RAG glossary
Key terms from this course, in plain English.
- Agentic RAG
- RAG where an LLM decides when and what to search, possibly running several searches before answering.
- Approximate nearest neighbour (ANN)
- Search methods that find very close vectors quickly, trading a little accuracy for speed.
- BM25
- A classic lexical ranking function that scores documents by matching query words, weighting rare words more.
- Chunk
- A passage of a document indexed as a unit for retrieval.
- Chunk overlap
- Repeating the end of one chunk at the start of the next to preserve context across boundaries.
- Citation
- A reference in an answer to the source passage that supports a claim.
- Contextual retrieval
- Prepending an LLM-written description of each chunk's place in its document before indexing, to improve retrieval.
- Cosine similarity
- A measure of how similar two vectors' directions are, from −1 to 1.
- Embedding
- A vector of numbers representing the meaning of a piece of text.
- Faithfulness
- Whether every claim in an answer is supported by the retrieved sources (also called groundedness).
- HNSW
- Hierarchical Navigable Small World — a popular graph-based ANN index.
- Hybrid search
- Combining lexical and semantic search results.
- Indexing
- Preparing documents for retrieval: cleaning, chunking, embedding and storing them.
- LLM-as-judge
- Using a language model, with a rubric, to grade outputs.
- Metadata
- Information stored with each chunk — source, date, section, permissions — used for filtering and citations.
- MRR
- Mean reciprocal rank: the average of 1/rank of the first relevant result.
- Query rewriting
- Transforming a user's question (e.g. a follow-up) into a better standalone search query.
- RAG
- Retrieval-augmented generation: retrieving relevant passages and giving them to an LLM to ground its answer.
- Recall@k
- The share of test questions where a relevant passage appears in the top k results.
- Reciprocal rank fusion (RRF)
- Merging ranked lists by summing 1/(k + rank) for each document.
- Reranker
- A model that scores query–passage pairs more accurately than first-stage retrieval, used on a shortlist.
- Semantic search
- Search by meaning using embeddings rather than exact words.
- Vector database
- A database designed to store vectors and run similarity search, often with metadata filtering.