Internal document search that works on a messy corpus
Most of this project is document work, not tool work.
AI Tools Hub Editorial · published · updated
The first failure
Answers grounded in a withdrawn policy. A person would notice it looks old; a retrieval system does not discount by date. The same happens with discontinued product specs, superseded price lists and procedures written by someone who left.
Narrow the scope
| Include | Exclude |
|---|---|
| Current policies with a named owner | Personal folders |
| Current product specifications | Finished project material |
| FAQ and help articles | Email attachments |
| Current price lists and contract forms | Work-in-progress drafts |
Verify permissions as a user, not an admin
Confirm the search layer honours the source permissions by searching for something sensitive from an ordinary employee account. Checking as an administrator proves nothing, and HR or salary documents surfacing company-wide is the common accident.
Always show citations
Configure answers to cite the document and location. Users can then verify, and a citation to an obsolete file is visible. With citations the system is usable before the corpus is perfect.
Evaluate on twenty real questions
Collect twenty of your most frequent internal questions and count correct answers. A vendor demo tells you nothing about your corpus.
FAQ
How long does the document work take?
Days, if you narrow scope. Forever, if you try to organise everything — so decide the boundary first.