Retrieval & RAG

Concepts for finding external information and supplying useful context to models.

Approximate nearest neighbor search aims to reduce search cost without guaranteeing the exact nearest neighbors.

BM25 ranks lexical matches using term rarity, diminishing returns from repetition, and document-length normalization.

Chunking divides source content into smaller units that can be indexed, retrieved, and placed in model context.

Dense retrieval finds candidate items by comparing learned, densely populated representations of queries and stored content.

An embedding is a numerical representation that places an input in a learned vector space.

HNSW is a layered graph index for approximate nearest neighbor search over vectors.

Hybrid search combines results from multiple retrieval methods, often sparse and dense, into one ranked set.

Indexing prepares searchable structures so a retriever can find relevant items in a corpus.

An inverted index looks up documents by term and can store term positions for phrase queries.

Lexical similarity measures how closely texts match in their wording or character sequences.

Query rewriting reformulates a request so a retrieval system can find more relevant content.

Reranking reorders retrieved candidates using a second, usually more expensive relevance step.

Retrieval top-k is the number of highest-ranked results requested or retained from a retrieval stage.

RAG retrieves relevant information at runtime and supplies it as context before a model generates an answer.

A retriever selects candidate information from a corpus for a query.

Sparse retrieval ranks items for a query using a small subset of features from a large feature space.

TF-IDF weights terms by their frequency in a document and how widely they appear across a collection.

A vector database stores vectors and metadata and provides indexed similarity search over them.

Vector search finds items whose stored vectors are close to a query vector.