What is Refining and Adjusting in Reranking?

In the context of reranking in Elastic, "refining" and "adjusting" refer to improving the initial search results to better meet the user's needs. Let's break down these terms:

Refining

Refining means enhancing the quality and accuracy of search results. This involves:

1. Improving Relevance:

   • By considering multiple relevance indicators, the system can better determine which documents are most pertinent to the query.
   • For example, combining keyword relevance with vector similarity can provide a more nuanced understanding of which documents are truly relevant.

2. Removing Noise:

   • Initial search results might contain documents that are not very relevant. Refining helps to filter out less relevant results.
   • This process helps in focusing on the most useful documents, enhancing the overall quality of the search results.

3. Combining Strengths:

   • Each search method has its own strengths. By refining, you can combine the best aspects of different search methodologies.
   • For instance, keyword searches might excel at matching specific terms, while vector searches might excel at finding similar concepts. Refining leverages both to improve the final result.

Adjusting

Adjusting involves modifying the initial ranking of documents based on additional criteria or methods. This involves:

1. Reordering Results:

   • Initial search results are ranked based on their individual relevance scores. Adjusting reorders these results to better reflect overall relevance.
   • For example, a document that ranks highly in multiple queries will be adjusted to appear higher in the final results list.

2. Applying Fusion Techniques:

   • Techniques like Reciprocal Rank Fusion (RRF) are used to adjust scores from different queries and combine them into a single ranking.
   • RRF calculates a new score for each document based on its ranks in the individual result sets, then adjusts the document’s position in the final list accordingly.

3. Balancing Relevance Signals:

   • Different queries might have different relevance signals. Adjusting balances these signals to create a unified ranking.
   • For example, a document might not rank very high in one query but might rank very high in another. Adjusting ensures that such a document’s combined relevance is accurately reflected in the final ranking.

Example

Imagine you search for "best smartphone" and get two sets of results:

• Query 1: A keyword search that finds documents mentioning "best smartphone".
• Query 2: A vector search that finds documents similar to a detailed review of a "top-rated smartphone".

Initial Results:

• Query 1: Doc A (rank 1), Doc B (rank 2), Doc C (rank 3)
• Query 2: Doc D (rank 1), Doc B (rank 2), Doc E (rank 3)

Refining:

• Identify that Doc B is relevant in both queries and should be given more importance.
• Remove less relevant documents from the initial sets.

Adjusting:

• Use RRF to combine the ranks and recalculate scores:
  • Doc B: 1/(60+2) + 1/(60+2) = higher combined score
  • Doc A: 1/(60+1) = moderate score
  • Doc D: 1/(60+1) = moderate score
  • Doc C: 1/(60+3) = lower score
  • Doc E: 1/(60+3) = lower score

Final Results:

• Doc B (most relevant, appears in both queries)
• Doc A (high keyword relevance)
• Doc D (high vector similarity)
• Other documents follow based on their adjusted scores.

In summary, refining enhances the quality of results by combining and filtering based on multiple criteria, while adjusting modifies the ranking to reflect a balanced view of overall relevance. Together, these processes aim to deliver the most pertinent and useful search results.