Knowledge Base
Route: /org/knowledgebase
Permission: View Knowledge Base Articles and Categories — or View Project Trainings
Adding and deleting sources needs Add, Edit, and Delete Knowledge Base Articles and Categories.
1. What this actually is
Despite the name, this screen manages AI training sources: documents you ingest so that ticket deflection, the assistant and AI workflows can answer from your own material.
The help articles your customers read at /support/kb are a different thing entirely, authored
under Settings → Knowledge Base. Two screens, two purposes, one name.
The shared permission with VOX (View Project Trainings) is the clue: the same ingested corpus feeds the voice agent.
| This screen | Settings → Knowledge Base | |
|---|---|---|
| Audience | Machines — AI retrieval | Humans — customers reading |
| Content | Whole documents you upload | Articles you write |
| Unit | Chunks for embedding | Articles and categories |
| Published to | The AI layer, and VOX | /support/kb (HELP_006, HELP_004) |
| Formatting matters? | No — only extractable text | Yes |
The page states its purpose: "Manage articles, FAQs, and documentation to deflect tickets and guide AI workflows."
2. How training sources spread across the application
A source ingested here is answerable material everywhere the AI layer reaches — which is the reason to curate rather than dump.
3. The screen
| Column | Meaning |
|---|---|
| Source name | The document's title, with an extract of its opening text beneath |
| Type | What was ingested — e.g. PDF Document |
| Chunks | How many segments the document was split into for retrieval |
| Status | Active once ingestion completed and the source is searchable |
| Added on | Ingestion date |
| Actions | Delete — removes the source and its chunks from retrieval |
Search sources… filters the list. + Train Agent ingests a new document.
4. Chunks — the number to actually read
A document is split into chunks before embedding; retrieval matches a chunk, not a whole file.
Check it immediately after ingesting. A long document that yields very few chunks means the text
did not extract — a scanned PDF with no text layer is the usual culprit. It will sit there
marked Active, contributing nothing, and you will not find out until the assistant fails to
answer something you know is in the file.
Rule of thumb: if the chunk count looks low for the document's length, re-check the source.
5. Curation
Ingesting a superseded policy or a retired runbook makes the assistant confidently wrong, which is worse than it having no answer. The AI cannot tell that a document is out of date.
Delete a source when it is retired: that removes its chunks from retrieval immediately. Treat this list as a curated corpus, not an archive.
6. Don't confuse this with…
| Screen | What it is |
|---|---|
| Knowledge Base (this page) | AI training sources. Machine-facing |
| Settings → Knowledge Base | Customer help articles, published to /support/kb |
| Ticket → Knowledge Articles | Articles linked to a specific ticket |
| Papers | Document management. Unrelated to either |
- A customer should be able to read it → Settings → Knowledge Base
- The assistant should be able to answer from it → this page
- Both → do both. They are not connected; ingesting an article here does not publish it, and publishing an article there does not train anything
7. Troubleshooting
| Symptom | Cause |
|---|---|
| Assistant cannot answer from a document you uploaded | Check the chunk count — likely a scanned PDF with no text layer |
| Assistant gives outdated answers | A superseded source is still ingested. Delete it |
| Customers cannot find an article you added here | This screen does not publish anything. Use Settings → Knowledge Base |
Source stuck without Active | Ingestion did not complete — re-add it |