How do I combine embeddings with keyword search?

A common way to combine embeddings with keyword search is to use a hybrid search pipeline:

1) Run both searches

• Keyword search: good for exact matches, IDs, names, rare terms, filters.
• Embedding/vector search: good for semantic similarity and paraphrases.

2) Merge the results

Typical approaches:

A. Weighted score fusion

Give each document:

• a keyword score
• a vector similarity score

Then combine them:

[ \text{final_score} = \alpha \cdot \text{vector_score} + (1-\alpha) \cdot \text{keyword_score} ]

Choose α based on your use case:

• higher α if meaning matters more
• lower α if exact term matching matters more

B. Reciprocal Rank Fusion (RRF)

If your keyword and vector systems produce ranked lists, combine ranks instead of raw scores:

[ \text{RRF}(d) = \sum_i \frac{1}{k + rank_i(d)} ]

This is popular because it’s robust even when score scales differ.

C. Two-stage retrieval

1. Use keyword search to get candidates
2. Re-rank candidates with embeddings

Or the reverse:

1. Vector search for candidates
2. Boost exact keyword matches in re-ranking

3) Use metadata filters

Before or during retrieval, filter by:

• date
• category
• language
• permissions
• product type

This keeps search relevant and efficient.

4) Practical best practices

• Normalize scores before blending; keyword and vector scores often live on different scales.
• Use keyword search for:
  • names
  • codes
  • error messages
  • URLs
  • exact phrases
• Use embeddings for:
  • paraphrases
  • conceptual matching
  • natural language queries
• Consider stemming, synonyms, and phrase queries in keyword search.

5) Simple implementation pattern

1. Build an inverted index for keyword search.
2. Store embeddings in a vector database or ANN index.
3. For each query:
   • search both indexes
   • retrieve top N from each
   • merge with RRF or weighted scoring
   • optionally rerank with a cross-encoder/LLM

Example

Query: “reset my account password”

• Keyword search finds documents with “reset” and “password”
• Embedding search also finds “forgot credentials” or “change login details”
• Hybrid search returns both exact and semantically similar results

If you want, I can also show:

• a Python example
• an example using Elasticsearch/OpenSearch
• or a vector DB + BM25 hybrid architecture