Vector Hybrid Search

Vector search understands meaning. Keyword search matches exact words. Each method fails in situations where the other succeeds. Hybrid search runs both methods and merges their results into one ranked list.

Where Each Method Struggles

CASE 1: Exact codes
  Query: "error E-4502"
  Keyword search: finds the page with "E-4502"      (strong)
  Vector search:  returns pages about errors        (weak)

CASE 2: Meaning without shared words
  Query: "how to make my laptop run faster"
  Keyword search: misses a page titled "Improve PC performance"  (weak)
  Vector search:  finds it by meaning                            (strong)

Product codes, names, abbreviations, and rare terms favor keyword search. Natural questions and paraphrases favor vector search. Real users type both kinds of queries.

Keyword Search in Brief

Keyword search counts how often query words appear in a document and how rare those words are across all documents. The most common scoring formula is BM25. A rare word such as "E-4502" earns a high score because few documents contain it.

The Hybrid Pipeline

HYBRID SEARCH FLOW

            User query
           /          \
          v            v
  +-------------+  +-------------+
  | Keyword     |  | Vector      |
  | search      |  | search      |
  | (BM25)      |  | (embeddings)|
  +-------------+  +-------------+
          \            /
           v          v
       +----------------+
       | Fusion step    |
       | merges lists   |
       +----------------+
               |
               v
        Final ranked results

Merging Two Result Lists

The two methods produce scores on different scales, so you cannot add them directly. Two common solutions exist.

Reciprocal Rank Fusion

Reciprocal Rank Fusion, or RRF, ignores raw scores and uses only rank positions. Each document earns points based on its rank in each list.

RRF score = 1 / (k + rank)      with k = 60

Document X: rank 1 in keyword list, rank 3 in vector list
  = 1/(60+1) + 1/(60+3)
  = 0.0164 + 0.0159
  = 0.0323

Document Y: rank 2 in keyword list, rank 10 in vector list
  = 1/(60+2) + 1/(60+10)
  = 0.0161 + 0.0143
  = 0.0304

Document X ranks above Document Y

Weighted Score Fusion

This method normalizes both score sets to a common range, then applies a weight.

final = (alpha x vector_score) + ((1 - alpha) x keyword_score)

alpha = 0.7  -> favors meaning
alpha = 0.3  -> favors exact words

Choosing the Balance

Worked Example

Query: "reset router model RX-200"

Keyword results:   1. RX-200 Setup Guide
                   2. RX-200 Reset Steps
                   3. RX-100 Reset Steps

Vector results:    1. How to restore factory settings on a router
                   2. RX-200 Reset Steps
                   3. Fixing connection problems

After fusion:      1. RX-200 Reset Steps       (strong in both)
                   2. RX-200 Setup Guide
                   3. How to restore factory settings on a router

The document that scores well in both lists rises to the top. That agreement gives hybrid search its reliability.

Implementation Tips

1. Check whether your database supports hybrid queries natively before you build your own.
2. Retrieve 20 to 50 candidates from each method before fusion.
3. Start with RRF because it needs no score tuning.
4. Build a test set of real queries and measure results for each balance setting.
5. Apply metadata filters to both searches equally.

Hybrid search covers the weaknesses of each method. It handles exact identifiers and natural language in a single system, which makes it a standard choice for production search.