Hybrid Search Explained

Question

What is hybrid search and why use it instead of just vector search?

Explanation

Hybrid search = running BM25 (keyword) and vector (semantic) search at the same time, then merging the results.

How it works in Weaviate

1. You send a query (text + vector)
2. Weaviate runs two searches in parallel: BM25 scores chunks by keyword matching, vector scores chunks by how close their meaning is to the query (cosine similarity)
3. Results are merged using a parameter called alpha

The alpha parameter

Controls the balance between the two:

• alpha = 0.0 - 100% BM25 (pure keyword)
• alpha = 0.5 - 50/50 (balanced - what we use)
• alpha = 1.0 - 100% vector (pure semantic)

Why 0.5 works well

• Exact term queries ("BM25 algorithm") - BM25 finds it precisely, vector might return vaguely related results
• Vague queries ("how to find relevant text") - vector understands the meaning, BM25 misses because no exact match
• Acronyms ("RAG") - BM25 finds the acronym, vector also finds "retrieval-augmented generation"

With alpha=0.5, you get the best of both worlds.

When to change alpha

• Your users search by exact names/codes - lower alpha (more BM25)
• Your users ask natural language questions - higher alpha (more vector)
• You're not sure - 0.5 is a safe default

Example

results = store.similarity_search(
    query,
    k=20,          # return top 20 candidates
    alpha=0.5,     # balanced hybrid
)

LangChain + Weaviate handles the merge automatically.