What RAG actually changes
A language model does not automatically know your current policies or manuals. Retrieval first finds relevant approved passages, then supplies them to the model with the question, so the answer is grounded in the material you chose rather than in a guess.
Perth heating firm's damage policy: a staff member asks “what if a supplied part arrives damaged?”. RAG retrieves the relevant clause from the approved policy and answers with a source: “Customer Service Policy v4.2 · §7.3”. The answer is traceable.
Question → retrieval → evidence → answerPermissions still matter
If employee A cannot open a confidential folder, a retrieval assistant should not reveal it either. Production retrieval needs permission-aware indexing and querying so that access boundaries are respected.
Freshness is part of accuracy
An answer grounded in an obsolete policy can still be wrong. Track source versions, effective dates and removal of superseded documents. Old knowledge presented confidently is still harmful.
Citations create a verification path
Show the document and passage that support an answer. This lets a person inspect the evidence and lets the system refuse when no suitable source exists — a safe failure rather than a confident guess.
Plain cloud LLM chat
- Answers come from the model's training, not your business
- No sources to open and check
- No enforcement of who may see what
Fine for general drafting; risky as a source of truth about your firm.
RAG over your own documents
- Grounded in the approved files you chose, like the damage policy in the example above
- Every answer carries a citation you can open, such as "Customer Service Policy v4.2 · §7.3"
- Permission-aware at index and query time
The better fit whenever the answer must be checkable.
The distinction matters most where a wrong answer has consequences: contracts, policies, care and compliance. If a person cannot verify the answer, the system is not finished.
Checklist for a safe RAG system
- Index only approved, current documents.
- Apply permissions at index and query time.
- Track source versions and remove superseded files.
- Require citations; allow the system to refuse when evidence is missing.
- Monitor answer quality and correctness.
RAG reduces unsupported answers by grounding, but it does not guarantee truth. Retrieval quality, permissions, source freshness and answer verification all still matter. It is a search of approved knowledge, not a source of truth.
Local firms, care providers, local authorities and professional practices can use RAG to make approved policies and procedures searchable without exposing confidential material. Permission-aware retrieval is a strong fit for sensitive sectors like health, finance and education.
Frequently asked questions
Can I connect RAG to my document files?
Often yes, by indexing approved documents into a retrieval store and controlling what is searchable. Start with a small, well-controlled set.
Is a RAG system safe with confidential documents?
It can be, if access control permissions are enforced and source freshness managed. Without those, retrieval can leak. Permissions are part of the system, not optional.
Do I need to train a model on my data?
No. RAG retrieves relevant passages from your approved documents at request time — the model is not trained on them, and you can keep the model external while keeping your documents under your control.
- ICO — AI and data protection
- NCSC — Secure AI system development
This guide's approach is being built in public. The tooling behind it is open source on github.com/BryanFiFife, so you can inspect rather than take our word for it:
Open source does not mean certified. Read the code, run the tests, and treat it like any tool: check it against your own requirements before you depend on it.
A reliable implementation is not “an LLM with a prompt”. It is a controlled business system with a model only where a model adds value.