Hybrid Search

What Is Hybrid Search?

Hybrid search combines lexical search, which looks for matching words or terms, with semantic search, which uses embeddings to find related meaning.

Why Combine Two Search Methods?

Hybrid search keeps useful signals from both retrieval approaches.

How Hybrid Search Works

The final results can then be filtered, reranked, or limited with Top-K.

Lexical Search

Common lexical approaches include inverted indexes and relevance algorithms such as BM25.

Semantic Vector Search

Semantic retrieval converts queries and chunks into embeddings and searches for vectors that are close under a selected metric.

Combining Search Scores

One simple approach is to normalize comparable signals and combine them with weights.

keyword_score = 0.80
semantic_score = 0.70

keyword_weight = 0.45
semantic_weight = 0.55

hybrid_score = (
    keyword_weight * keyword_score
    + semantic_weight * semantic_score
)

print(round(hybrid_score, 3))

This is a teaching example. Production systems must account for the different score scales and meanings of their retrievers.

Why Score Normalization Matters

Lexical and vector systems may produce scores with different ranges. Adding raw scores directly can make one signal dominate.

Rank Fusion with RRF

Reciprocal Rank Fusion (RRF) combines ranked lists without requiring raw scores to use the same scale.

def rrf_score(rank, k=60):
    return 1 / (k + rank)

keyword_rank = 2
semantic_rank = 4

score = rrf_score(keyword_rank) + rrf_score(semantic_rank)

print(round(score, 5))

A document appearing high in both lists receives contributions from both rankings.

Hybrid Search + Top-K

Hybrid retrieval can merge candidates, rank them, and keep the final Top-K results.

Hybrid Search + Metadata Filtering

A support system can restrict results to the correct product and region while using both retrieval signals.

Simple Python Candidate Fusion

keyword_results = ["A", "B", "C"]
semantic_results = ["C", "A", "D"]

combined = []

for item in keyword_results + semantic_results:
    if item not in combined:
        combined.append(item)

print(combined)

A production implementation should use a formal fusion method and preserve ranks or scores for evaluation.

Hybrid Search in RAG

Hybrid retrieval is useful for enterprise content where questions mix natural language with exact terms, codes, names, or identifiers.

When Hybrid Search Works Well

When Hybrid Search May Not Be Necessary

Hybrid retrieval adds complexity, so justify it with measured improvement.

Common Hybrid Search Mistakes

Evaluate Hybrid Search