Prepare
Deduplicate, date and classify files, and assign ownership and access rules.
Practical guide
Build local RAG from reliable sources, not from a polished demonstration
A useful retrieval system needs a controlled corpus, permission-aware search and tests that verify the supporting evidence.
Direct answer
Retrieval-augmented generation first finds relevant passages in an authorised document set, then gives those passages to the model as context for its answer. The documents, index and model can be operated in the company's environment, depending on the architecture. This improves traceability but does not eliminate errors.
The retrieval path
Split
Create passages that are precise enough to retrieve while retaining useful context.
Index
Represent passages for search while preserving their origin and metadata.
Retrieve
Select relevant passages after applying the user's access rights.
Answer
Generate from the selected context with citations and refusal rules.
Evaluate
Check whether sources actually support the answer and whether it helps the process.
Begin with ten questions and three measures
Choose ten to thirty real questions with an expected source and known refusal cases. Measure source recall, faithfulness to the cited passage and usefulness for the business task.
A fast answer is not a successful answer if the source is wrong. Evaluation must include outdated documents, conflicting versions and users with different permissions.
Common mistakes
- Indexing every shared folder before selecting a use case.
- Mixing obsolete and current versions without priority rules.
- Hiding results in the interface instead of enforcing retrieval rights.
- Measuring fluency but not supporting evidence.
- Treating local processing as automatic GDPR compliance.
- Automating actions before defining approval and rollback.
Frequently asked questions
Does RAG prevent hallucinations?
No. It can improve access to evidence and make answers easier to verify, but retrieval can miss a passage and the model can misinterpret it.
Should an SME index all its documents?
No. Begin with a limited, useful and maintained corpus whose ownership and access rules are understood.
Sources and references
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks — the research paper that introduced the RAG term
- CNIL recommendations for AI systems and GDPR — official French data-protection guidance
Continue exploring
Start with one measurable process
Choose one use case, a controlled document set and a result that your team can verify.
Request a demo