Use External Tools, Knowledge

An agent that can only talk is a script with a nicer voice. This page connects it to two things it does not know: your live systems, and your written policies.

Which one you need depends on whether the answer changes:

	Use a tool	Use the Knowledge Base
The answer	Changes per caller, per minute	Is the same every time
Examples	Order status, slot availability, account balance	Return policy, shipping FAQ, price list
Mechanism	An HTTP request the agent makes mid-call	Text retrieved from a document you uploaded

The example on this page is Maya, an order-support agent for a store called QuickMart. She can check what is in a caller’s cart (a tool) and answer questions about returns (a document).

Create the tool
Open the node where the conversation happens — Maya’s main support node — scroll to Tools, and click Create a Tool → External HTTP API. Name it get_cart_details.
The description does all the work. It is the only thing the model sees when deciding whether to call this tool. Write a sentence that says when to use it and what comes back:

Use this tool when the caller asks about cart details. It returns the cart’s products, quantities, total, and discounted total.

A description like “gets cart” leaves the model guessing, and a model guessing about tools either calls the wrong one or calls none.
For the endpoint, this example uses a public sample API so you can try it with no backend of your own:
```
GET https://dummyjson.com/carts/1
```
No authentication, no parameters. It is a GET because it reads rather than changes anything. In your own build this would point at your order system.
Save the tool.
Attach the tool to the node
Creating a tool does not give it to anybody. Go back to Maya’s node, find Tools, and select get_cart_details. Save.
This is the step people skip. An unattached tool exists in your workspace, shows up in the tools list, and will never be called — because the agent is only told about tools attached to the node it is currently on. Attaching per node is deliberate: it stops a checkout tool being reachable during the greeting.
Upload a Knowledge Base document
Click Upload Document, or go to Knowledge Base Files. Pick something your customers actually ask about — a returns policy, a shipping FAQ.
Decibyl asks how the agent should read it:
- Full Document — the agent sees the whole file. Best for anything short. Needs no embedding model.
- Chunked Search — the agent retrieves only the relevant section. For long documents. Configure an embedding model first, under Model Configurations; without one, processing or retrieval fails.
A two-page policy is a Full Document case. Upload it and wait for the status to read Completed before going further.
Attach the document to the node
Same rule as the tool. Back on Maya’s node, scroll to Knowledge Base Documents, and attach the file.
Uploading makes the document available to your workspace. Attaching makes it available to the conversation.

Tell the agent which is which

The model does not infer your architecture. Say it in the node’s prompt:

If the caller asks about cart contents, order items, quantities, total, or
discounts, use the get_cart_details tool before answering. Never guess at
cart contents.

For questions about returns, refunds, or cancellations, answer from the
attached Knowledge Base document.

Save the node.

Test both paths
Start a Test Agent → Test Audio call and try one of each:
- “Hi Maya, what’s in my cart?” — watch the transcript. She decides the tool applies, calls the API, and answers from the response.
- “What’s your return policy?” — she answers from the document instead.
If she answers a cart question without calling the tool, the description is the thing to fix — not the prompt.

Next steps

Full tool reference

Parameters, authentication, timeouts, and what a tool should never be allowed to do.

Full Knowledge Base reference

Chunking behaviour, embedding setup, and when each retrieval mode wins.

Built-in tools

Call transfer, end call, and Google Calendar — no endpoint to write.

Connect an MCP server

Expose a whole set of existing tools to the agent at once.