Internal support
IT, HR, finance — sourced answers from policies.
What is RAG? Definition and explanations
RAG (Retrieval-Augmented Generation) combines document retrieval with a generative model: the system first fetches the relevant passages from your own sources, then generates an answer grounded on those excerpts. The benefit is twofold — fewer inventions, and answers traceable back to the source document. Its quality depends first on the retrieval step: an excellent model never rescues poor excerpts.
In one sentence
RAG grounds a model's answer on passages retrieved from your own sources.
Key points
- Two phases: retrieval (often vector + keyword) then conditioned generation.
- Quality depends on chunking, metadata, and access control.
- Citations let you verify sensitive claims.
- No guarantee if sources are incomplete or stale.
- Complements fine-tuning and system prompts.
Term at a glance
- RAG
- Retrieval-Augmented Generation
- French term
- Génération augmentée par récupération
- Domain
- Artificial intelligence
- Category
- Applied generative AI
- Level
- Intermediate to advanced
Why RAG became essential
Quality depends on indexing, document chunking and passage ranking.
Sources are often cited so users can verify.
It is the preferred approach for internal knowledge bases, HR policies or product docs.
Building a RAG pipeline
- 01
Curate sources
Approved docs, dated versions, exclude unvalidated drafts.
- 02
Index well
Coherent chunks, metadata, recall tests on typical questions.
- 03
Secure queries
Filter chunks by user identity; log questions and served passages.
- 04
Evaluate continuously
Faithfulness metrics, FAQ coverage, expert feedback loops.
A concrete RAG example
An employee asks about the expense policy; the system retrieves the current HR page and answers with a link to the source.
RAG use cases
Sales
Specs and lead times from PIM/ERP exports.
Compliance
Query registers with audit trail.
Assisted legal research
First-pass triage, not legal advice.
Benefits and limits of RAG
- Updatable without costly retraining
- Traceability via citations
- Can merge multiple repositories
- Bad index → silently wrong answers
- Added latency
- Demands document governance
RAG or fine-tuning?
| RAG | Fine-tuning | |
|---|---|---|
| Goal | Inject up-to-date facts and docs | Change model style or behaviour |
| Update cost | Reindex documents | Training runs |
| Traceability | Strong with mandatory citations | Weak — weights dispersed |
| Data needed | Searchable corpus | Labelled example sets |
RAG and business value
RAG turns a generic LLM into a front-end for your document estate — if that estate is trustworthy. Start with a narrow scope and weekly eval rather than indexing all of SharePoint at once.
Frequently asked questions
Does RAG remove hallucinations?
It reduces them when sources cover the question and prompts enforce fidelity. Extrapolation still happens.
Do you need a vector database?
Often for semantics, sometimes plus BM25. Depends on volume, language, budget.
RAG with agents?
Yes — the agent calls the retriever as a tool before other actions.
Related terms
Sources and references
Want an internal assistant that cites your docs — not random web text?
Design your RAG