What is function calling (tool use)?

Function calling: definition

Language models know only their training knowledge and the texts supplied to them. To query current data or trigger actions, they need tools. Function calling is the mechanism for that: developers describe the available functions to the model with name, purpose and parameters, usually as a JSON schema.

The model does not run the function itself. It only decides which function with which values makes sense and returns a structured call. The software checks it, executes it and returns the result. This separation is important for security and control.

How a call works

A customer asks in chat: “When will my order arrive?” The model recognizes that it needs the function “retrieve order status” and generates the call with the order number from the conversation. The application queries the inventory system and returns the status. From that, the model formulates the answer.

Several functions can be used one after another or in parallel. AI agents are built on this principle. The same technique is also used to get outputs in a fixed format, such as fields from an email as JSON that are written straight into a database. This saves error-prone text parsing.

Real-world example: an assistant for furniture retail

A furniture retailer uses an assistant on its website. It has three functions: check an item’s availability, retrieve order status and create a callback request in the ticket system. It answers questions with real data from the systems instead of guesses. The customer gets an answer that matches the current stock level, and employees are relieved of routine inquiries.

Read-only functions run automatically; the callback request is created only after the customer confirms. The assistant sees only the data that belongs to the logged-in customer. Every call is logged so that it can later be traced what happened. New functions can be added later without rebuilding the assistant.

Security, limits and imple­menta­tion

The model can supply wrong or manipulated parameters, for example when a third party hides instructions in inputs through prompt injection. Therefore: validate arguments, apply the rights of the respective user, have actions with consequences approved and limit calls. Functions should also be described clearly and unambiguously.

Nano Concepts designs functions, interfaces and permissions so that AI assistants access your systems in a useful but controlled way. This includes a clean breakdown of functions, a validation layer between model and system, and tests with typical as well as deliberately faulty inputs before the assistant goes into production.

• Keep functions small, clearly named and well documented
• Check parameters on the server before execution
• Use the rights of the logged-in user, not blanket system rights
• Have write actions confirmed
• Log calls and set limits