Explain hybrid search: why combine vector search and BM25? What is RRF (Reciprocal Rank Fusion)?

Hybrid search combines the advantages of two approaches: vector search and traditional BM25 text search.

• Vector search works well with semantics and meaning, allowing to find relevant documents by sense, even if there are no exact word matches.

• BM25 is a classic ranking algorithm based on word frequencies and document length, suitable for exact keyword matching.

By combining them, you can achieve a more quality search that considers both semantics and exact matches.

RRF (Reciprocal Rank Fusion) is a method of combining results from multiple search systems or models. It assigns each result position a weight inversely proportional to its rank and sums these weights for the final ranking.

The RRF formula for document d:

[ score(d) = \sum_{i=1}^n \frac{1}{k + rank_i(d)} ]

where:

• ( n ) — number of systems,
• ( rank_i(d) ) — position of the document in the i-th system,
• ( k ) — a smoothing constant (usually 60).

Thus, RRF effectively combines vector search and BM25 results, improving the quality of the final ranking.