Knowledge Base & RAG
Priya messages Acme Diagnostics at 9pm and asks, "What should I not eat before my lipid test?" The answer is in a PDF Acme uploaded weeks ago — and the agent answers from it, in plain words, with nobody on the team awake. This page explains how: how your documents become searchable, and how each agent answers from exactly the material you chose.
It's for admins like Asha who set up what the agent knows. To attach folders to a specific agent, see the Knowledge sub-node.
The Knowledge Base is a search index, not your system of record
Think of the Knowledge Base as search over your documents. Each file is split into passages and indexed by meaning as well as by the words it contains, so the agent finds what a question is about: Priya asks what not to eat, and the agent finds the passage headed "fasting requirements".
Your original files are kept safely in storage; the Knowledge Base is the index laid over them. Knowledge Base files are kept until you delete them — routine data-retention clean-up doesn't remove them. Anything that changes all the time and lives in your own systems — orders, balances, bookings — is better fetched live through an Integration than copied into documents.
Three ways in, one process
| How files arrive | Where from |
|---|---|
| Studio | Upload or drag files and folders onto Connect → Knowledge Base, or keep a folder in sync with a connector. |
| Public REST API | Send files with an API key that has Knowledge Base write permission. |
| Chat attachments | A visitor attaches a file in chat; it's indexed for that customer. |
However a file arrives, it goes through the same steps: the original is stored, the file appears as Pending, it's read, split into passages and indexed in the background, and it turns Active when it can be searched.
Every file is read the right way
Before a file can be searched, its content has to be read out. Text-based files and PDFs with real text are read directly. Scans, images, audio and video are read by AI — tables, charts and handwriting included, and recordings transcribed. Long files are processed in parts, so nothing at the end is lost. The AI model used for this is the KB model on the Knowledge Base page; the default suits most workspaces.
Folders can go further with a use case, which extracts typed fields — prices, dates, severities — so the agent can filter and compare rather than only search. See Extraction & Understanding.
Answers come only from the folders you attach
An agent searches only the folders ticked on its Knowledge sub-node — not your whole Knowledge Base. Folders are live: new files in them are included automatically. Superseded versions are ignored, and a conversation only sees shared documents plus that customer's own.
On the AI Agent node you choose when it searches: Always attach (before every reply, the most reliable) or On demand (only when the agent decides it needs to, cheaper for agents where most turns need no documents). See Retrieval & RAG.
Worked example — Priya's fasting question
Setup. Asha, an admin at Acme Diagnostics, uploads lab-prep-guide.pdf into a Test Preparation folder and ticks only that folder on the agent's Knowledge sub-node. The staff roster and pricing sheets are in other folders she leaves unticked.
Action. At 9pm Priya asks: "What should I not eat before my lipid test?"
Result. The agent replies from the prep guide:
For a lipid (cholesterol) test, please fast for 9–12 hours beforehand. You can drink water, but avoid food, coffee, tea and alcohol during that window. Take your regular medicines unless your doctor advises otherwise.
What just happened. The agent searched only Test Preparation, found the passages about fasting and answered from them. The pricing and roster folders were never searched, so nothing from them could leak into a medical answer.
Live search for connected folders
Some folders don't need to be copied at all. A OneDrive (Microsoft 365) or Google Drive connector can search in place: nothing is downloaded or indexed, and the source is searched live when a question comes in.
- Microsoft 365 native search — searches the connector's OneDrive, SharePoint and Teams folders live in Microsoft 365. Files never leave Microsoft 365.
- Google Drive native search — searches the Drive folder live. It matches keywords only, so it finds less than indexing; choose Download & index when you need the best recall.
Switching a connector to search in place removes any copy that was indexed before, so you never get stale results.
Where to go next
- Knowledge sub-node — attach folders to an agent.
- Knowledge Base Overview — manage folders, files and connectors.
- Retrieval & RAG — how the agent finds the right passages.
- File Upload (widget) — let visitors attach files in chat.