Large Language Models

Tools are usually sent in your request to the provider, tools are nothing more than identifiers that your code can handle if the llm requests it.

An example would be, you want to allow the llm to turn on or off the light in your bedroom. You could pass a tool (literally just the word toggle_light in this case) to the llm request along with whatever your input is (this could be a part of a chat, voice command, whatever) then if the llm determines that the input has somehow warranted the light toggle (say the input is a user saying "turn on the light") it will "use" the tool. Using the tool means literally sending a response back to your code that initiated the required that says "hey run the light toggle code with these parameters" your code is responsible for actually executing that action and telling the llm the response

Tool use is the single most useful thing llms do today imo

it is close to being prompt. You have to look for models which are trianed with tool calling in training (if i am not wrong, mistral 7b 0.3 is one such model, many big models have that, probably phi also, and gemma 3 11/27 b might also have it (this is a thing you can check in their huggingface readme).

If it is in their training, then they can ouptut a tool calling token (i do not remember what is called, but lets say it is [call-tool]). So when you would ask it do some calculation (many models available on chat interface of huggingface have selectable tools you can decide from, enable or disable them, math is one of them) then they would decide if it is appropriate to call a tool (you would not call math tool if you wrote hello). Once they emit a [call-tool] token, then they would prepare some "structred data" (most models do json) lets say prompt was, find integral of sin for 3/4 of its period. And lets consider the model is smart enough to call a tool, and also smart enough to prepare a reasonable structure of it, for example, for maths stuff, they are usually converted to one of either latex or python compatible first, depending on the interface it was trained. lets say this one is trained for latex, then it would generate [call-token] { type_of_tool: "math"; prompt: "$integral_0^(3/2 * pi) sin (x) dx$"; } then the interface (where you are running code) is expected to take over from here, depending on interface, the generation of response would either be frozen, and then resumed, with the result of input now part of current chain, or some would let it continue its yapping, and then pasted a littel afterwards, and then continued to generate output. the json would now be processed by interface, and a relavent tool would be called, which would process the input, and produce its own output (most likely also json), and then given to input.

If you want to implement it yourselves, you can probably take a look at llama cpp implementation, if not use it in llama cpp https://github.com/ggml-org/llama.cpp/blob/master/docs/function-calling.md

also, it would most likely be available in ollama too.

Most agentic model stuff is basically tool-calling in steroids.

Have you seen https://ollama.com/blog/tool-support