Glossary

Canonical terms for this subject. Every lesson uses these words.

Pipeline

RAG (retrieval-augmented generation): Answering with a model whose prompt is augmented by documents retrieved from an external index at query time. Avoid: "search + AI", grounding (that is the goal, not the mechanism).
Chunk: The unit of text that gets embedded, indexed and retrieved. One chunk is one retrievable fact-bundle.
Embedding: A fixed-length vector encoding a text's meaning, so semantic similarity becomes vector distance.
Index: The data structure that makes retrieval fast: a vector index for embeddings, an inverted index for keywords.
Contextual retrieval: Prepending a short chunk-specific context (what document, what section, what entity) before embedding and indexing the chunk. Avoid: context stuffing.

Dense retrieval: Ranking chunks by embedding similarity to the query embedding. Finds meaning, misses exact tokens.
Sparse retrieval: Ranking by weighted keyword overlap, canonically BM25. Finds exact tokens (codes, names, addresses), misses paraphrase.
Hybrid retrieval: Running dense and sparse retrieval in parallel and fusing the ranked lists.
RRF (reciprocal rank fusion): Fusing ranked lists by scoring each item 1 / (k + rank) per list and summing, k typically 60.
Metadata filter: A hard predicate (price ≤ 400000, city = Ghent, tenant = this agency) applied by the index before or with similarity ranking. Filters are exact; similarity is approximate.
Reranker: A second-stage model that rescores the top candidates by reading query and chunk together (a cross-encoder), trading latency for precision.
HyDE: Query rewriting that embeds a hypothetical answer instead of the raw question, so the query vector lives where answers live.
Top-k: The k highest-ranked chunks a retrieval stage passes on. Every stage has its own k.

Gold set: A fixed list of real questions, each with the chunks that should be retrieved and the facts a correct answer must state.
Recall@k: The fraction of gold chunks that appear in the top k retrieved. The retrieval ceiling: the generator cannot cite what retrieval never returned.
MRR (mean reciprocal rank): Average of 1 / rank of the first relevant chunk. Measures whether the right chunk is near the top, not merely present.
Faithfulness: The fraction of claims in the answer supported by the retrieved chunks. Low faithfulness with good retrieval means the generator is inventing.
Context precision: How much of what was retrieved was actually relevant. Low precision wastes context and invites distraction.
LLM-as-judge: Using a model to grade answers. Useful and biased: position, verbosity and self-enhancement biases are documented.

Indirect prompt injection: Instructions hidden inside retrieved content that the model mistakes for operator instructions. The signature RAG attack. Avoid: jailbreak (that is a different attack, by the user).
Row-level access control: Filtering retrievable chunks by the caller's permissions at query time, inside retrieval, never after generation.

Terms are promoted here as lessons cover them; definitions tighten as understanding deepens.