Retrievers & Re-ranking — Advanced

ColBERT/SPLADE internals, DPR/RocketQA training, hard-negative denoising, cross-encoder distillation, RRF tuning, LLM rerankers, lost-in-the-middle, and IR metrics.

Hard50 questions

Sample questions

On the retrieval cost/accuracy spectrum, how do bi-encoder, ColBERT, and cross-encoder compare per query-document pair?

All three cost the same; only accuracy differs
Bi-encoder: dot product; ColBERT: MaxSim; cross-encoder: full forward pass
Cross-encoder is cheapest; bi-encoder is most expensive
ColBERT runs a forward pass per pair, like a cross-encoder

Why

A bi-encoder compares two pooled vectors (one dot product), ColBERT does token-level MaxSim over precomputed vectors (moderate), and a cross-encoder runs a joint forward pass per pair (most expensive, most accurate). The other options invert or flatten this ordering.

ColBERT Sentence-BERT

ColBERTv2 makes late interaction storage-efficient. What is its main compression idea?

Storing token vectors as residuals to nearby cluster centroids
Pooling all token vectors into a single mean vector
Dropping every token vector except the first
Converting the vectors into BM25 term weights

Why

ColBERTv2 clusters token embeddings and stores each as a centroid id plus a quantized residual, drastically shrinking the multi-vector index while keeping late-interaction quality. It doesn't mean-pool (that would lose late interaction), keep only one token, or become BM25.

ColBERTv2 ColBERT

SPLADE derives sparse term weights from a transformer. Where do the per-term scores come from?

From the document's raw term counts, like BM25
From cosine similarity between two dense vectors
From the masked-language-model output logits over the vocabulary
From the document's creation timestamp

Why

SPLADE reuses the MLM head: it aggregates the model's vocabulary logits (with a log-saturation and ReLU) into a sparse weight per term, including expansion terms. It is not raw counts, dense cosine, or timestamps.

SPLADE SPLADE v2

SPLADE adds a FLOPS (sparsity) regularizer during training. Why is it necessary?

To increase the embedding dimensionality over time
To force the model to output dense vectors
To translate expansion terms into other languages
To keep postings lists sparse for fast lookup

Why

Without a sparsity penalty the learned term weights would be dense, defeating the inverted-index efficiency SPLADE aims for; the FLOPS regularizer pushes most weights to zero. It doesn't change dimensionality, force density, or translate terms.

SPLADE v2 SPLADE

RocketQA improved dense retrieval training with 'cross-batch negatives'. What does that mean?

Reusing negatives across GPUs to enlarge the pool
Removing all negatives to speed up training
Using only the single hardest negative per query
Sampling negatives from a different language corpus

Why

Cross-batch negatives gather in-batch negatives from all GPUs (and steps), giving each query far more negatives than a single device's batch — a stronger contrastive signal. It's not negative removal, single-negative, or cross-lingual sampling.

RocketQA Dense Passage Retrieval (DPR)

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