Skip to main content
A knowledge base lets your agent answer from your own documents, such as a price list, a policy or a product manual, instead of only what’s in its prompt. Use it when the facts are too long for the prompt or change often.

​
How it works

Before each reply, Autopilot searches your documents for the passages closest to what the caller just said. It gives the two best passages to the model for that reply only. They never go into the conversation history, so the next turn gets a fresh search. The search has 800 ms. If it takes longer, the agent answers without it, so a slow search never holds the caller up. Autopilot reads your documents when each call starts. On a call that begins with a question, the first answer can come before the documents are ready. In that case the agent answers from its prompt alone. The caller asks “How much is the premium plan?”. Autopilot searches your PDFs in the workspace bucket for up to 800 ms (slower than that, the agent answers without it) and gets the two best passages. They join the system prompt and the conversation history in the model request for this reply, the agent speaks, and the passages are dropped after the reply, never added to the history. The caller asks “How much is the premium plan?”. Autopilot searches your PDFs in the workspace bucket for up to 800 ms (slower than that, the agent answers without it) and gets the two best passages. They join the system prompt and the conversation history in the model request for this reply, the agent speaks, and the passages are dropped after the reply, never added to the history.

​
Who sets up what

A knowledge base needs storage for the documents, a document parser and an embeddings model. The operator of the install sets these up once. You, the developer, list the documents your agent uses. If the operator hasn’t set up knowledge storage, Autopilot ignores the list, and your agent answers from its prompt alone.

​
Before you begin

You need:
  • An Autopilot agent. See Build an Autopilot agent.
  • Your documents as PDF files. Other formats aren’t supported.
  • A way to put files in your workspace’s storage. Each workspace has its own bucket, named after the workspace ref in lowercase. Fonoster has no upload in the dashboard, the SDK or the CLI, so someone with access to the storage puts the files there.

​
Add documents to your agent

1

Store the documents

Put each PDF in your workspace’s bucket. Note each file’s key (its path in the bucket), such as kb/price-list-2026.pdf.
2

List them in the agent

Add an entry for each document to languageModel.knowledgeBase. Each entry has three fields:The dashboard doesn’t show the knowledge base, but it keeps the list when you save the agent there. Set it with the SDK or the CLI.
Add the list to the agent’s languageModel:
3

Tell the model how to use it

Add a line to the system prompt so the model leans on the passages and doesn’t guess:
4

Try it

Call the agent with Talk to agent and ask about something that’s only in your documents. Give the call a moment after it connects, so the documents are ready.

​
Tips

  • Keep documents focused. A short, current price list answers better than a long catalog where prices are buried.
  • Put facts that every call needs, such as opening hours, in the prompt instead. The prompt is always there; a search can miss.
  • If you need live data, such as an order’s status, use an HTTP tool. A knowledge base is for documents that change rarely.
Every knowledge base field is in Autopilot configuration.

​
Next steps

Give your agent tools

Fetch live data from your API during the call.

Write prompts for voice

Tell the model when to rely on your documents.