Reranking

Reranking is an information retrieval process in which an already retrieved set of candidate results is scored and reordered using an additional ranking method so that the most relevant results appear nearer the top. It is widely used in web search, enterprise search, ecommerce, recommendation systems, and Retrieval-Augmented Generation (RAG) applications.

Reranking solves an important computational problem. Searching a collection containing millions or billions of items with the most powerful available model would often be too slow and expensive. Instead, a faster retrieval system first identifies a smaller group of potentially relevant candidates. A more precise reranker can then spend additional computing resources evaluating only that reduced set.

The simplified process is:

The initial retrieval method does not have to be simple keyword matching. Modern systems can retrieve candidates through lexical search, vector search, semantic retrieval, or hybrid search before reranking them. Google Cloud, for example, documents a system in which retrieved and combined candidates are subsequently rescored with a dedicated semantic reranker.

Why Advanced Search Systems Use Reranking

Reranking allows large search systems to balance speed, computing cost, and relevance rather than relying on one model to perform every stage of the search.

Fast retrieval methods can search enormous collections efficiently, while more computationally expensive models are reserved for a much smaller candidate pool.

Two-Stage Information Retrieval

A common search architecture can be simplified into two stages.

Stage 1 – Retrieval: A fast system searches a large collection and returns a manageable candidate set. Retrieval may use lexical methods such as BM25, dense vector search, bi-encoders, or a combination of techniques.

Stage 2 – Reranking: A more precise model evaluates those candidates and assigns new relevance scores, changing their order before the final results are returned.

Candidate pools vary considerably between systems. There is no universal process in which exactly 1,000 results are always reduced to 10. Google Cloud’s current VertexRanker, for example, can receive anywhere from 1 to 1,000 candidates for semantic reranking.

Balance Between Speed and Computing Cost

Keyword-based retrieval is relatively inexpensive because it can match terms across huge indexes very quickly. More advanced semantic and neural models are computationally expensive, especially if they must compare a query against millions or billions of documents. Reranking therefore combines the speed of first-stage keyword or other lightweight retrieval with the accuracy of a more sophisticated second-stage model.

Reranking can improve the relevance of top results by applying a more detailed model to the smaller candidate set and jointly evaluating the query against each candidate. This makes advanced neural models and LLM-based relevance scoring practical because they are applied only to selected candidates rather than the entire collection.

How Reranking Works

The exact process depends on the search system. One common neural-search design combines a fast bi-encoder during retrieval with a more precise cross-encoder during reranking.

The Retrieval Phase

A bi-encoder processes the query and documents separately and converts them into numerical representations called embeddings. Document embeddings can be calculated and stored before a search occurs.

When a query arrives, the system creates its embedding and compares it with stored document embeddings to find nearby matches. Because the documents do not need to be processed from scratch for every query, this approach can search large collections efficiently.

Bi-encoders are only one retrieval option. A system can instead use traditional keyword retrieval or combine lexical and semantic results. Sentence Transformers, for example, documents both lexical search and bi-encoder semantic search as possible first stages before cross-encoder reranking.

The Reranking Phase

A cross-encoder processes the query and each candidate document together rather than representing them independently. This allows the model to examine the relationship between the query and candidate more closely and produce a relevance score.

The additional analysis can improve ranking accuracy, but it also requires more computation. Cross-encoders are therefore normally applied only to a limited number of candidates rather than the entire collection.

Some systems also use large language models as rerankers. Instead of relying only on similarity scores, an LLM can evaluate a smaller set of retrieved passages against the user’s actual question and reorder them according to relevance.

Reranking in Google Search and Modern Search Systems

Google Search uses many ranking systems rather than one simple retrieval-and-reranking model. Google does not publicly document its complete internal ranking pipeline, so individual technologies should not be arranged into an assumed sequence of fixed L1, L2, or L3 reranking stages.

Google has, however, publicly described several systems relevant to retrieval and ranking.

• Neural Matching: Google describes neural matching as an AI system that understands representations of concepts in queries and pages and matches them together.
• RankBrain: RankBrain helps Google understand relationships between words and concepts so that relevant content can be returned even when it does not contain every exact word in the query. Google’s earlier explanation also describes RankBrain as helping determine the order of search results.
• BERT: Google’s ranking documentation says BERT helps understand how combinations of words express different meanings and intent.
• DeepRank and NavBoost: Documents released during the U.S. antitrust proceedings against Google provide additional information about internal ranking technologies. The filings identify DeepRank and RankBrain as important deep-learning ranking systems and describe NavBoost as an influential ranking component using user-interaction data (pp. 65–66).

These technologies should not be interpreted as evidence of a simple fixed cascade in which one always reranks the output of another.

Google Cloud provides a particularly clear real-world example: its Agent Retrieval system can combine candidates from several searches and then use VertexRanker to rescore the merged results against the natural-language query.

Where Reranking Is Used

Reranking extends well beyond conventional web search.

• Search Engines and Enterprise Search: A fast first stage can retrieve potentially relevant webpages, files, emails, or knowledge-base documents before a more precise system determines their final order.
• RAG and AI Applications: A RAG system may retrieve passages through keyword, vector, or hybrid search and then rerank them before sending the strongest passages to a language model. This can reduce irrelevant context and improve answer quality, although reranking does not guarantee that an LLM will never hallucinate.
• Ecommerce Search: A product-search engine can retrieve products related to the query and apply additional ranking models to determine which products are most relevant. Reranking may consider semantic relevance alongside other permitted product or business signals.
• Recommendation Systems: Large recommendation platforms often use a similar candidate-generation-and-ranking architecture. A 2016 Google research paper on YouTube recommendations describes one neural network for candidate generation and a separate model for ranking the resulting candidates.

• Jobs and Matching Systems: Search and recommendation systems can use retrieval and ranking methods to match job listings with candidates or résumés with relevant vacancies. However, not every applicant tracking system uses AI reranking, and automated hiring systems require additional fairness and compliance considerations.

Similar ranking approaches also appear outside conventional search systems:

• Computer Vision: Some image-retrieval and object-detection systems use reranking to refine an initial set of visual matches or detections, although it is not a universal stage in computer-vision pipelines.
• Logistics: Amazon has used learning-to-rank methods to identify suitable package-delivery locations from noisy location data.
• Robotics: Reranking can be used to score and reorder candidate robot grasps so that more promising options are attempted first.
• Drug Discovery: Computational screening systems can first narrow very large molecular collections and then apply more precise scoring or ranking to the most promising compounds.

Frequently Asked Questions

What is the difference between vector search with a bi-encoder and reranking with a cross-encoder?

A bi-encoder usually represents the query and documents separately, making it efficient for searching large collections through vector similarity. A cross-encoder evaluates the query and candidate together, allowing more detailed relevance analysis but requiring considerably more computation. This is why cross-encoders are commonly applied after retrieval rather than to every document.

Does adding a reranker make a search system faster or slower?

The reranking stage itself adds processing time because candidates must be evaluated again. However, the two-stage architecture makes sophisticated ranking practical because the expensive model processes only a small candidate set instead of the entire collection.

What technologies and algorithms does Google use for reranking?

Google publicly documents systems including RankBrain, BERT, and neural matching, while antitrust court materials have revealed additional information about technologies such as DeepRank and NavBoost. Google’s complete retrieval and ranking architecture is not publicly documented, so these systems should not be assigned to a precise reranking sequence unless Google has documented that relationship.

How to know whether an enterprise search application or AI app needs a reranker?

A reranker may be useful when the retrieval system regularly finds the correct documents but places them too low, vector search returns broadly related but imprecise matches, or only a small number of retrieved passages can be provided to an LLM. The benefit should be tested against the additional latency, computing cost, and implementation complexity introduced by reranking.