THE PRACTICAL GUIDES
RAG consultants in Australia: brief knowledge search with answers you can check
An AI knowledge-search project should help a named user answer a defined set of questions from approved material. A fluent answer with a link is not enough: the cited passage must support the claim, the user must be permitted to see it, and a missing answer must remain missing. Put those conditions into the supplier brief before choosing a search platform or model.
This guide helps Australian buyers commission retrieval-augmented generation, usually shortened to RAG. It focuses on a custom knowledge-answering service, rather than the separate Microsoft workplace rollout decision. Browse implementation providers for the broader inventory.
Name the question and the usable evidence
AWS's RAG explanation describes retrieving information from an external source and supplying it with the question to a language model. It also discusses updating that information. This architecture does not, by itself, establish that your documents are current, permissions are correct or every answer is supported.
A bounded first brief could be: “Help our service team find current product-care instructions, quote the relevant passage and open the approved source.” That is different from an assistant allowed to change a customer record or promise a refund. Start with finding and explaining information; specify any later action separately.
Define the source collection, audience, supported languages, document formats and update process. Identify the owner who decides which version is authoritative. If those foundations are unresolved, use the data-readiness guide to scope that work first.
Two advertised capabilities to discuss
| Existing provider | Published scope | Question for the proposal |
|---|---|---|
| Ippon Australia | Intelligent search, knowledge discovery and retrieval-augmented generation, alongside evaluation and production delivery. | Which source, citation, permission and release checks will be delivered for our specific collection? |
| DiUS | Production AI applications including RAG, data access patterns, evaluation and operating handover. | Who maintains ingestion, access boundaries and answer-quality checks after the initial build? |
Ippon's service page and DiUS's capability page establish advertised scope. They do not demonstrate performance on your questions or establish a comparable fixed fee. Ask each supplier to quote the same collection, users, evaluation work and support period.
Separate a working link from a supported answer
Use this fictional example in a proposal review: the assistant says an appointment can be cancelled with 24 hours' notice and links a policy. The linked policy actually says 48 hours. The link opens successfully, but the answer is still wrong.
Ask the supplier to show the cited passage, document identifier and version beside each important claim. Establish how a reviewer records an unsupported statement, a wrong source, a missed answer and a legitimate “not found”. Checking that a URL exists is a different test from checking that its contents justify the answer.
Specify what happens with conflicting policies, an unreadable attachment or a question outside the collection. A useful response can identify the missing information and route the question to an owner. An invented answer should not be accepted merely because it sounds helpful.
An original twelve-case commissioning pack
Prepare an approved synthetic collection with invented product instructions, one withdrawn document and a restricted file. Agree the expected result for each test identity before any later trial. The following is a fictional exercise, not a system we built or supplier score:
- Six answerable questions: include ordinary wording and paraphrases. Each answer needs the correct current passage, not just a relevant-looking document.
- Two questions with no answer in the collection: expect an explicit limit and the agreed human route.
- Two questions about restricted material: use an identity without access. The response must follow the agreed permission boundary, including what may be revealed about the file's existence.
- One question whose only source was withdrawn: after the agreed removal process, check that the old material is no longer used and the question is handled appropriately.
- One document containing an irrelevant instruction: check that source text is treated as evidence to inspect, rather than permission to change the assistant's task or disclose other material.
Six plus two plus two plus one plus one gives twelve cases. Retain the question, identity, expected outcome, retrieved passages, final response, version and reviewer decision. Report failures by kind instead of hiding permission leaks or confident unsupported claims inside one average. The buyer must agree the real acceptance criteria and broader test coverage.
Make refresh and maintenance visible
The handover should let an owner add, correct and withdraw an approved document, see when the change becomes searchable, and recognise an incomplete or failed refresh. Ask what happens to stored copies and citations after deletion or an access change. A successful upload is not evidence that the answer service has applied the change.
Price the initial source preparation, connectors, evaluation and interface separately from continuing search, model usage, storage, monitoring and support. Record volume assumptions and the cost of another collection or permission model. Ask for ownership, export, recovery instructions and a way to pause unreliable answers while people use the original sources.
Sources reviewed 25 October 2026, Melbourne time. Provider capabilities remain published claims. No business collection, account or model trial was accessed. Use the quote worksheet, the pilot handover guide or all AI buying guides.
Editorial guidance with linked sources. Reviewed 2026-10-25.