AI tools

What are AI tools and tool calling?

AI tools are services that an AI model can use to gain new abilities and information. Tool calling is the AI's internal decision-making process where it recognizes that a request requires it to temporarily stop its normal text generation and instead use a tool to assist the user. This allows the AI to fetch real-time data, run code, or take actions in the outside world, making it a much more useful assistant.

You can find AI tools under EbbotGPT > Configurations. Go into your active configuration and set it to prompt version: v2 (the tool calling version) and the AI tools section of the configuration will appear.

Handover to Human

Transfers the conversation to a human agent.

Trigger scenario

This capability enables the AI to recognize and launch your predefined scenarios or workflows. To ensure accurate execution, provide a clear, detailed description for each scenario, then simply activate the tool and select the scenarios the AI is authorized to use.

How to:

1. Add a description to the scenario to make it easy for the AI to understand when it should be triggered

2. Activate trigger scenario tool

3. Select what scenario that can be triggered

Get Website

Use this tool to perform live web scrapes for information that is time-sensitive or frequently changing, ensuring the most current data is acquired on demand.

Allowed URLs: Specify what URLs the tool should be allowed to scrape

Calculator

Use this tool to perform accurate mathematical calculations. This bypasses the LLM's core weakness, which is its reliance on statistical probability rather than logical computation for numerical tasks.

Get User Info

Fetch information that already exist in Ebbot about the user.

Set User Info

Add information about the user to the user's profile in ebbot.

Search dataset

Search for information in the dataset by allowing the AI to create the query and perform the search.

Extra description: In the inputfield you can define when a specific data set should be used

Searchable datasets: Set one or more datasets the LLM should be able to access

Number of results: How many documents should be retrieved (more documents means slower LLM response)

Search Google

Let the AI find information from a google search

End Chat

Let the AI be able to close the chat.

Reject

Let the AI block harmful questions and topics.

HTTP Request

Allow your AI agent to access information on the internet by making HTTP requests to external APIs and websites. This tool allows you to fetch real-time data or send information to third-party services instantly.

Use this tool as a fast way to do light weight integrations. For instance, it can execute a GET request to your product recommendation engine in order to suggest products.

Technical instruction on how to use the http request tool

HTTP Request Tool (http_request)

The http_request tool lets you define one or more preconfigured HTTP requests that the assistant can call as safe, named tools. Each request you configure becomes its own tool (for example, request_search_products).

This chapter explains the JSON config for http_request and how to set it up.

How it works (short version)

• You provide a requests array in the tool config.

• Each request object defines a method, URL, headers/body, and optional placeholders.

• Placeholders are turned into arguments that the assistant must supply.

• The tool performs the HTTP call, then returns a JSON string with metadata and the response data.

Config JSON

The tool config accepts a single key: requests.

{
  "requests": [
    {
      "name": "Search catalog",
      "description": "Search our product catalog by keyword.",
      "method": "POST",
      "url": "https://api.example.com/catalog/search?limit={{limit}}",
      "headers": {
        "Authorization": "Bearer {{token}}",
        "X-Client": "assistant"
      },
      "body": {
        "query": "{{query}}",
        "includeOutOfStock": false
      },
      "placeholders": [
        { "key": "query", "type": "string", "description": "Search term" },
        { "key": "limit", "type": "number", "default": 10 },
        { "key": "token", "type": "string", "description": "API token" }
      ],
      "responsePath": "items"
    }
  ]
}

Request object fields

Field	Type	Required	Description
name	string	Yes	Human-readable name. Used to generate the tool name (e.g. request_search_catalog).
description	string	No	Custom tool description shown to the assistant.
method	string	Yes	HTTP method. Allowed: GET, POST, PUT, PATCH, DELETE (case-insensitive).
url	string	Yes	Target URL. Must be HTTPS and public. You can include placeholders like {{id}}.
headers	object	No	Header key/value map. Placeholders are supported in both keys and values.
body	any	No	Request body for non-GET methods. Placeholders are supported anywhere in the JSON.
placeholders	array	No	Defines the arguments the assistant must provide. See below.
responsePath	string	No	JMESPath expression to extract a specific part of a JSON response.

Placeholders (dynamic arguments)

Use placeholders to insert user-provided values into the request.

• Placeholder format: {{key}}

• Supported in: url, headers, and body (including nested fields and object keys).

• Placeholders create tool arguments for the assistant.

Each entry in placeholders can include:

Field	Type	Required	Description
key	string	Yes	Argument name used in {{key}} placeholders.
type	string	No	Argument type for the schema. Defaults to string. Options: string, number, boolean, object, array.
description	string	No	Helpful explanation for the assistant.
default	any	No	If provided, the argument becomes optional.

Example: optional placeholder with default

{
  "placeholders": [
    { "key": "limit", "type": "number", "default": 10 }
  ]
}

If the assistant does not pass limit, the tool uses 10.

Response selection with responsePath

If the response is JSON, you can extract a nested value using a JMESPath expression. If the path is invalid, the full JSON is returned instead.

Example:

{
  "responsePath": "data.items[0].name"
}

Safety limits and behavior

The tool enforces safe and predictable HTTP behavior:

• HTTPS only.

• No localhost, private IPs, single-label hostnames, or .local domains.

• Max response size: 1 MB.

• Max redirects: 3.

• Timeout: 15 seconds.

• Only certain content types are accepted (HTML, text, JSON, XML, etc.).

• Responses are returned as a JSON string and truncated to 16,384 characters.

The returned JSON string looks like this:

{
  "url": "https://...",
  "method": "POST",
  "status": 200,
  "statusText": "OK",
  "contentType": "application/json",
  "headers": { "content-type": "application/json" },
  "data": { "...": "..." }
}

Tips

• Use clear name values so the generated tool name is understandable.

• Define placeholders for anything that should be dynamic (query, IDs, auth tokens).

• Prefer responsePath to keep responses small and focused.

• For GET requests, body is ignored.

How to see if a tool was used

To see exactly how your AI agent used its tools, go to your chat history. In the top right corner of the AI's reply you can see if a tool was used. Click on that text to see more information about the tool's output and how it was used.