Hybrid Search Concepts

1. Why Choose?

We saw that Dense Search (Vectors) and Sparse Search (Keywords) have opposite strengths. Hybrid Search runs both and combines the results.

2. Theoretical Workflow

1. User Query: "Error 501 on login page."
2. Dense Retriever: Finds docs about "Login Issues" (Concepts).
3. Sparse Retriever: Finds docs containing "Error 501" (Exact Match).
4. Ensemble: Combine the two lists.

3. Reciprocal Rank Fusion (RRF)

How do you combine the lists? You can't just add the scores (Cosine Similarity score 0.8 is not comparable to BM25 score 15.0). We use Rank Fusion.

We look at the rank, not the score.

• Doc A: Rank 1 in Dense, Rank 5 in Sparse.
• Doc B: Rank 10 in Dense, Rank 1 in Sparse.

RRF gives points for being high on either list.

from langchain.retrievers import EnsembleRetriever

ensemble_retriever = EnsembleRetriever(
  retrievers=[bm25_retriever, vector_retriever],
  weights=[0.5, 0.5]
)

docs = ensemble_retriever.get_relevant_documents("query")

4. Summary

Hybrid Search is the current state-of-the-art for production RAG. It provides the "Safety" of keywords with the "Magic" of vectors.

Key Intuition: "Don't put all your eggs in one index."