Advanced Retrieval & RAG quiz

Question 1

What is the fundamental difference between dense retrieval and sparse (BM25) retrieval?

Accepted Answer

BM25 uses exact term matching; dense retrieval uses neural embeddings capturing semantic similarity beyond lexical overlap

Answer

Dense retrieval is faster at query time than BM25 because it skips inverted index lookups

Answer

BM25 is a deep learning method; dense retrieval is a traditional statistical method

Answer

Both methods produce identical results on paraphrased queries

Question 2

What is a bi-encoder vs. a cross-encoder architecture and when do you use each?

Accepted Answer

Bi-encoder enables fast pre-computed ANN retrieval; cross-encoder enables accurate re-ranking on small candidate sets

Answer

Cross-encoders are used for first-stage retrieval because they are faster

Answer

Bi-encoders apply full attention between query and document

Answer

Both architectures have identical latency characteristics

Question 3

What is hybrid search and how is RRF (Reciprocal Rank Fusion) used to combine retrieval results?

Accepted Answer

RRF combines ranked lists by summing 1/(k+rank) per list, making it scale-invariant and avoiding score normalization

Answer

RRF requires normalizing BM25 and vector scores to [0,1] before combining

Answer

RRF only works with two retrieval systems; more systems require different fusion methods

Answer

Hybrid search always outperforms pure BM25 regardless of domain

Question 4

What is HyDE (Hypothetical Document Embeddings) and how does it improve query-document alignment?

Accepted Answer

HyDE generates a hypothetical document from the query using an LLM, then retrieves using the generated document's embedding

Answer

HyDE splits the query into hypothesis and evidence parts before retrieval

Answer

HyDE uses BM25 only—it doesn't involve LLM generation

Answer

HyDE retrieves documents and generates answers in one combined step

Question 5

How do you evaluate a RAG pipeline end-to-end and what metrics does RAGAs use?

Accepted Answer

RAGAs measures Context Precision, Context Recall, Faithfulness, and Answer Relevancy using LLM-as-judge without human labels

Answer

RAGAs uses BLEU and ROUGE scores to evaluate generated answers

Answer

RAGAs only evaluates retrieval quality, not generation quality

Answer

RAGAs requires manually labeled answer pairs for every evaluation

Question 6

What is contextual retrieval and how does Anthropic's implementation improve chunk relevance?

Accepted Answer

Contextual retrieval prepends an LLM-generated document context summary to each chunk before embedding to prevent ambiguity

Answer

Contextual retrieval uses user query context to reweight BM25 term frequencies

Answer

Contextual retrieval embeds the query and context jointly using a cross-encoder

Answer

Contextual retrieval is identical to standard RAG but uses larger chunks

Question 7

How does agentic RAG differ from naive RAG and what tools enable iterative retrieval?

Accepted Answer

Agentic RAG uses an LLM agent to iteratively decide when and what to retrieve, unlike naive one-shot RAG

Answer

Agentic RAG retrieves all documents upfront and caches them for the session

Answer

Agentic RAG is identical to RAG but uses a larger context window

Answer

Agentic RAG removes the retrieval step entirely, using only parametric knowledge

Question 8

What is passage re-ranking with a cross-encoder and which pre-trained models are commonly used?

Accepted Answer

Cross-encoders re-rank bi-encoder candidates by scoring (query, passage) pairs jointly via full cross-attention for higher accuracy

Answer

Cross-encoders are faster than bi-encoders at inference because they skip vector math

Answer

Re-ranking always degrades retrieval quality by introducing cross-encoder bias

Answer

Re-ranking is only useful when BM25 is the first-stage retriever

Question 9

How does product quantization (PQ) reduce the memory footprint of large vector indexes?

Accepted Answer

PQ splits vectors into sub-vectors and quantizes each to codebook centroids, compressing storage by 10-32× with recall tradeoff

Answer

PQ compresses vectors by rounding floats to integers with no accuracy loss

Answer

PQ requires retraining the embedding model to produce smaller vectors

Answer

PQ is only applicable to binary embeddings, not float vectors

Question 10

How do you implement RAG over tabular data (CSVs, SQL databases)?

Accepted Answer

Text-to-SQL converts natural language to SQL queries; table-to-text serializes rows for embedding—both enable RAG over tabular data

Answer

Standard chunk embedding works equally well for tabular and text data

Answer

Tabular RAG only works with CSV files, not SQL databases

Answer

You must fine-tune a separate model for each table schema

Question 11

What is the role of LLM prompt compression in reducing RAG latency and cost?

Accepted Answer

LLMLingua uses a small perplexity model to prune less important tokens from retrieved context, reducing cost 2-5x with minimal quality loss

Answer

Prompt compression always improves answer quality by removing noise

Answer

LLMLingua retrains the retrieval model to generate shorter documents

Answer

Prompt compression works by summarizing retrieved documents using GPT-4

Question 12

What is the needle-in-a-haystack (NIAH) test and what does it reveal about LLM retrieval from context?

Accepted Answer

NIAH tests recall of a specific fact at varying positions within long contexts, revealing lost-in-the-middle degradation in LLMs

Answer

NIAH measures retrieval speed from a vector database under load

Answer

NIAH evaluates the precision of BM25 sparse retrieval on legal documents

Answer

NIAH is a benchmark for multilingual retrieval quality

Question 13

What is vector database namespace isolation and why is it important for multi-tenant RAG?

Accepted Answer

Namespace isolation partitions vector indexes per tenant, preventing cross-tenant data leakage in multi-tenant RAG systems

Answer

Namespace isolation is only relevant for systems with more than 1 million documents

Answer

Multi-tenant RAG requires separate model deployments per tenant

Answer

Metadata filtering alone is always sufficient for multi-tenant security

Question 14

What is the difference between closed-domain and open-domain QA in the context of RAG?

Accepted Answer

Open-domain QA retrieves from large unrestricted corpora; closed-domain QA retrieves from curated domain-specific corpora

Answer

Open-domain QA is easier because more documents are available for the answer

Answer

Closed-domain QA requires web search while open-domain QA does not

Answer

Both have identical retrieval architecture requirements

Test yourself