Skip to main content
Hybrid search combines results from multiple search queries into a single, ranked result set using fusion algorithms. Actian VectorAI DB provides built-in fusion methods that merge results from different query vectors, embedding models, or search strategies to improve retrieval quality beyond what a single search achieves. Hybrid search works by running multiple independent searches and then fusing their results into a unified ranking. Each search can use a different query vector, embedding model, or search strategy. The fusion algorithm resolves differences in scoring and ranking to produce a final result set that captures relevance signals from all sources. Common use cases include:
  • Combining results from multiple embedding models
  • Merging semantic and keyword search results
  • Query expansion with reformulated questions
  • Multi-perspective retrieval for RAG applications
  • Improving robustness through query variations

​
Fusion methods

VectorAI DB supports two fusion algorithms for combining search results:

​
Reciprocal Rank Fusion

RRF scores each result based on its rank position across multiple result lists rather than raw similarity scores. This makes RRF effective when combining results from different models or search types where scores are not directly comparable. The ranking_constant_k parameter controls how much weight is given to higher-ranked results. The default value of 60 works well for most use cases. Lower values emphasize top-ranked results more heavily, while higher values distribute weight more evenly across ranks.

​
Distribution-Based Score Fusion

DBSF normalizes scores based on the statistical distribution of each result set before combining them. This approach produces more balanced rankings when the score distributions from different searches have different means and variances. DBSF is particularly effective when combining results from searches that use different vector distributions or when one search consistently produces higher scores than another.

​
Hybrid search patterns

Several patterns leverage hybrid search for improved retrieval:

​
Multi-model ensemble

Run the same query through multiple embedding models and fuse the results. Different models capture different aspects of semantic meaning, and fusion combines their strengths.

​
Query variation

Generate slight perturbations of a query vector and search with each variation. Fusing the results improves robustness by reducing sensitivity to the exact query representation.

​
Hybrid RAG retrieval

Combine searches from an original question embedding and a reformulated question embedding. This retrieves relevant context that a single query might miss, improving RAG pipeline quality.

​
Performance considerations

Hybrid search runs multiple queries, so latency increases roughly linearly with the number of searches. The fusion step itself adds minimal overhead. Consider these tradeoffs:
  • More searches improve retrieval quality but increase latency
  • Fusion algorithms operate on result lists in memory and are fast relative to search
  • Use limit on individual searches to control the candidate pool size
  • Benchmark single versus hybrid search to find the right balance for your use case

​
Task guides

See step-by-step guides for hybrid search operations:

​
Next steps

Explore related topics to deepen your understanding of search in VectorAI DB:
  • Search — Understand the underlying vector search operations.
  • Filtering — Apply metadata conditions to hybrid search queries.
  • Distance metrics — Understand how similarity scores are calculated.
  • Points — Manage the data being searched.