Knowledge retrieval

Trusted answers from your own documents.

Most organisations already have the answers their staff need. They're buried in policies, procedures, contracts and old email. Retrieval augmented generation (RAG) lets a language model search your documents and answer with citations, so people can check the source instead of trusting the model.

Opportunities

Policy and procedure search

Staff ask questions in plain English and get an answer that links to the exact clause in the current policy.

Onboarding

New starters find how things are done without waiting for someone to be free.

Technical knowledge

Engineering standards, manuals and past project records become searchable by meaning, not just keywords.

What to get right

Access control

The assistant must only retrieve documents the person asking is allowed to see. Retrieval that ignores existing permissions is the most common way these systems leak information internally.

Privacy Act 1988

If your documents contain personal information, using them in a new system is a use under the Australian Privacy Principles. Check it is related to the purpose you collected it for (APP 6) and keep it secure (APP 11).

Accuracy

Answers can still be wrong. Show citations, make it easy to report bad answers, and keep humans responsible for decisions.

This is general information, not legal advice. See our Australian AI compliance guide for the wider landscape and links to official sources.

Where to run it

Retrieval systems handle your most sensitive internal content, so they are one of the strongest cases for running the model and the vector database inside your own network. A modest open weight model is often enough, because the answer comes from your documents rather than the model's memory.

First steps

  1. Pick one well maintained document set with a clear owner, such as HR policies.
  2. Map who can see which documents today, and carry those permissions into the index.
  3. Test with real questions from staff and measure how often the cited source actually supports the answer.
  4. Set a review cycle so outdated documents are removed from the index.

Related guides