Milvex.Ranker (milvex v0.10.2)

Copy Markdown

Builder functions for hybrid search rerankers.

Provides three reranking strategies:

  • weighted/1 - Weighted average scoring
  • rrf/1 - Reciprocal Rank Fusion
  • decay/2 - Decay function scoring based on field proximity

Examples

{:ok, ranker} = Ranker.weighted([0.7, 0.3])
{:ok, ranker} = Ranker.rrf(k: 60)
{:ok, ranker} = Ranker.decay(:gauss, field: "timestamp", origin: 1_710_000_000, scale: 86_400)

Summary

Functions

decay(function, opts)

Creates a decay function ranker for hybrid search.

rrf(opts \\ [])

Creates an RRF (Reciprocal Rank Fusion) ranker.

weighted(weights)

Creates a weighted ranker with the given weights.

Functions

decay(function, opts)

@spec decay(
  :gauss | :exp | :linear,
  keyword()
) :: {:ok, Milvex.Ranker.DecayRanker.t()} | {:error, Milvex.Errors.Invalid.t()}

Creates a decay function ranker for hybrid search.

Applies a decay curve to a numeric field, scoring documents higher when the field value is closer to the origin.

Parameters

  • function - Decay curve: :gauss, :exp, or :linear
  • opts - Keyword list of options

Options

  • :field - (required) Name of the numeric field
  • :origin - (required) Center point (integer)
  • :scale - (required) Distance from origin where score = decay (positive integer)
  • :offset - Zone around origin with full score (default: 0)
  • :decay - Score at scale distance (default: 0.5, must be between 0 and 1 exclusive)

Examples

{:ok, ranker} = Ranker.decay(:gauss, field: "timestamp", origin: 1_710_000_000, scale: 86_400)
{:ok, ranker} = Ranker.decay(:exp, field: "publish_time", origin: 1_710_000_000, scale: 3600, offset: 300, decay: 0.3)

rrf(opts \\ [])

@spec rrf(keyword()) ::
  {:ok, Milvex.Ranker.RRFRanker.t()} | {:error, Milvex.Errors.Invalid.t()}

Creates an RRF (Reciprocal Rank Fusion) ranker.

Options

  • :k - Smoothness parameter (default: 60, must be positive)

Examples

{:ok, ranker} = Ranker.rrf()
{:ok, ranker} = Ranker.rrf(k: 100)

weighted(weights)

@spec weighted([number()]) ::
  {:ok, Milvex.Ranker.WeightedRanker.t()} | {:error, Milvex.Errors.Invalid.t()}

Creates a weighted ranker with the given weights.

Each weight corresponds to a sub-search in the hybrid search. The number of weights must match the number of searches.

Examples

{:ok, ranker} = Ranker.weighted([0.8, 0.2])