Abstention
Choosing not to answer when evidence is missing, conflicting or unauthorized.
Engineering note: Treat it as a valid product state with an explicit reason and next step.
Beginner-friendly reference
RAG Engineering Glossary
Use this as a translation layer between plain-English learning and production engineering vocabulary.
ACL / access control
Rules that decide which data a user is allowed to access.
Engineering note: Enforce deterministically before restricted text reaches the model.
BM25 / lexical retrieval
Keyword-oriented ranking that rewards matching terms using corpus statistics.
Engineering note: Often strong for identifiers, commands, error codes and exact phrases.
Chunk
A searchable piece of a larger source document.
Engineering note: Keep stable source/chunk identity and evaluate boundary choices.
Chunk overlap
Text repeated between neighbouring chunks.
Engineering note: It can preserve context but also create duplicates and waste context budget.
Context
The evidence and instructions supplied to the model for a request.
Engineering note: Context construction is a separate stage with ordering, budget and authority decisions.
Context precision
How much retrieved context is actually useful/relevant.
Engineering note: Low precision means the model is being given noise even if the answer sometimes succeeds.
Context recall
Whether the retrieved context contains the information needed to answer.
Engineering note: A generator cannot reliably recover evidence the retriever never supplied.
Dense retrieval
Retrieval using vector representations of semantic meaning.
Engineering note: Useful for paraphrase; it still needs filters, evaluation and authority controls.
Embedding
A numeric vector representation used to compare semantic similarity.
Engineering note: Record the embedding model/version with the index; different vector spaces are not interchangeable.
Evidence
Source material used to support an answer.
Engineering note: Evidence should retain provenance and stable identifiers.
Faithfulness / groundedness
Whether generated claims are supported by supplied evidence.
Engineering note: Measure this separately from retrieval quality and answer style.
Filter
A deterministic constraint applied to eligible records, such as tenant, date or status.
Engineering note: Authorization filters should happen before restricted text can be exposed downstream.
Grounding
Constraining an answer to retrieved evidence and making support inspectable.
Engineering note: Grounding can still fail even when retrieval succeeds.
Hybrid retrieval
Combining lexical and semantic retrieval signals.
Engineering note: Use it when evaluation shows the methods cover complementary query types.
Index
A data structure optimized for finding candidate records quickly.
Engineering note: An index is a derived artifact; version and rebuild it from authoritative sources.
Ingestion
Turning source material into normalized, searchable knowledge records.
Engineering note: Parsing quality, metadata and provenance begin here.
Knowledge lifecycle
How sources are created, updated, superseded, deleted and re-indexed.
Engineering note: Freshness and deletion are operational requirements, not one-time setup tasks.
MRR
Mean Reciprocal Rank; rewards placing the first relevant result near the top.
Engineering note: Useful when the first relevant hit matters, but it does not describe all relevant results.
nDCG
Normalized Discounted Cumulative Gain; rewards useful ranking when relevance can be graded.
Engineering note: Helpful when multiple results have different degrees of relevance.
Prompt injection
Untrusted input attempts to alter model behaviour or instructions.
Engineering note: Retrieved documents can contain indirect injection; RAG alone does not eliminate the risk.
Provenance
The origin and history of a source or derived chunk.
Engineering note: Include source ID, version, location and enough lineage to reproduce the evidence path.
RAG
Retrieval-augmented generation: retrieve external evidence, then generate using that evidence.
Engineering note: It is a system pattern, not a guarantee of correctness.
Recall@k
Whether relevant evidence appears within the first k retrieved results.
Engineering note: Use labelled cases and choose k based on the candidate-generation trade-off.
Reranker
A second-stage scorer that reorders a smaller candidate set using richer signals.
Engineering note: Relevance reranking does not replace authorization or source authority policy.
RRF
Reciprocal Rank Fusion; combines rankings using rank positions rather than raw scores.
Engineering note: Useful when retrievers produce scores that are not directly comparable.
Source authority
Rules that decide which relevant source should be trusted when sources differ.
Engineering note: Status, owner, effective date and source class often matter more than similarity.
Sparse retrieval
Retrieval based on sparse term features, often keyword-oriented.
Engineering note: Lexical/BM25 systems are common sparse retrieval approaches.
Top-k
The number of leading retrieval candidates kept for a later stage.
Engineering note: Candidate k and final context size should not be assumed to be the same.
Vector database
A store/index optimized for vector similarity and associated metadata.
Engineering note: It is infrastructure, not the RAG architecture itself.
GraphRAG
Retrieval that uses graph-structured entities/relationships or community structure in addition to text/vector retrieval.
Engineering note: Useful for some multi-entity and whole-dataset questions; adds indexing/query complexity.
Agentic retrieval
A bounded process that can route, decompose or iteratively retrieve evidence.
Engineering note: Require step budgets, stop rules, traceability and measured benefit over simpler retrieval.