How to use terms correctly
- Retrieval quality is not the same as answer quality.
- Grounding means source support, not guaranteed truth.
- Citations should point to verifiable evidence.
- Evaluation should combine metrics and expert review.
Glossary
RAG terminology, alphabetically searchable
Use the glossary to align teams around retrieval, evaluation, grounding, citations, and architecture language.
How to use the glossary
Use these terms when writing project briefs, evaluation plans, procurement requirements, and tutorials. Shared vocabulary prevents confusion between retrieval quality, answer quality, source authority, and model behavior.
A simple RAG example
When RAG is the wrong tool
A strong first project
Agentic RAG
A RAG pattern where an agent plans retrieval steps, calls tools, reflects on results, or performs multi-step reasoning before answering.
Related: agent workflow, query rewriting
Answer faithfulness
The degree to which an answer is supported by the retrieved context rather than unsupported model inference.
Related: grounding, citation
Chunking
The process of splitting documents into retrievable units, usually with attention to size, overlap, structure, and metadata.
Related: embedding, retrieval
Citation
A reference that links generated claims back to source passages, documents, or records used in retrieval.
Related: traceability, grounding
Context window
The maximum amount of text or tokens a model can consider at once during generation.
Related: long-context models, prompt assembly
Embedding
A numerical representation of text, images, or other data used to compare semantic similarity.
Related: vector database, semantic search
Evaluation
The practice of measuring retrieval quality, answer quality, faithfulness, latency, cost, and user outcomes.
Related: recall, precision
Graph RAG
A RAG approach that uses graph structures, entities, and relationships to improve retrieval and reasoning across connected knowledge.
Related: knowledge graph, metadata
Grounding
Connecting model output to external evidence so answers are constrained by sources rather than only model memory.
Related: hallucination, citation
Hallucination
A generated statement that is false, unsupported, or not justified by the available context.
Related: faithfulness, evaluation
Hybrid search
Retrieval that combines keyword matching with vector similarity, often improving recall and precision for mixed queries.
Related: semantic search, metadata filtering
Knowledge base
A curated collection of documents, records, or facts prepared for retrieval by an AI system.
Related: ingestion, metadata
Metadata filtering
Restricting retrieval using attributes such as date, author, department, language, jurisdiction, or document type.
Related: retrieval, hybrid search
Multimodal RAG
RAG over multiple data types such as text, images, audio, video, tables, or scanned documents.
Related: document parsing, embeddings
Precision
The share of retrieved items that are relevant to the query.
Related: recall, evaluation
Prompt injection
An attack or failure mode where retrieved or user-supplied text attempts to override system instructions or safety boundaries.
Related: security, retrieval
RAG
Retrieval-Augmented Generation: connecting language models with external knowledge sources to improve grounding, accuracy, traceability, and domain relevance.
Related: retrieval, grounding
Recall
The share of relevant items that the retrieval system successfully finds.
Related: precision, evaluation
Reranking
A second-stage retrieval step that reorders initially retrieved candidates using a more precise relevance model.
Related: retrieval, hybrid search
Retrieval
Finding relevant passages, documents, or data records for a user query before generation.
Related: vector database, semantic search
Semantic search
Search based on meaning similarity rather than exact keyword overlap.
Related: embedding, vector database
Vector database
A database designed to store embeddings and retrieve nearby vectors efficiently.
Related: embedding, semantic search