Interface IReranker<T>

Namespace

AiDotNet.Interfaces

Assembly

AiDotNet.dll

Defines the contract for reranking retrieved documents to improve relevance ordering.

public interface IReranker<T>

Type Parameters

T

The numeric data type used for relevance scoring.

Remarks

A reranker refines the ordering of initially retrieved documents using more sophisticated relevance scoring. While initial retrieval must be fast and may use simple similarity metrics, reranking can employ computationally expensive methods like cross-encoders or large language models to achieve better relevance rankings for the final result set.

For Beginners: A reranker is like a second opinion on search results.

Think of it like a two-stage hiring process:

Stage 1 (Initial Retrieval):

• Quick screening of 1000 applicants
• Filter to top 20 based on resume keywords
• Fast but might miss some good candidates

Stage 2 (Reranking):

• Detailed review of those 20 candidates
• Deeper analysis of experience and fit
• Slower but more accurate
• Final ranking of best 5

Similarly, reranking takes the initial search results and re-orders them using more sophisticated analysis, ensuring the best results appear first.

Properties

ModifiesScores

Gets a value indicating whether this reranker modifies relevance scores.

bool ModifiesScores { get; }

Property Value

bool

Remarks

Some rerankers (like identity/pass-through rerankers) don't modify scores, while others compute new scores. This property indicates the behavior.

For Beginners: This tells you if the reranker actually does something.

• False: Pass-through reranker (no changes, just returns documents as-is)
• True: Active reranker (recalculates scores and reorders documents)

A pass-through reranker is useful as a default when you don't need reranking, but want to keep the code structure consistent.

Methods

Rerank(string, IEnumerable<Document<T>>)

Reranks a collection of documents based on their relevance to a query.

IEnumerable<Document<T>> Rerank(string query, IEnumerable<Document<T>> documents)

Parameters

query string

The query text used to assess relevance.

documents IEnumerable<Document<T>>

The documents to rerank.

Returns

IEnumerable<Document<T>>

The documents reordered by relevance, with updated relevance scores.

Remarks

This method takes documents (typically from an initial retrieval stage) and reorders them using the reranker's scoring strategy. The returned documents have their RelevanceScore properties updated to reflect the reranker's assessment. Documents are returned in descending order of relevance.

For Beginners: This re-sorts search results to put the best matches first.

For example:

• Input: 10 documents from initial search (roughly ordered)
• Processing: Reranker deeply analyzes each one
• Output: Same 10 documents, but better ordered with updated scores

The first document in the output is now the absolute best match, not just a good match like it might have been before reranking.

Rerank(string, IEnumerable<Document<T>>, int)

Reranks documents and returns only the top-k highest scoring results.

IEnumerable<Document<T>> Rerank(string query, IEnumerable<Document<T>> documents, int topK)

Parameters

query string

The query text used to assess relevance.

documents IEnumerable<Document<T>>

The documents to rerank.

topK int

The number of top-ranked documents to return.

Returns

IEnumerable<Document<T>>

The top-k documents ordered by relevance, with updated relevance scores.

Remarks

This method combines reranking with result limiting. It's more efficient than reranking all documents when only the top results are needed, as some implementations can optimize by not fully scoring lower-ranked candidates.

For Beginners: This re-sorts and then keeps only the best results.

For example:

• Input: 20 documents from initial search
• Rerank: Analyze and re-sort all 20
• Output: Only the top 5 best matches (if topK = 5)

This is useful when you retrieved many documents initially but only need the very best ones after careful analysis.