The weights to apply to each input search. The number of weights must match the number of ANN search requests.

METHODS:

  • WeightedReranker(params float[] weights) (constructor) -

    Creates a new weighted reranker with the specified weights.

BoostRerank

A function-score reranker that boosts results matching a filter (or random score on a field), mapped to the proto schema.FunctionScore.

public sealed class BoostRerank : FunctionReranker<BoostRerank>
{
    public BoostRerank(string name) { ... }
    public BoostRerank WithFilter(string filter) { ... }
    public BoostRerank WithWeight(float weight) { ... }
    public BoostRerank WithRandomScoreField(string field) { ... }
    public BoostRerank WithRandomScoreSeed(long seed) { ... }
}

METHODS:

  • BoostRerank(string name) (constructor) -

    Creates a boost reranker.

  • WithFilter(string filter) -

    Sets the filter expression; results matching it get the boost.

  • WithWeight(float weight) -

    Sets the boost weight applied to matching results.

  • WithRandomScoreField(string field) -

    Sets the field used to generate a random score.

  • WithRandomScoreSeed(long seed) -

    Sets the seed for the random score.

DecayRerank

A function-score reranker applying a gaussian/exponential/linear decay to results by a scalar field, mapped to the proto schema.FunctionScore.

public sealed class DecayRerank : FunctionReranker<DecayRerank>
{
    public DecayRerank(string name) { ... }
    public DecayRerank WithInputField(string field) { ... }
    public DecayRerank WithFunction(string function) { ... }
    public DecayRerank WithOrigin<T>(T origin) { ... }
    public DecayRerank WithOffset<T>(T offset) { ... }
    public DecayRerank WithScale<T>(T scale) { ... }
    public DecayRerank WithDecay(float decay) { ... }
}

METHODS:

  • DecayRerank(string name) (constructor) -

    Creates a decay reranker.

  • WithInputField(string field) -

    Sets the field the decay applies to; the server requires exactly one input field for decay functions.

  • WithFunction(string function) -

    Sets the decay function: "gauss", "exp" or "linear".

  • WithOrigin<T>(T origin) -

    Sets the reference point (origin) of the decay curve.

  • WithOffset<T>(T offset) -

    Sets the no-decay zone around the origin.

  • WithScale<T>(T scale) -

    Sets the scale: the position at which relevance drops to the decay value.

  • WithDecay(float decay) -

    Sets the decay value (the score at the scale position).

ModelRerank

A function-score reranker that uses a model service to recompute relevance, mapped to the proto schema.FunctionScore.

public sealed class ModelRerank : FunctionReranker<ModelRerank>
{
    public ModelRerank(string name) { ... }
    public ModelRerank WithProvider(string provider) { ... }
    public ModelRerank WithQueries(IEnumerable<string> queries) { ... }
    public ModelRerank WithEndpoint(string endpoint) { ... }
}

METHODS:

  • ModelRerank(string name) (constructor) -

    Creates a model reranker.

  • WithProvider(string provider) -

    Sets the model service provider.

  • WithQueries(IEnumerable<string> queries) -

    Sets the query strings used by the model to compute relevance; the count must equal the number of queries in the search.

  • WithEndpoint(string endpoint) -

    Sets the model service endpoint URL.

Example

using Milvus.Client.V2;
using Milvus.Client.V2.Requests.Dql;
using Milvus.Client.V2.Responses.Dql;
using Milvus.Client.V2.Types;

using var client = new MilvusClientV2(new ConnectConfig
{
    Uri = "localhost:19530",
    Username = "root",
    Password = "Milvus"
});

SearchResp results = await client.HybridSearchAsync(new HybridSearchReq
{
    CollectionName = "my_collection",
    SearchRequests =
    [
        new AnnSearchReq
        {
            VectorFieldName = "embedding",
            Vectors = [new[] { 0.1f, 0.2f, 0.3f, 0.4f }],
            MetricType = MetricType.L2,
            Limit = 10
        },
        new AnnSearchReq
        {
            VectorFieldName = "sparse",
            Texts = ["The Hobbit"],
            Limit = 10
        }
    ],
    Reranker = new RrfReranker(60),
    Limit = 5
});

Try Managed Milvus for Free

Zilliz Cloud is hassle-free, powered by Milvus and 10x faster.

Get Started
Feedback

Was this page helpful?