Combining Keyword and Semantic Search

Why Hybrid?

Pure semantic search misses exact matches (product IDs, names). Pure keyword search misses conceptual similarity. Hybrid combines both.

Architecture

User Query
  ├── Keyword Search (BM25) → Results A
  └── Semantic Search (Vector) → Results B
       ↓
  Reciprocal Rank Fusion (RRF)
       ↓
  Combined Results

Reciprocal Rank Fusion

def rrf(results_lists, k=60):
    scores = {}
    for results in results_lists:
        for rank, doc in enumerate(results):
            if doc.id not in scores:
                scores[doc.id] = 0
            scores[doc.id] += 1 / (k + rank + 1)
    return sorted(scores.items(), key=lambda x: -x[1])

Implementation Options

• Elasticsearch + Vector Plugin: BM25 + kNN in one system
• PostgreSQL + pgvector: Full-text search + vector similarity
• Separate systems: Dedicated keyword + vector DBs, merge results
• Weaviate/Qdrant: Built-in hybrid search

Tuning Hybrid Weight

Balance between keyword and semantic results:

• alpha=0: Pure keyword search
• alpha=0.5: Equal weight
• alpha=1: Pure semantic search
• Test different values on your evaluation dataset

Common Pitfall: Implementing Hybrid Search Without Normalizing Scores

BM25 scores and cosine similarity scores are on completely different scales. BM25 produces scores that can range from 0 to 20+, while cosine similarity is bounded between 0 and 1. Merging them without normalization gives BM25 an unfair weight advantage. Normalize both to [0, 1] before applying weights:

def normalize_scores(results: list) -> list:
    if not results:
        return results
    max_score = max(r["score"] for r in results)
    min_score = min(r["score"] for r in results)
    span = max_score - min_score or 1.0
    for r in results:
        r["score"] = (r["score"] - min_score) / span
    return results

keyword_results = normalize_scores(bm25_results)
semantic_results = normalize_scores(vector_results)
# Now safe to merge with equal weights

Best Practice: Use Weaviate or Qdrant for Built-In Hybrid Search

If you are building a new search system, choose a vector database with native hybrid search rather than maintaining two separate systems. Weaviate provides BM25 + vector search with a single query. Qdrant supports sparse + dense vector hybrid search. Both eliminate the need to run and sync a separate keyword search engine like Elasticsearch.

Named Product Reference: pgvector for PostgreSQL Hybrid Search

For teams already running PostgreSQL, the pgvector extension adds vector similarity search directly to the database. Combined with PostgreSQL's built-in full-text search (tsvector/tsquery), you get hybrid search in a single database without any additional infrastructure. This is the lowest-complexity hybrid search solution for teams with existing Postgres infrastructure.