Prepare a document inventory
Record each document's title, business purpose, version, effective date, visibility and maintenance owner. Prioritize current documents with known sources. Identify duplicates, drafts and superseded policies separately so one question does not retrieve conflicting passages.
Check supported formats, file sizes and parsing methods for your release. For scans, tables embedded in images and manuals with complex layouts, validate representative pages first. A filename extension alone says little about the quality of the extracted content.
Track each processing stage
| Stage | Result to inspect |
|---|---|
| Ingestion | The complete source file is readable and has an identifiable version. |
| Parsing | The body text, figures, tables and reading order can be checked. |
| Indexing | Usable passages retain source locations and access information. |
| Retrieval | Test questions return relevant material. |
| Citation | The reader can open the cited location with their own permissions. |
Test a small sample before expanding
- Choose pages containing headings, body text, tables and illustrations as a test set.
- Ask separate questions about locating facts, interpreting rules and relating text to an illustration.
- Check that the answer uses the correct version and that the cited passage contains the supporting evidence.
- Repeat the tests with an identity that cannot read the material. Check both retrieval and citation access.
Verify updates and deletion
After replacing a file, verify that new content can be retrieved and that any older passages are identified by version. After deletion or access revocation, also check how caches, earlier citations and exports are handled.
If an answer still cites old material, check the file version and index status, then check whether the question reused an earlier conversation. Record the question, expected version, actual citation and time of occurrence to diagnose the cause before uploading the file again.
