Searching Your Knowledge Base

The real power of the knowledge base isn't uploading documents — it's searching them. Instead of scrolling through pages looking for an answer, you ask the AI a question and it finds what you need, citing the source so you can verify.

How to search

Navigate to Knowledge Base → Chat and use the chat input at the bottom. Just type your question the way you'd ask a colleague.

The chat is intentionally simple. There's a:

• Textarea to type your question
• Model selector to pick which AI model answers (default is set by your admin)
• Send button to submit

That's it. No tabs, no toggles, no advanced retrieval modes — the agent decides when to search and how deep to go based on your question.

What good questions look like

Here are some effective ways to ask:

• Specific questions — "What's our refund policy for enterprise customers?"
• Summaries — "Summarize the key points from the Q4 sales report"
• Comparisons — "How does our pricing differ between the starter and pro plans?"
• Lookups — "What's the deadline mentioned in the partnership agreement?"

You don't need special syntax or keywords. Just ask like you're talking to someone who's read all your documents.

Reading the answer

When the agent answers, you'll see:

• The response in plain English, with [1]-style numbered citations inline where it pulled facts from your documents
• Source pills at the bottom of the response — one per cited document, showing the document title and (where available) the page or slide number
• A reasoning trail if the agent did multiple searches to assemble the answer

Click a source pill to jump to the document. Citations always link back to a specific chunk so you can verify the claim.

If the agent can't find a relevant answer in your documents, it'll say so honestly rather than guessing. This is intentional — it keeps responses grounded and trustworthy.

How the agent decides to search

Behind the scenes, the agent has one tool: search_kb. It calls this tool whenever it needs evidence from your documents. For a fresh question, expect a search. For a follow-up that's clearly about the same topic ("What about for contractors?"), the agent often answers conversationally without searching again — it has the relevant chunks in context already.

If the topic shifts, the agent will run a fresh search. You don't manage this; it's just the default behavior.

Per-workspace isolation

The knowledge base is scoped per workspace. The agent can only search documents in your current workspace's index. If you have multiple workspaces (e.g., personal and team), each has its own knowledge base and the agent never crosses between them.

This is enforced at the index level — even if you asked the agent to look at another workspace, it would have no way to do so.

Tips for better results

• Be specific — "What's the SLA for premium support?" works better than "Tell me about support"
• Include context — "In the onboarding guide, what are the first three steps?" helps narrow the search
• Ask follow-ups — The chat remembers the conversation, so you can drill down naturally
• Upload more documents — The more relevant content the AI has, the better its answers
• Check the citation — If a claim feels surprising, click the source pill and verify against the original