vector search with keyword hybrid

A hybrid search system combines:

• Keyword search: matches exact terms, stems, filters, phrases
• Vector search: finds semantically similar documents using embeddings

This usually gives better results than either alone.

Typical architecture

1. Index documents twice

   • Lexical index for keywords (e.g. BM25 / inverted index)
   • Vector index for embeddings (e.g. FAISS, Milvus, Pinecone, pgvector, Elasticsearch kNN)

2. At query time

   • Convert query to embedding
   • Run keyword retrieval and vector retrieval in parallel
   • Combine/rerank results

Common fusion strategies

1) Weighted score fusion

Normalize scores from both systems and combine:

[ score = \alpha \cdot score_{vector} + (1 - \alpha) \cdot score_{keyword} ]

Good when you can calibrate scores.

2) Reciprocal Rank Fusion (RRF)

Very common because it’s simple and robust:

[ RRF(d) = \sum_{i \in systems} \frac{1}{k + rank_i(d)} ]

Usually works well even if the score scales differ.

3) Two-stage retrieval + rerank

• Retrieve top N from keyword and top N from vector search
• Merge candidates
• Rerank with a cross-encoder / LLM / learning-to-rank model

This often gives the best quality.

When hybrid helps

• Queries with exact entities: names, IDs, SKUs, error codes
• Queries with vague intent: “best way to reduce latency”
• Long-tail vocabulary or synonyms
• Sparse metadata and semantic relevance together

Example

Query: apple battery replacement cost

• Keyword search catches docs with “battery replacement” and “cost”
• Vector search catches docs about “repair pricing” or “service fees” even if wording differs
• Hybrid combines both, improving recall and precision

Implementation options

Elasticsearch / OpenSearch

• BM25 + dense vector kNN
• Use function score, RRF, or reranking

PostgreSQL

• tsvector / full-text search + pgvector
• Combine with SQL ranking logic

Dedicated vector DB + search engine

• Example: Elasticsearch for keyword, Pinecone/Milvus for vector
• Merge results in application layer

Practical tips

• Use filters first (tenant, permissions, date)
• Normalize scores before weighted fusion
• Evaluate with Recall@K, nDCG@K, MRR
• Tune retrieval sizes separately for keyword and vector results
• Cache embeddings for frequent queries

If you want, I can also provide:

1. a Python example of hybrid search,
2. an Elasticsearch/OpenSearch query, or
3. a pgvector + full-text search example.