Practice English vocabulary for RAG retrieval strategies: dense retrieval, sparse retrieval, hybrid retrieval, top-k similarity, and reranking.
0 / 5 completed
1 / 5
What is 'dense retrieval' and how does it work?
Dense retrieval (e.g., using bi-encoders like sentence-transformers) encodes text into high-dimensional vectors. Documents semantically similar to the query score high even without sharing keywords — enabling retrieval like 'find documents about car engines' matching 'automobile motor specifications'.
2 / 5
What is 'sparse retrieval using BM25 keyword matching'?
BM25 (Best Match 25) is a strong baseline for keyword retrieval. It rewards documents that contain query terms (TF) while penalizing common terms that appear everywhere (IDF). It's fast, interpretable, and still competitive with neural methods for many queries.
3 / 5
What is 'hybrid retrieval that combines dense and sparse scores'?
Hybrid retrieval addresses the weaknesses of each approach individually. Dense retrieval misses exact keyword queries (product codes, names); sparse retrieval misses paraphrases. Combining both with score normalization and fusion (e.g., RRF — Reciprocal Rank Fusion) consistently outperforms either alone.
4 / 5
What does 'retrieval is top-k by cosine similarity' mean?
Cosine similarity measures the angle between two vectors (ignoring magnitude). Values range from -1 to 1; higher values mean more similar directions in embedding space, indicating semantic similarity. Top-k retrieval returns the k highest-scoring documents as context for the LLM.
5 / 5
What does 'the reranker reorders retrieval results' do?
Two-stage retrieval: a fast bi-encoder retrieves top-k candidates, then a slower but more accurate cross-encoder reranks them. Cross-encoders consider query-document interactions jointly, producing more accurate relevance scores but are too slow to use across the full document corpus.