--- title: "Retrieval-augmented generation: RAG system guide" description: "Retrieval-augmented generation gives a language model selected external context, helping applications answer from current, private or cited sources." canonical: "https://zephior.com/glossary/retrieval-augmented-generation" last-updated: 2026-07-28 --- # Retrieval-augmented generation: RAG system guide > Retrieval-augmented generation gives a language model selected external context, helping applications answer from current, private or cited sources. By [Tony Kim](https://zephior.com/authors/tony-kim). Published 2026-07-28; updated 2026-07-28. 5 minute read. ## Definition Retrieval-augmented generation, or RAG, is an application pattern that retrieves relevant external information for a user task and supplies selected context to a generative model. The model then produces an answer conditioned on that context rather than relying only on its learned parameters. ## Problem Adding a vector database does not make answers factual. The system can index the wrong content, retrieve an irrelevant passage, omit a decisive clause, cross a permission boundary or generate a claim that the passage does not support. End-to-end quality hides whether the failure came from retrieval or generation. ## Point of view RAG is a controlled evidence pipeline. Manage source authority, permissions, parsing, retrieval, context assembly, citations and answer policy as distinct stages, and evaluate each stage with questions drawn from the real workflow. ## RAG has multiple failure points The ingest path discovers content, checks access, parses structure, segments text, attaches metadata and updates the index. The query path interprets the task, filters permitted content, retrieves candidates, reranks passages, assembles context and generates an answer. Citation rendering and feedback happen after generation. A failure at an early stage propagates. If the table parser loses a column, no retriever can recover the missing relationship. If the permission filter is late, sensitive content may already be exposed. If the source link is lost during chunking, the answer cannot provide useful provenance. | Stage | Failure example | Diagnostic | | --- | --- | --- | | Ingestion | Current file not indexed | Is the authoritative source present? | | Retrieval | Required clause absent | Was evidence in the candidate set? | | Context | Near-duplicates crowd out detail | Did the prompt receive enough evidence? | | Generation | Claim exceeds the passages | Does each material claim follow? | | Citation | Link opens wrong version | Can the user verify the source? | ## RAG improves access to evidence, not automatic truth RAG is useful for knowledge that changes, belongs to the organisation or needs citations. It does not replace deterministic databases for exact transaction state, policy engines for permissions or calculation code for financial results. Those systems can be exposed as tools when their output is needed. Retrieval also cannot resolve a source that is wrong or genuinely ambiguous. The application should represent source authority and conflict, and it should decline to synthesize certainty when the evidence does not support it. Human owners remain necessary for high-impact interpretations. - Authorize before adding context. - Preserve source version and exact location. - Evaluate retrieval before blaming the model. - Validate citations for support, not mere overlap. - Use abstention for missing or conflicting evidence. ## Workflow 1. **Define source authority and access.** Identify which repositories are authoritative for each question and who may see them. Preserve document identity, version, effective date and ownership. Apply user permissions before or during retrieval so a relevant but unauthorized passage never reaches the model. 2. **Prepare retrievable content.** Parse headings, tables, lists and metadata without discarding structure. Choose chunk boundaries that preserve complete meaning and retain links to the source location. Detect duplicates, superseded documents and failed extraction before indexing. 3. **Retrieve and assemble context.** Use lexical, semantic or hybrid search with metadata filters and, where useful, reranking. Inspect whether the required evidence appears in the selected context, not merely somewhere in the corpus. Allocate context to diverse supporting passages rather than repeated near-duplicates. 4. **Generate, cite and evaluate.** Instruct the model to answer within the supplied evidence, expose uncertainty and cite source locations. Validate citation entailment and structured fields. Evaluate retrieval recall, context precision, answer correctness, unsupported claims and abstention separately. ## Key decisions - Which source is authoritative when repositories conflict? - What chunk and metadata preserve the meaning needed for this task? - Should retrieval be lexical, semantic, hybrid or tool-based? - How much evidence must be present before the system may answer? - What citation granularity lets a reviewer verify the claim quickly? ## Risks - Unauthorized content enters the prompt through retrieval. - Poor parsing breaks tables and separates qualifiers from their clauses. - Top-ranked passages can be relevant in topic but insufficient for the answer. - The model can cite a real source that does not entail its statement. - Stale duplicates outrank the current authoritative document. ## Metrics - retrieval recall for required evidence - precision and diversity of selected context - answer correctness and completeness - citation entailment for material claims - unsupported answer and appropriate abstention rate - index freshness and extraction failure rate ## Frequently asked questions ### What does RAG mean in AI? RAG means retrieval-augmented generation. An application retrieves relevant external information and gives it to a generative model as context for the answer. ### Does RAG stop hallucinations? No. It can provide better evidence, but retrieval may fail and the model may still make unsupported claims. Evaluate retrieval, answer correctness, citations and abstention separately. ### Is RAG the same as a vector database? No. A vector index can be one retrieval component. A complete RAG system also includes source governance, parsing, permissions, query handling, ranking, context assembly, generation, citations and evaluation. ### When should a company use RAG? Use it when answers need current, private, domain-specific or citable knowledge. Prefer direct database queries or deterministic tools for exact transactions, calculations and rule enforcement. ## Primary sources - [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401), arXiv