Retrievers & Re-ranking — Core

Sparse (BM25) vs dense retrieval, hybrid search and RRF, bi- vs cross-encoders, ColBERT late interaction, SPLADE, query expansion (HyDE), and retrieval metrics.

Medium50 questions

Sample questions

What does BM25 primarily rank documents by?

The semantic similarity of dense sentence embeddings
The order in which documents were added to the index
The number of hyperlinks pointing to each document
Term-frequency and inverse-document-frequency of the query's words

Why

BM25 is a lexical scoring function over term frequency and IDF (with length normalization), rewarding documents that contain the query's rare, informative terms. It is not an embedding method, insertion-order, or a link-graph score like PageRank.

Okapi BM25 BEIR benchmark

What is the defining characteristic of dense retrieval versus sparse (BM25) retrieval?

It matches by learned embedding similarity, not exact term overlap
It requires the query and document to share at least one word
It stores documents as compressed images
It ranks purely by document length

Why

Dense retrieval embeds queries and documents into vectors and matches by semantic similarity, so it can find relevant text with no shared words — unlike BM25's exact-term matching. It doesn't require word overlap, store images, or rank by length.

Dense Passage Retrieval (DPR)Sentence-BERT

A user searches for a specific error code like 'ERR_2048'. Why might BM25 beat a dense retriever here?

BM25 embeds the code into a richer vector
BM25 exact-matches the rare token dense vectors blur
Dense retrievers cannot handle short queries at all
BM25 always outperforms dense retrieval on every query

Why

Exact identifiers and rare tokens are BM25's strength — it matches the literal term, whereas a dense model may map the code near semantically 'similar' but wrong text. BM25 doesn't embed, dense models handle short queries, and neither wins universally (hence hybrid search).

BEIR benchmark Okapi BM25

Hybrid search combines dense and sparse retrieval. Why is this often better than either alone?

It removes the need to embed the query at all
It guarantees the exact same results as BM25
It captures both semantic matches and exact-term matches
It doubles the embedding dimensionality automatically

Why

Dense retrieval finds meaning-based matches while sparse retrieval catches exact terms, rare tokens, and identifiers, so fusing them covers each other's blind spots. It still embeds the query, differs from pure BM25, and doesn't change dimensionality.

BEIR benchmark Contriever

Reciprocal Rank Fusion (RRF) merges two ranked lists. What does it primarily use from each list?

The raw similarity scores added directly together
Each document's rank position, not its raw score
Only the single top result from each list
The document lengths to break ties

Why

RRF sums 1/(k + rank) across lists, using each document's rank position, which sidesteps the problem of combining incomparable score scales (e.g. cosine vs BM25). It doesn't add raw scores, keep only the top result, or use length.

BEIR benchmark Sentence-Transformers docs

Free account

Take the full module

These are the first few of 50 questions. A free account opens the rest as a scored drill.

Every question in this module
Instant feedback and supporting reading
Your score and progress, saved

Free · your email is used for progress only.

About Nybble^™

The AI space moves fast.
Nybble^™ is how you keep up — and stay sharp.

Snack

What happened. In two minutes.

The AI news cycle moves at a pace no one can keep up with. Snack distills what launched, what shipped, and what matters — every day, without the filler.

Go to Snack

Stack

The concepts behind the headlines.

News tells you what. Stack tells you why and how. From RAG architectures to agentic evals, these are the ideas that will shape what you build next.

Go to Stack

Hack

Prove you actually get it.

Reading about LangChain is not the same as knowing it. Hack challenges you with production-grade questions, then shows you the references that make the answer stick.

Go to Hack