Hybrid BM25 and vector search ranking methods

Hybrid ranking merges exact keyword matches from BM25 with semantic similarity scores from vector queries to order retrieved documents. In parallel pipelines, each retrieval system independently scores candidates before combining outputs through fusion techniques such as Reciprocal Rank Fusion or Relative Score Fusion. This dual approach balances strict lexical constraints with conceptual relevance while mitigating individual retrieval blind spots.

Architectural mechanisms of hybrid retrieval

Lexical search retrieves documents by matching query terms against an inverted index that maps terms to the documents containing them ^[1][2]. In engines such as Apache Lucene and MongoDB Atlas Search, BM25 serves as the default scoring algorithm for keyword relevance ^[3]. BM25 scores documents based on term frequency saturation, which prevents keyword-stuffed documents from dominating by ensuring each additional occurrence has diminishing impact ^[4]. In addition, BM25 applies document length normalization relative to the average collection document length so a concise answer can outrank a lengthy text ^[5]. It also incorporates inverse document frequency, giving higher weight to rare terms than common terms ^[6]. Conversely, dense vector search represents text as numerical vectors to evaluate semantic similarity using metrics such as cosine similarity, Euclidean distance, or dot product ^[7][8].

Relying exclusively on one retrieval approach exposes blind spots, as full-text search misses relevant documents when wording changes, while vector search can be unreliable for exact identifiers ^[9]. Standard hybrid retrieval pipelines address these issues by executing a multi-stage process where metadata filters run first, full-text and vector retrievers run in parallel to generate two separate ranked lists with scores, and the outputs are merged and reranked ^[10][11]. Alternatively, full-text search can be used to narrow candidate sets by keyword, phrase, or metadata fields before running vector search inside that smaller set to rank by meaning ^[12].

Score fusion and ranking algorithms

When parallel retrieval executes, hybrid systems generate separate ranked lists with scores from keyword search and vector search ^[11]. In MongoDB Atlas, reciprocal rank fusion and relative score fusion are recommended techniques for combining BM25 and vector search scores into a unified ranking ^[8][13]. Relative score fusion normalizes scores from each search method, scaling them to a common range, typically 0 to 1, before combining them ^[14]. This normalization ensures that the relative importance of each search method is preserved even when original score distributions differ ^[14].

Reciprocal rank fusion calculates the reciprocal rank of each document across different search methods and combines these ranks into a unified score for each document ^[15]. An RRF algorithm calculates the reciprocal rank score as 1/(rank + k), where rank is the position of the document in the list and k is a constant that balances the influence of individual rankings and decides sensitivity to rank positions ^[16][17]. Experimental results demonstrate that RRF performs best when k is set to a small value, such as 60 ^[17]. The engine then sums the reciprocal rank scores obtained from each strategy to produce a single combined score, reranking documents so those with higher scores appear at the top of the final list ^[18]. In Azure AI Search, RRF merges and homogenizes rankings into a single result set whenever two or more queries execute in parallel, such as hybrid queries combining full-text and vector queries ^[19].

In Azure AI Search, full-text search scored with BM25 has no upper limit ^[20]. Vector search scored with HNSW yields score ranges of 0.333 to 1.00 for cosine similarity, and 0 to 1 for Euclidean and dot product metrics ^[7]. For hybrid search using RRF, the upper limit is bounded by the number of queries being fused, with each query contributing a maximum of approximately 1/k to the score ^[21]. Systems can assign different weights to search strategies or vector queries before combining scores to increase or decrease their importance ^[22][23].

Multilingual retrieval and secondary reranking

Hybrid retrieval workflows often combine scores and follow with reranking ^[11]. In Azure AI Search, semantic ranking can occur after RRF merging of results, returning a separate reranker score between 0.00 and 4.00 in the query response ^[24].

Research investigating embedding APIs on standard retrieval benchmarks, BEIR and MIRACL, demonstrates that ranking performance depends on target languages ^[25][26][27]. On English retrieval, reranking BM25 results using embedding APIs provides a budget-friendly approach that is most effective compared to employing embeddings as first-stage retrievers ^[25]. For non-English retrieval, reranking improves results, but a hybrid model combined with BM25 achieves the best performance, albeit at a higher cost ^[26].

Key facts

• BM25 calculates relevance using term frequency saturation, document length normalization, and inverse document frequency ^[4][5][6].
• Standard hybrid retrieval pipelines run metadata filters first, execute full-text and vector retrievers in parallel, and merge and rerank results ^[10][11].
• Full-text search misses relevant documents when wording changes, while vector search can be unreliable for exact identifiers ^[9].
• Reciprocal rank fusion calculates scores using 1/(rank + k), where rank is document position and k is a constant that performs best when set to a small value such as 60 ^[16][17].
• In RRF, reciprocal rank scores obtained from each strategy are summed to create a single score, placing higher-scoring documents at the top of the final ranking ^[18].
• Relative score fusion normalizes scores from each method to a common range, typically 0 to 1, preserving relative importance across different score distributions ^[14].
• In Azure AI Search, BM25 scores have no upper limit, vector similarity scores range from 0.333 to 1.00 for cosine and 0 to 1 for Euclidean and dot product, and RRF scores are bounded by the number of queries fused with each contributing at most approximately 1/k ^[20][7][21].
• Different retrieval strategies and vector queries can be weighted to adjust their influence before combining scores ^[22][23].
• Semantic ranking can take place after RRF merging in Azure AI Search, producing a separate score between 0.00 and 4.00 ^[24].
• Benchmark evaluations on BEIR and MIRACL show that reranking BM25 results is most effective for English, while a hybrid model with BM25 performs best for non-English retrieval ^[25][26][27].