Advanced RAG

Question 1

What does hybrid search combine?

Accepted Answer

Dense vector and sparse keyword search — Hybrid search fuses semantic (vector) similarity with keyword (BM25) matching, so both meaning and exact terms contribute to ranking.

Question 2

Advanced RAG techniques mainly aim to improve…

Accepted Answer

Retrieval quality and relevance — Naive RAG often retrieves mediocre context; advanced techniques refine query, retrieval, and ranking to feed the LLM better context.

Question 3

RERANKING with a cross-encoder improves results by…

Accepted Answer

Re-scoring candidates for relevance — A cross-encoder jointly reads query+document for precise relevance, reordering the fast bi-encoder's candidates so the LLM sees the most relevant first.

Question 4

QUERY REWRITING / expansion helps because…

Accepted Answer

The raw query often retrieves poorly — Reformulating vague or short queries (adding synonyms/context) surfaces more relevant documents than the original wording.

Question 5

Multi-query retrieval generates…

Accepted Answer

Several variations, fused results — Different phrasings retrieve different relevant docs; merging their results (e.g. reciprocal rank fusion) improves recall.

Question 6

HyDE (Hypothetical Document Embeddings) works by…

Accepted Answer

Drafting a hypothetical answer first — A hypothetical answer is often closer in embedding space to relevant documents than the bare question, improving retrieval.

Question 7

'Small-to-big' / parent-child chunking retrieves on small chunks but…

Accepted Answer

Returns the larger parent context — Small chunks give precise matching; expanding to surrounding parent context gives the LLM enough information to answer well.

Question 8

In Reciprocal Rank Fusion the constant k (default 60) controls…

Accepted Answer

How sharply top-ranked items are rewarded over lower-ranked ones — Smaller k amplifies the advantage of top ranks; larger k flattens the distribution, making mid-ranked results more competitive. k=60 is a robust default across corpora because it prevents any single top-rank from dominating the fused score.

Question 9

Contextual compression in RAG…

Accepted Answer

Trims chunks to relevant parts — Removing irrelevant text from retrieved chunks reduces noise and context-window usage, focusing the LLM on what matters.

Question 10

Why does naive (top-k vector) RAG often underperform?

Accepted Answer

Retrieval returns irrelevant chunks — Pure similarity can miss precise relevance, context, or exact terms; advanced techniques address these retrieval failures.

Key insight

Where does each technique run?

Test your knowledge

What does hybrid search combine?

Interview Questions

References & Further Reading

Related topics

Love this visualization?