Stop 14 · Retrieval-Augmented Generation & Vector DatabasesTopics · 5/9

Hybrid Search (Keyword + Vector)

In short

Vector search understands meaning but can miss exact terms like 'CS3401' or 'Section 80C'. Keyword search nails exact terms but misses synonyms. Hybrid search runs both and merges the results.

Combining traditional keyword search (BM25) with vector similarity search — often outperforms either method alone, especially for queries with exact terms like product codes or names.

The blob on a railway platform calls through a megaphone while holding up a friend's photo.

Think of it like this

Finding a friend at the station

Looking for a friend on a crowded platform, you shout her name on a megaphone and also hold up her photo. The name catches her if she hears it. The photo helps people who recognise her face. Keyword search is the name, vector search is the photo.

Watch it happen

One query, two searches, one list

Query

BM25

Vector

Fuse

Results

1/5Query: 'CS3401 syllabus for compiler design'.

The big ideas

What you need to know

BM25 for exact terms

A classic keyword ranking method. Great for codes, names, rare words and acronyms.

Vectors for meaning

Catches paraphrases and synonyms, and handles questions phrased very differently from the text.

Fuse by rank

RRF gives each doc 1/(k + rank) from each list and adds them. A k around 60 is a common default.

Built into many tools

Weaviate, Elasticsearch, OpenSearch and others support hybrid search directly.

See it in code

Walk through the code

python

1def rrf(keyword_ids: list[str], vector_ids: list[str], k: int = 60) -> list[str]:2    scores: dict[str, float] = {}3    for ranking in (keyword_ids, vector_ids):4        for rank, doc_id in enumerate(ranking, start=1):5            scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank)6    return sorted(scores, key=scores.get, reverse=True)7 8kw = ["d7", "d2", "d9"]     # from BM259vec = ["d2", "d4", "d7"]    # from vector search10print(rrf(kw, vec))         # ['d2', 'd7', 'd4', 'd9']

1/3Each document earns 1/(k + rank) from every list it appears in.

Reciprocal Rank Fusion in a few lines. d2 and d7 appear in both lists, so they win.

Save yourself the pain

Common mistakes

Oops

Adding BM25 scores and cosine scores directly.

Do this instead

They're on different scales. Fuse by rank with RRF, or normalise scores first.

Oops

Assuming vector search alone handles product codes and IDs.

Do this instead

Test queries with exact codes. If they fail, add keyword search.

Your mission

Hybrid on your notes

~35 min

Quick check

A user searches 'Section 80C deduction'. Vector search returns general tax articles but misses the exact section. What helps most?

Go deeper

1Implement hybrid search combining BM25 and vector similarity scores.
2Test on queries where keyword search wins (exact terms) vs. where semantic search wins (paraphrased queries).
3Learn common score-fusion techniques (e.g. reciprocal rank fusion).