Module 3: Sparse vs Dense vs Hybrid Search on Qdrant - Vector Search Engine

Module Overview

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Sparse vs Dense vs Hybrid Search

Understand dense versus sparse retrieval, their strengths, and how a hybrid approach can combine them.

Follow-along code: Module 3 notebook

Overview

Module 2 showed you where your data lives and how Qdrant retrieves it. In this module, you’ll learn what that retrieval misses and how to cover the gap. You’ll explore dense and sparse vectors, BM25, and the inverted index, then see why a product code defeats either one alone. You’ll also learn how fusion merges two ranked lists, and where a filter belongs so both retrievers respect it. By the end, you’ll have built a hybrid collection, run a fused query, and filtered it correctly.

Where We Left Off

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Where We Left Off

In Module 2, you built a complete ingestion and retrieval pipeline: raw text, vector, store, top-K query. Dense retrieval handles meaning well. It gets shaky on the part of a query that has to be exact.

Here is a shoe catalog with two products one digit apart. Searching it dense-only for Nike Pegasus 40:

Result	Dense score
Nike Pegasus 40 running shoes	0.8713
Nike Pegasus 41 running shoes	0.8626
Nike Pegasus 40 womens running shoes	0.7830
Nike Pegasus Trail 4 trail running shoes	0.7425

The Problem

Dense ranks the right shoe first, so nothing here looks broken. Look at the margin: 0.0087, about one percent of the top score. To the model, “Pegasus 40” and “Pegasus 41” are near-identical statements about running shoes, because that is what they are. The digit a shopper cares about is one token out of five, averaged into a vector describing the whole phrase.

The Two Families of Search

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

The Two Families of Search

Every retrieval system is built from one or both of these.

Dense Search

A dense vector has a small, fixed number of dimensions, 384 for the model used here, and every one of them holds a value. Two texts with similar meaning land close together whether or not they share any words:

Hybrid Search: Dense and Sparse

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Hybrid Search: Dense and Sparse

Hybrid search runs both retrievers in the same request and combines their ranked lists into one result set.

Run the same Nike Pegasus 40 query through sparse alone and the picture inverts. The top three:

Result	Sparse score
Nike Pegasus 40 running shoes	3.8396
Nike Pegasus 40 womens running shoes	3.8293
Nike Pegasus 41 running shoes	1.7007

Sparse pushes the 41 down to third, because 40 is a different token from 41. But it now has dense’s problem on a different pair: the men’s and women’s Pegasus 40 share every token the query contains, so it separates them by 0.0103 on a 3.8396 top score, roughly a quarter of a percent. Dense had those two 0.09 apart and no trouble at all.

Setting Up Hybrid Search in Qdrant

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Setting Up Hybrid Search in Qdrant

Hybrid search uses named vectors, dense and sparse on the same point, and the Query API to run a sub-query against each before fusing. A prefetch is one of those sub-queries: it produces a candidate list that fusion then merges.

Step 1: Create a Hybrid Collection

Two things are new since Module 2. The collection declares a sparse config alongside the dense one, so both vectors live on the same point. And that sparse config carries a modifier, which has no dense equivalent: it tells Qdrant to compute the second half of the BM25 score at query time, the inverse document frequency. That half is what makes a rare token like 40 outweigh a common one like shoes, so without it BM25 scoring is wrong rather than merely untuned. miniCOIL needs the same modifier.

Fusion Strategies

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Fusion Strategies

The difference between the two strategies is what each does with magnitude. RRF knows only that a document came first, second, or third, so a runaway top match and a photo finish look identical to it. DBSF rescales each retriever’s scores onto a comparable range before combining them, which keeps that information at the cost of depending on how those scores are distributed.

Strategy	How it works	When to use it
RRF (Reciprocal Rank Fusion)	Merges by rank position, discarding raw scores	The default, and the safe choice whenever the two score scales differ, which is nearly always
DBSF (Distribution-Based Score Fusion)	Normalizes each retriever’s score distribution, then combines	When the size of the gaps between scores carries information worth keeping

Neither reliably beats the other, so treat the choice as an evaluation result rather than a preference: start with RRF and switch only after measuring on a set of queries with known-good answers. The Hybrid Queries documentation covers both, along with their tuning parameters.

Filtering: Works with Any Retrieval Method

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Filtering: Works with Any Retrieval Method

Payload filters are not a hybrid-only feature. The same conditions apply to dense-only, sparse-only, and hybrid retrieval, and Qdrant evaluates them during the search rather than after it, so a filtered search still returns a full result set. What changes is where the filter goes.

Take the constraints a real shopper has: in stock, and available in their size.

shopper_filter = models.Filter(
 must=[
 models.FieldCondition(key="in_stock", match=models.MatchValue(value=True)),
 models.FieldCondition(key="sizes", match=models.MatchValue(value=11)),
 ]
)

sizes is a list on each point, and MatchValue on a list matches when any element matches, so this reads as “size 11 is among the sizes stocked”.

Knowledge Check

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

Knowledge Check

Q: A shopper searches your catalog for iPad Air, and dense-only search returns iPad Mini first. Both are reasonable matches for the words, but the ranking is wrong. What would you add, and why would it fix this specific failure?

Show answer

Add sparse (BM25) retrieval alongside dense, combined through hybrid search. Dense embeds the whole phrase into one vector, so “Air” and “Mini” barely move the score, the same failure mode as Pegasus 40 versus 41. Sparse treats “Air” and “Mini” as distinct tokens with no relationship, so it separates the two products cleanly. Hybrid fusion lets the sparse side catch what dense alone misses.

References and Further Reading

info@qdrant.tech (Andrey Vasnetsov) — Mon, 01 Jan 0001 00:00:00 +0000

Module 3

References and Further Reading

Hybrid Queries: prefetch semantics, both fusion strategies, and their tuning parameters.
Understanding SPLADE and Sparse Vectors: how sparse vectors work and how SPLADE compares to BM25.
miniCOIL: Sparse Neural Retrieval: why miniCOIL exists and how it extends BM25 with contextual meaning.
Filtering: full filter syntax and the payload index each condition needs.
Named Vectors: configuring and querying more than one vector on a single point.

What’s Next: Module 4

Eight products rebuild in seconds. On a collection that takes hours to embed, one wrong decision means embedding everything again.