Tools and knowledge sources in the AI4CALL assistant

In the AI4CALL assistant configuration you will find two families of resources: tools and knowledge sources. Plugins and custom tools are tools: they let the assistant perform actions. RAG and the Knowledge Prompt are knowledge sources: they provide the information to use in answers. The video above shows how to tell them apart and where to configure them.

The difference is practical: to have an action performed you need a tool that is available and configured; to answer questions about your own content you need the right knowledge source. Let's go through each tab and see where every resource is set up.

Plugins: purchase, configuration and assignment

For plugins, the tab separates three steps: purchase, configuration and assignment. The plugin must be available in your account and have a configured instance. Only then do you choose which tools this assistant can use.

Opening the email service, we see the selected instance and the available tool. You can pick a single tool or select all the tools of that instance. Always check which instance you are assigning: two instances of the same plugin can work on different accounts. In the prompt, explain when to use the tools and which data to collect before calling them.

Custom tools: connecting external services with REST and MCP

With custom tools you connect external services. A REST tool describes a single endpoint. An MCP connection lets you retrieve the catalogue of tools exposed by a server. In both cases the configuration also includes the authentication needed to access the service.

How to create a new REST tool

Let's open New REST Tool. The name identifies the tool and must follow the format indicated by the field. In the example we define an availability check: the instructions ask to collect date and time first, report the outcome and not create bookings. Then you enter the endpoint, method and authentication required by your API.

The body parameters describe the data to send to the service. For each one you define name, type, description and whether it is required. The description must also make the expected format clear: for a date, for example, specify the format required by the endpoint. That way a correct value is never rejected simply because it was sent in a form the service does not accept.

How to configure an MCP tool

In the MCP Tool tab you configure the server, the method, the endpoint and the instructions. The server exposes a catalogue of actions that the assistant can discover. Use the instructions to clarify what that service offers and when to call it. Once the resource is ready, go back to the assistant and assign it.

RAG: answering with your own documents

RAG lets the assistant answer from your own documents: manuals, policies, catalogues. The content is split into fragments and, during the conversation, the fragments relevant to the question are retrieved. Organise your documents into collections dedicated to one topic, then assign the collection to the assistant.

When you create a collection you set the name, the description and how documents are split. You can limit the size of the fragments or choose a number of parts. Fragments that are too small risk losing context; fragments that are too large carry information that is irrelevant to the question. Note that changes to the splitting apply to files uploaded afterwards, as the screen reminds you.

Knowledge Prompt: short, recurring answers

The Knowledge Prompt suits short, recurring answers. You define a default answer and the expressions a customer might use to ask that question; an expression can also have its own specific answer. This content goes straight into the prompt, so it is best to keep it short.

For example, let's create a knowledge entry for office hours: the default answer states the opening days and times. After creating it you can prepare the question variants and assign the knowledge to the assistants concerned. Use verified information and update it whenever it changes.

• Plugins: perform actions; they need to be available in your account, have a configured instance and have their tools assigned to the assistant.
• REST tools: describe a single endpoint, with name, instructions, method, authentication and body parameters.
• MCP tools: connect a server that exposes a catalogue of actions the assistant can discover.
• RAG: uses documents split into fragments and organised into collections dedicated to one topic.
• Knowledge Prompt: short, recurring answers inserted directly into the prompt, with their question variants.

Before saving: check the assigned resources

Before saving, check the assigned resources and the summary on the right. The rule stays the same: to have an action performed you need a tool that is available and configured, to answer questions about your content you need the right knowledge source. Complete the instructions in the prompt and save the assistant.

Frequently asked questions about assistant configuration

What is the difference between plugins and custom tools?

Both are tools that let the assistant perform actions. A plugin must be available in your account and have a configured instance; you then assign the individual tools to the assistant. Custom tools connect external services: a REST tool describes a single endpoint, while an MCP connection retrieves the catalogue of tools exposed by a server.

When should I use RAG and when the Knowledge Prompt?

RAG is for documents such as manuals, policies and catalogues: the content is split into fragments and retrieved during the conversation. The Knowledge Prompt suits short, recurring answers, which are inserted directly into the prompt and should therefore be kept short.

Why must the parameter description state the format?

Because the description of each body parameter must also make the format expected by the endpoint clear. For a date, for example, the required format has to be specified. That way a correct value is never rejected simply because it was sent in a form the service does not accept.