A hybrid query combines vector search with lexical search to achieve better relevance by leveraging both semantic similarity and keyword matching. Full-text search and vector search use fundamentally different scoring mechanisms - text search typically uses relevance scores based on term frequency and document frequency, while vector search uses similarity distances. Without proper normalization, one search method may dominate the results, leading to suboptimal ranking. Hybrid queries address this by combining and normalizing scores from both methods. LambdaDB supports three rescoring methods to combine results from different query types: rrf (Reciprocal Rank Fusion), mm (min-max), and l2 (l2_norm).
  • RRF combines rankings by taking the reciprocal of each result’s rank position, providing balanced weighting across different search methods.
  • MinMax normalization scales scores to a 0-1 range before combining.
  • L2 norm uses Euclidean distance-based normalization to merge relevance scores from multiple query sources.
Regardless of the rescoring method used, the final combined score is always normalized to a value between 0 and 1.

​
Parameters

​
Rescoring methods

MethodDescriptionBoost support
rrfReciprocal Rank Fusion - balanced ranking combinationNo
mmMin-Max normalization scalingYes
l2Euclidean distance-based normalizationYes

​
Query object parameters

Each query object within the rescoring method array can contain:
ParameterDescriptionTypeRequiredDefault
queryQuery object (queryString/knn/sparseVector/bool)object✓
boostScore multiplier for relevancefloat0.5
A hybrid query can include up to two query objects. If you need to express more complex logic within a query object, use a boolean query to combine multiple conditions.

​
Boost constraints

  • The boost parameter is only available for mm and l2 rescoring methods.
  • The sum of all boost values must equal to 1.0.
  • Each individual boost value must be between 0 and 1.

​
Examples

​
L2-norm hybrid query

{
  "l2": [
    {
      "queryString": {
        "query": "content:LambdaDB"
      },
      "boost": 0.7
    },
    {
      "knn": {
        "filter": {
          "queryString": {
            "query": '"https://lambdadb.ai"',
            "defaultField": "metadata.url"
          }
        },
        "field": "text_embedding",
        "queryVector": [0.1, 0.2, 0.3],
        "k": 5
      },
      "boost": 0.3
    }
  ]
}
This example combines:
  • A queryString query with 0.7 boost weight.
  • A knn vector query with pre-filtering and 0.3 boost weight.
  • Uses L2-norm rescoring to merge the results.

​
MinMax hybrid query with sparse vector

{
  "mm": [
    {
      "sparseVector": {
        "field": "sparse_embedding",
        "queryVector": {
          "machine": 0.8,
          "learning": 0.6,
          "AI": 0.4
        }
      },
      "boost": 0.6
    },
    {
      "knn": {
        "field": "dense_embedding",
        "queryVector": [0.1, 0.2, 0.3],
        "k": 10
      },
      "boost": 0.4
    }
  ]
}
This example combines:
  • A sparseVector query with 0.6 boost weight.
  • A knn dense vector query with 0.4 boost weight.
  • Uses MinMax normalization to merge the results.

​
RRF hybrid query

{
  "rrf": [
    {
      "queryString": {
        "query": "machine learning",
        "defaultField": "content"
      }
    },
    {
      "knn": {
        "field": "content_embedding",
        "queryVector": [0.1, 0.2, 0.3],
        "k": 10
      }
    }
  ]
}
This example uses Reciprocal Rank Fusion to balance text and vector search results.
The final returned documents may not include the requested number of documents from the knn query. This is because the scores of documents returned solely from other queries may be higher than those of the top k documents returned from the knn.