A document assistant needs identifiable, current sources with enforceable access boundaries. Organise the corpus before optimising answer wording.
Give each source an identity
In a fictional internal-policy assistant, similarly named PDFs may describe different rules. Retain ownership, version, effective date and authorised audience. A cited answer can still be wrong when its source is obsolete.
| Check | Required decision |
|---|---|
| Duplicate | Which version is authoritative? |
| Retired document | Remove from retrieval and derived copies |
| Restricted access | Filter for the requesting user |
| Tables and appendices | Verify extraction on a sample |
| Unknown origin | Exclude or obtain owner review |
Evaluate retrieval before generation
Create questions with known relevant documents and check whether the correct passages are found. Better instructions cannot recover a missing source. Include questions the corpus cannot answer, as well as region-specific policies that should not be applied globally.
Can the entire shared drive be ingested?
That shortcut often mixes drafts, archives and confidential material. Start with a controlled collection and a clear update process. Assign responsibility for approving a policy, retiring its predecessor and ensuring the assistant stops using the old version. Corpus maintenance is an ongoing product responsibility, not a one-time upload step.
Regional policy conflicts
A travel policy for one region may contradict another while both documents are valid. Retrieval needs the relevant audience context, and the answer should identify which policy applies. Combining both into a generic global answer can create a misleading instruction.
Sources and further reading
From this guide to your project
AI ↗
We start with a task and a quality criterion. AI belongs where it meaningfully improves the service.
Tell us about your ideaExamples are educational and do not describe client results. Content prepared with AI assistance; see our editorial policy. Our method ↗