Class ReciprocalRankFusion

Namespace

MailFathom.Application.Emails.Search

Assembly

MailFathom.Application.dll

Combines two rankings of the same emails into one, reading only where each ranking placed them.

public static class ReciprocalRankFusion

Inheritance

object

ReciprocalRankFusion

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

Reciprocal Rank Fusion is the method the architecture draft names, and the reason it names it is that the two inputs are not on one scale and never will be. A full-text rank is a function of term frequency and document length; a vector distance is a function of a model's geometry. Adding, averaging, or min-max normalizing them would ask which number is worth more, and every answer to that would be a constant that a change of embedding model silently invalidates. Fusion by rank asks nothing of the numbers at all — only of the order they put the documents in — so changing the model changes which documents are found and never how the two findings are weighed.

The formula is the published one: a document scores 1 / (k + rank) in each ranking that returned it, counting ranks from one, and its fused score is the sum over both. A document only one ranking found therefore still scores, which is what lets a semantic match with no shared word rank at all — and a document both found outranks either of them, which is what keeps an exact phrase from being displaced by a merely related message.

The result is deterministic for a given pair of inputs. Equal fused scores are not rare here and are not a tie-breaking detail: two documents at symmetric places — first lexically and fifth semantically against fifth and first — score identically by construction. The timeline order settles those, which is the same order that settles a tie in the lexical ranking alone, so a search returns one sequence rather than whichever the sums happened to produce.

Fields

RankConstant

The constant that flattens the difference between the very top places.

Methods

Fuse(IReadOnlyList<RankedEmailCandidate>, IReadOnlyList<RankedEmailCandidate>, int)

Fuses a lexical ranking and a semantic ranking into one bounded ordering.