Knowledge base
Everything your agent knows lives here: your documents and your data, indexed so the agent finds the right answer and cites exactly where it comes from.
The knowledge base is your agent's memory. You put in what your company knows, the agent indexes it, and every answer draws on a specific passage from your content. The guiding principle is simple: what isn't in your knowledge is never made up. A single space, shared by all your agents, where each agent then chooses what it uses.
Two ways to know
Your knowledge lives in two compartments, depending on the nature of the information. The right choice comes down to one question: text that explains, or data that changes.
| Compartment | What you put in it | How it's queried |
|---|---|---|
| Documents | Unstructured content: policies, FAQs, guides, fact sheets, pages from your website. | Indexed for search by meaning. The agent finds the right passage even when it's phrased differently. |
| Live data | Structured content that changes: catalog, pricing, stock, availability. | Queried in real time through a connector. The answer reflects the current state of the source. |
For anything that's prose, stick with Documents. As soon as a piece of data has to be accurate to the minute (a price, a stock level, an availability), it belongs in Live data, plugged into your existing tool. Setting up connectors is covered in connect live data.
Add a document
A guided assistant lets you add a document in four ways. Nothing is saved until you confirm at the final step: you can cancel at any point without leaving a trace.
| Source | Formats and limits | Review step |
|---|---|---|
| A file | PDF, DOCX, TXT or MD, up to 20 MB. | Added directly. A preview of the indexed text stays available afterward. |
| A web page | The address of a page. | We read the page's actual text and show it to you in full before you confirm. |
| An entire site | The address of the site. | We detect the pages, you check which ones, then you browse their actual content page by page and keep the ones to add. |
| Pasted text | Minimum 50 characters. | Added directly, with a title of your choice. |
For an entire site, empty or blocked pages are discarded automatically, and everything is grouped into a single folder in the library rather than scattered across a multitude of links. A scanned PDF with no real text (just an image) can't be read: use a file whose text is selectable, or paste the content directly. The assistant always ends by connecting the document to the agents that should use it.
One idea per document or per section. Direct answers ("Check-out is at 12 p.m."). And spell out the edge cases: holidays, exceptions, off-season. The agent only returns what's written in black and white. If an answer needs to exist, it has to be somewhere in your content.
How your content is understood and kept up to date
Once added, a document is indexed for search by meaning, not by exact keywords. When a customer asks a question, the agent first pulls a broad set of candidate passages, then a smart ranking keeps only the few passages that truly answer, best first. That ranking reasons about meaning: it links synonyms, crosses languages, and tolerates typos. A customer writing in another language or in shorthand still finds the right answer.
For a web source, you can turn on "Keep up to date automatically" and pick a cadence: never, daily, weekly or monthly. When it's due, we revisit the page and only re-index what has actually changed. An unchanged page therefore costs nothing and stays available without interruption. For a file or text, you update by replacing the file or editing the text, re-indexed as soon as you save.
Citations, for your team only
This is an important point, often misunderstood. Under every AI answer, in "Conversations" as well as in the test sandbox, your team sees a "Sources" line: numbered markers. On hover or tap, a small card shows the document used, the exact excerpt the answer draws on, its freshness ("Updated three days ago") and a link to open it.
The customer never sees these sources. They are strictly reserved for your team. It's a proof of provenance: it lets you check where each answer comes from and spot, where needed, a document to fix, without ever cluttering the conversation on the customer's side.
If an AI answer shows no source, it means it wasn't drawing on any document. On factual questions, that's a signal that some content may be missing. The "Sources" line is therefore as much a trust tool as a gap detector.
Choosing sources agent by agent
One source, shared: each agent chooses what it uses. You manage all your content in one place, the library. Adding, editing, re-indexing or deleting a document happens here, in the Hub, once.
In an agent's editor, the "Knowledge" tab is only there to check the sources that agent is allowed to use. It's a selection, not management: you don't create or delete documents there. A single document can thus serve several agents, and you decide, agent by agent, what each one sees. Editable at any time, without touching the other agents.
Test your knowledge
Before a real customer asks the question, ask it yourself. The "Test your knowledge" bench lets you type a question the way a customer would, choose the agent, and see exactly what it would find: passages ranked by relevance, with the source document, the excerpt and its freshness.
This test replays the real production search, the same engine used in real conversations. The result therefore can't diverge from what your customers get: it's a faithful window, not a rough simulation.
When the test returns "No match", you've found a content gap: add a document that answers this question. It's the same signal surfaced by "Unanswered questions" in Analytics, except that here you trigger it on demand, to close the gaps before they cost you a customer.