RAG and vector databases quiz

Question 1

In a vector database, why is HNSW (Hierarchical Navigable Small World) generally preferred over a flat (brute-force) index for large-scale ANN search?

Accepted Answer

It gives sub-linear query time by navigating a layered proximity graph

Answer

It guarantees exact nearest neighbors with no recall loss

Answer

It uses less memory than a flat index in all cases

Answer

It eliminates the need for a distance metric

Question 2

What is the primary trade-off controlled by the HNSW parameter `ef_search` (a.k.a. `ef`) at query time?

Accepted Answer

Recall/accuracy vs. query latency

Answer

Index build memory vs. disk usage

Answer

Embedding dimensionality vs. storage

Answer

Number of shards vs. replicas

Question 3

In RAG, why is cosine similarity often used instead of raw dot product when embeddings have not been normalized?

Accepted Answer

Cosine ignores magnitude and compares only direction

Answer

Cosine is always faster to compute than dot product

Answer

Dot product cannot be used in vector databases

Answer

Cosine requires no embeddings

Question 4

What problem does IVF (Inverted File Index) with PQ (Product Quantization) primarily solve in vector search?

Accepted Answer

They reduce search space and memory via clustering plus compression

Answer

They guarantee exact search at billion scale

Answer

They eliminate the need for embeddings

Answer

They convert sparse vectors to dense vectors

Question 5

What is 'hybrid search' in the context of RAG retrieval?

Accepted Answer

Combining dense semantic search with sparse lexical (keyword) search

Answer

Searching across two vector databases simultaneously

Answer

Mixing two embedding models in one vector

Answer

Running search on both GPU and CPU

Question 6

Explain Reciprocal Rank Fusion (RRF) and why it is popular for combining retrieval result lists.

Accepted Answer

It sums 1/(k+rank) across lists, needing no score normalization

Answer

It averages the raw similarity scores of each retriever

Answer

It multiplies the ranks from each retriever

Answer

It picks the single highest-scoring document only

Question 7

Why does a cross-encoder reranker typically outperform a bi-encoder retriever on relevance, and what is the cost?

Accepted Answer

Cross-encoders model query-document interactions jointly but cannot be precomputed

Answer

Cross-encoders are faster because they precompute document vectors

Answer

Bi-encoders always beat cross-encoders on relevance

Answer

Cross-encoders require no query at indexing time

Question 8

What is the 'lost in the middle' phenomenon and how does it affect RAG context construction?

Accepted Answer

LLMs underweight information placed in the middle of long contexts

Answer

LLMs forget the system prompt over long conversations

Answer

Vector indexes corrupt middle-ranked results

Answer

Embeddings lose precision for mid-length documents

Question 9

In chunking for RAG, what is the purpose of using overlapping windows between adjacent chunks?

Accepted Answer

To preserve context for information spanning chunk boundaries

Answer

To reduce the total number of vectors stored

Answer

To increase embedding dimensionality

Answer

To enforce exact-match keyword search

Question 10

What is the curse of dimensionality's impact on nearest-neighbor search in high-dimensional embedding spaces?

Accepted Answer

Distances concentrate, reducing contrast between near and far points

Answer

Higher dimensions always make search faster

Answer

Embeddings become exactly orthogonal and unsearchable

Answer

Cosine similarity is undefined above 512 dimensions

Question 11

What is Matryoshka Representation Learning (MRL) and why is it useful for vector databases?

Accepted Answer

Embeddings that can be truncated to shorter dims with graceful degradation

Answer

A technique to compress text before embedding

Answer

A graph index replacing HNSW

Answer

A method to merge multiple vector databases

Question 12

What does scalar/binary quantization of embeddings achieve, and what is the main risk?

Accepted Answer

It reduces memory/compute at the cost of some recall

Answer

It increases embedding dimensionality for better accuracy

Answer

It guarantees identical results to float32 search

Answer

It converts dense vectors to sparse vectors

Question 13

In RAG evaluation, what does the metric 'faithfulness' (a.k.a. groundedness) measure?

Accepted Answer

Whether the answer's claims are supported by the retrieved context

Answer

Whether retrieved chunks are ranked correctly

Answer

Whether the embedding model is well-calibrated

Answer

The latency of the generation step

Question 14

What is the difference between 'context recall' and 'context precision' in RAG retrieval evaluation?

Accepted Answer

Recall = needed info retrieved; precision = retrieved chunks are relevant/ranked well

Answer

Recall is for generation, precision is for embeddings

Answer

They are identical metrics with different names

Answer

Precision measures latency, recall measures cost

Question 15

What is HyDE (Hypothetical Document Embeddings) and what retrieval problem does it address?

Accepted Answer

Generating a hypothetical answer and embedding it to search

Answer

Encrypting documents before embedding

Answer

Hashing documents into buckets

Answer

A reranking algorithm based on dot product

Question 16

What is query expansion / multi-query retrieval and why does it improve recall in RAG?

Accepted Answer

Generating multiple query variants and merging their retrievals

Answer

Compressing the query to fewer tokens

Answer

Searching only the most frequent keyword

Answer

Increasing the embedding dimension dynamically

Question 17

What is the role of a 'parent document' or 'small-to-big' retrieval strategy?

Accepted Answer

Embedding small chunks for matching but returning larger parent context

Answer

Embedding large chunks but returning small ones

Answer

Storing documents twice for redundancy

Answer

Using two embedding models per chunk

Question 18

Why might you choose a disk-based ANN index (e.g., DiskANN) over a fully in-memory index?

Accepted Answer

It enables billion-scale search with low RAM at the cost of latency

Answer

It is always faster than in-memory HNSW

Answer

It requires no SSD

Answer

It eliminates the need for quantization

Question 19

In pgvector, what is the difference between the `vector_cosine_ops`, `vector_l2_ops`, and `vector_ip_ops` operator classes when creating an index?

Accepted Answer

They specify which distance metric the index optimizes

Answer

They set the embedding dimension

Answer

They control the number of index replicas

Answer

They define table partitioning

Question 20

What is metadata filtering in vector search, and why is 'pre-filtering' vs 'post-filtering' an important distinction?

Accepted Answer

Pre-filtering restricts candidates before ANN; post-filtering drops after and may under-return

Answer

Pre-filtering is always slower and never used

Answer

They produce identical results and counts

Answer

Post-filtering happens at index build time

Question 21

Why is multi-tenancy in vector databases often implemented with namespaces/collections rather than a single shared index with a tenant filter?

Accepted Answer

Namespaces give isolation and avoid selective-filter recall issues

Answer

Namespaces are always cheaper than filtering

Answer

Filtering is impossible in vector databases

Answer

Namespaces remove the need for embeddings

Question 22

What does the `M` parameter control in an HNSW index, and how does it affect the index?

Accepted Answer

The max neighbors per node, trading memory/build time for recall

Answer

The number of query results returned

Answer

The embedding dimension

Answer

The number of shards

Question 23

What is 'semantic chunking' and how does it differ from fixed-size chunking?

Accepted Answer

Splitting at semantic boundaries detected via embedding similarity shifts

Answer

Splitting by exact character count

Answer

Splitting only at page breaks

Answer

Removing stopwords before embedding

Question 24

In RAG, why can a high embedding model dimensionality not directly compensate for a poor domain match of the embedding model?

Accepted Answer

Domain mismatch is about learned semantics, which dimensions alone do not fix

Answer

More dimensions always fix domain mismatch

Answer

Dimensionality is irrelevant to retrieval

Answer

Fine-tuning reduces dimensionality

Question 25

What is the purpose of an embedding 'rescoring' or 'oversampling' step when using quantized vectors?

Accepted Answer

To recover recall by re-ranking an oversampled shortlist with full precision

Answer

To increase index build speed

Answer

To reduce the number of stored vectors

Answer

To encrypt the candidate vectors

Question 26

Why is it important to use the same embedding model (and version) for both indexing documents and embedding queries?

Accepted Answer

Query and document embeddings must share the same model/space to be comparable

Answer

Different models can be mixed if dimensions match

Answer

Only the query model matters

Answer

Re-embedding is never needed after a model change

Question 27

What is 'contextual retrieval' (e.g., contextual chunk embeddings) and how does it improve retrieval accuracy?

Accepted Answer

Prepending LLM-generated document context to each chunk before embedding

Answer

Caching query results for reuse

Answer

Encrypting chunks for privacy

Answer

Searching only the most recent documents

Question 28

In a RAG system, what is the main risk of setting the retrieval top-k too high?

Accepted Answer

It adds noise/cost and can degrade generation via distraction

Answer

It always improves answer quality

Answer

It reduces recall

Answer

It has no effect on latency

Question 29

What is ColBERT-style 'late interaction' retrieval and how does it differ from single-vector dense retrieval?

Accepted Answer

It stores per-token vectors and scores via MaxSim for fine-grained matching

Answer

It uses one vector per document like standard dense retrieval

Answer

It is a keyword-only method

Answer

It eliminates the need for an index

Question 30

Why might cosine similarity scores from an embedding model be unreliable as an absolute confidence/threshold signal?

Accepted Answer

Scores are uncalibrated and model-specific, so absolute thresholds are brittle

Answer

Scores are perfectly calibrated probabilities

Answer

Cosine always ranges 0 to 1 with fixed meaning

Answer

Thresholds transfer across all models

Test yourself