Support agent on your own knowledge base
Publish an agent that answers questions from the company's internal documentation.
This is the platform's most common use case: an Agent that answers recurring questions about the expense policy, the vacation process or the product catalog, from the documents the company already has.
This guide goes beyond your first agent: it covers what separates an agent that demos well from one that holds up under real use.
Prerequisites
- The Owner, Curator or Editor role in the workspace.
- At least one model enabled under Settings › Models.
- The documents that back the answers, in a format with selectable text.
Scope the collections
Split by subject, not by origin
One Collection per topic, such as "HR Policies" or "Product Catalog", works better than a single Collection with everything.
Splitting by subject makes search more precise and lets you take one set out of circulation without affecting the rest.
Discard what is outdated
Contradictory documents are the most common way for an Agent to give wrong answers that look right. Two versions of the same policy, both indexed, and the Agent cites the old one.
Before uploading, confirm which version of each document is current.
Wait for indexing
Check under Collections › Files that every document has reached the Active state. Documents still processing don't take part in answers, and the Agent screen gives no warning about it.
Write the prompt
The form's default text covers tone and format. For a support agent, add three things:
- The role and the audience: who it serves and what it answers about.
- The order of priority: answer from the retrieved documents, not from the model's general knowledge.
- The behavior when there's a gap: say it didn't find the answer, and point to whom to ask.
The third one saves the most rework. Without it, the model fills the gap with what it learned in training, and the answer comes out plausible, with no references, and wrong.
Test before publishing
Test with real questions, not the ones you would write. Three categories that reveal problems:
- Ambiguous questions, which could refer to two documents.
- Out-of-scope questions, to check that it declines instead of making things up.
- Questions whose answer changed recently, to check that the right version is indexed.
For each answer, open the References block and check that the cited passage supports what was said.
Publish and monitor
Check Publish on save. The Agent then shows up in the Chat selector for the workspace.
In the first weeks, two screens tell you whether it is working:
- Logs › Conversations shows what people actually ask, which is rarely what you imagined.
- Logs › Response feedback gathers the cases that went wrong.
Reading these two usually yields the second version of the prompt and a list of missing documents.
Next steps
- Adjust the instruction text in Write the prompt.
- Cap the agent's cost in Quotas.