What is Reranking? What exactly happens during Reranking?

Reranking in Elastic refers to the process of refining and adjusting the initial search results by combining multiple result sets obtained from different search queries or methods, using techniques like Reciprocal Rank Fusion (RRF). The goal of reranking is to improve the relevance and quality of the final ranked list of documents by considering various relevance indicators from multiple searches.

• Reranking is a process that takes initial search results and reorders them to improve their relevance.
• Imagine you have a list of search results. Reranking looks at these results and adjusts their order to better match what you’re looking for.

How Does Reranking Work?

• After the initial search, the system examines the results and applies additional criteria to reorder them.
• For example, it might combine results from multiple searches using a technique like Reciprocal Rank Fusion (RRF), which adjusts the ranking based on how documents are scored across different searches.
• This helps in pushing the most relevant documents to the top of the list.

Benefits of Reranking:

• Refined Results: It fine-tunes the list of results to better meet the user’s needs.
• Higher Quality: By considering multiple relevance signals, it ensures that the most pertinent documents are ranked higher.

Key Points about Reranking in Elastic:

1. Combination of Result Sets: Reranking involves combining result sets from different queries or sub-searches. These could include traditional keyword searches, k-nearest neighbors (kNN) searches, and others.

2. Reciprocal Rank Fusion (RRF): One common method used in reranking is Reciprocal Rank Fusion. RRF combines the rankings of documents from different result sets based on a specific formula to determine their final score.

In reranking, specifically in the context of Elastic, the following steps are typically involved:

Steps in the Reranking Process:

1. Initial Search Execution:

   • Multiple search queries or sub-searches are executed independently. Each query generates its own result set.
   • These queries can be of different types, such as keyword searches, k-nearest neighbors (kNN) searches, or other types of queries supported by Elastic.

2. Result Collection:

   • The result sets from each query are collected. Each result set contains documents ranked according to the relevance criteria specific to that query.
   • For instance, a keyword search might rank documents based on keyword frequency and relevance, while a kNN search might rank them based on vector similarity.

3. Reciprocal Rank Fusion (RRF):

   • The RRF algorithm is applied to combine the individual result sets into a single, reranked result set.
   • The RRF formula is used to calculate a combined score for each document. The formula considers the rank of each document in the individual result sets and adjusts the score accordingly:

     python
     score = 0.0 for q in queries: if d in result(q): score += 1.0 / (k + rank(result(q), d)) return score

     • k is a ranking constant.
     • q is a query from the set of queries.
     • d is a document in the result set of q.
     • result(q) is the result set of q.
     • rank(result(q), d) is the rank of document d in the result set of q, starting from 1.

4. Score Calculation:

   • For each document, the RRF formula calculates a score based on its ranks in the individual result sets.
   • Documents that appear higher in multiple result sets receive higher combined scores, reflecting their overall relevance across different queries.

5. Final Ranking:

   • The documents are sorted based on their combined scores to produce the final ranked list.
   • The final result set presents the most relevant documents at the top, combining the strengths of the different search methods used.

Example:

Let's consider a practical example to illustrate the reranking process using RRF.

Step 1: Initial Searches

• Query 1 (Keyword Search):

  • Search for "blue shoes".
  • Result set: Doc A (rank 1), Doc B (rank 2), Doc C (rank 3).

• Query 2 (kNN Search):

  • Search based on vector similarity for a specific query vector.
  • Result set: Doc B (rank 1), Doc D (rank 2), Doc A (rank 3).

Step 2: Result Collection

• Collect the result sets from both queries.

Step 3: Apply RRF

• Use the RRF formula to combine ranks:
  • score_A = 1/(k + rank_Q1(A)) + 1/(k + rank_Q2(A))
  • score_B = 1/(k + rank_Q1(B)) + 1/(k + rank_Q2(B))
  • score_C = 1/(k + rank_Q1(C))
  • score_D = 1/(k + rank_Q2(D))
  Assuming k = 60 and ranks starting from 1:
  • score_A = 1/(60+1) + 1/(60+3) = 0.0161 + 0.0159 = 0.0320
  • score_B = 1/(60+2) + 1/(60+1) = 0.0161 + 0.0161 = 0.0322
  • score_C = 1/(60+3) = 0.0159
  • score_D = 1/(60+2) = 0.0161

Step 4: Final Ranking

• Sort documents based on combined scores:
  • Doc B (score 0.0322)
  • Doc A (score 0.0320)
  • Doc D (score 0.0161)
  • Doc C (score 0.0159)

The final ranked list reflects the overall relevance considering both keyword and kNN search results.

Key Points:

• Combining Multiple Queries: Reranking leverages multiple search methodologies to improve the relevance of search results.
• Reciprocal Rank Fusion: The RRF algorithm is central to reranking, providing a systematic way to combine and score results from different queries.
• Final Relevant Results: The process yields a final ranked list that is more relevant and accurate compared to any single search method used alone.

Reranking enhances search quality by integrating diverse relevance signals, thereby delivering better search experiences for users.