Gemma 4 Tool Calling Explained: Build AI Agents with Function Calling (Step-by-Step Guide)

Think about asking your AI mannequin, “What’s the climate in Tokyo proper now?” and as a substitute of hallucinating a solution, it calls your precise Python operate, fetches reside knowledge, and responds accurately. That’s how empowering the device name features within the Gemma 4 from Google are. A really thrilling addition to open-weight AI: this operate calling is structured, dependable, and constructed straight into the AI mannequin!

Coupled with Ollama for native referencing, it lets you develop non-cloud-dependent AI brokers. The very best half – these brokers have entry to real-world APIs and companies domestically, with none subscription. On this information, we are going to cowl the idea and implementation structure in addition to three duties which you can experiment with instantly.

Additionally learn: Operating Claude Code for Free with Gemma 4 and Ollama

Conversational language fashions have a restricted data primarily based on after they had been developed. Therefore, they will provide solely an approximate reply if you ask for present market costs or present climate circumstances. This lack was addressed by offering an API wrapper round widespread fashions (features). The goal – to unravel all these questions by way of (tool-calling) service(s).

By enabling tool-calling, the mannequin can acknowledge:

• When it’s essential to retrieve outdoors info
• Determine the right operate primarily based on the supplied API
• Compile accurately formatted technique calls (with arguments)

It then waits till the execution of that code block returns the output. It then composes an assessed reply primarily based on the obtained output.

To make clear: the mannequin by no means executes the strategy calls which have been created by the consumer. It solely determines which strategies to name and the right way to construction the strategy name argument checklist. The consumer’s code will execute the strategies that they referred to as by way of the API operate. On this state of affairs, the mannequin represents the mind of a human, whereas the features being referred to as signify the palms.

Earlier than you start writing code, it’s helpful to grasp how all the things works. Right here is the loop that every device in Gemma 4 will observe, because it makes device calls:

1. Outline features in Python to carry out precise duties (i.e., retrieve climate knowledge from an exterior supply, question a database, convert cash from one foreign money to a different).
2. Create a JSON schema for every of the features you will have created. The schema ought to include the identify of the operate and what its parameters are (together with their sorts).
3. When the system sends a message to you, you ship each the tool-schemas you will have created and the system’s message to the Ollama API.
4. The Ollama API returns knowledge in a tool_calls block quite than plain textual content.
5. You execute the operate utilizing the parameters despatched to you by the Ollama API.
6. You come the end result again to the Ollama API as a ‘function’:’device’ response.
7. The Ollama API receives the end result and returns the reply to you in pure language.

This two-pass sample is the inspiration for each function-calling AI agent, together with the examples proven under.

To execute these duties, you have to two parts: Ollama should be put in domestically in your machine, and you have to to obtain the Gemma 4 Edge 2B mannequin. There are not any dependencies past what is supplied with the usual set up of Python, so that you don’t want to fret about putting in Pip packages in any respect.

1. To put in Ollama with Homebrew or MacOS:

# Set up Ollama (macOS/Linux) 
curl --fail -fsSL  | sh 

2. To obtain the mannequin (which is roughly 2.5 GB):

# Obtain the Gemma 4 Edge Mannequin – E2B 
ollama pull gemma4:e2b

After downloading the mannequin, use the Ollama checklist to verify it exists within the checklist of fashions. Now you can connect with the working API on the URL and run requests towards it utilizing the helper operate we are going to create:

import json, urllib.request, urllib.parse
def call_ollama(payload: dict) -> dict:
    knowledge = json.dumps(payload).encode("utf-8")
    req = urllib.request.Request(
        "/api/chat",
        knowledge=knowledge,
        headers={"Content material-Kind": "utility/json"},
    )
    with urllib.request.urlopen(req) as resp:
        return json.masses(resp.learn().decode("utf-8"))

No third-party libraries are wanted; due to this fact, the agent can run independently and offers full transparency.

Additionally learn: How you can Run Gemma 4 on Your Cellphone: A Fingers-On Information

Fingers-on Process 01: Dwell Climate Lookup

The primary of our strategies makes use of open-meteo that pulls reside knowledge for any location by a free climate API that doesn’t want a key with a purpose to pull the data all the way down to the native space primarily based on longitude/latitude coordinates. When you’re going to make use of this API, you’ll have to carry out a collection of steps :

1. Write your operate in Python

def get_current_weather(metropolis: str, unit: str = "celsius") -> str:
    geo_url = f"
    with urllib.request.urlopen(geo_url) as r:
        geo = json.masses(r.learn())
    loc = geo["results"][0]
    lat, lon = loc["latitude"], loc["longitude"] 
    url = (f"
           f"?latitude={lat}&longitude={lon}"
           f"&present=temperature_2m,wind_speed_10m"
           f"&temperature_unit={unit}")
    with urllib.request.urlopen(url) as r:
        knowledge = json.masses(r.learn())
    c = knowledge["current"]
    return f"{metropolis}: {c['temperature_2m']}°, wind {c['wind_speed_10m']} km/h" 

2. Outline your JSON schema

This offers the data to the mannequin in order that Gemma 4 is aware of precisely what the operate will probably be doing/anticipating when it’s referred to as.

 weather_tool = { 

    "sort": "operate",
    "operate": {
        "identify": "get_current_weather",
        "description": "Get reside temperature and wind pace for a metropolis.",
        "parameters": {
            "sort": "object",
            "properties": {
                "metropolis": {"sort": "string", "description": "Metropolis identify, e.g. Mumbai"},
                "unit": {"sort": "string", "enum": ["celsius", "fahrenheit"]}
            },
            "required": ["city"]
        }
    }

3. Create a question to your device name (in addition to deal with and course of the response again) 

messages = [{"role": "user", "content": "What's the weather in Mumbai right now?"}] response = call_ollama({"mannequin": "gemma4:e2b", "messages": messages, "instruments": [weather_tool], "stream": False}) msg = response["message"]
if "tool_calls" in msg: tc = msg["tool_calls"][0] fn = tc["function"]["name"] args = tc["function"]["arguments"] end result = get_current_weather(**args) # executed domestically
messages.append(msg) 
messages.append({"function": "device", "content material": end result, "identify": fn})
remaining = call_ollama({"mannequin": "gemma4:e2b", "messages": messages, "instruments": [weather_tool], "stream": False}) 
print(remaining["message"]["content"])

Output

Fingers-on Process 02: Dwell Forex Converter

The basic LLM fails by hallucinating foreign money values and never having the ability to present correct, up-to-date foreign money conversion. With the assistance of ExchangeRate-API, the converter can get the newest international alternate charges and convert precisely between two currencies.

When you full Steps 1-3 under, you’ll have a totally functioning converter in Gemma 4:

1. Write your Python operate

def convert_currency(quantity: float, from_curr: str, to_curr: str) -> str:
    url = f"
    with urllib.request.urlopen(url) as r:
        knowledge = json.masses(r.learn())
    fee = knowledge["rates"].get(to_curr.higher())
    if not fee:
        return f"Forex {to_curr} not discovered."
    transformed = spherical(quantity * fee, 2)
    return f"{quantity} {from_curr.higher()} = {transformed} {to_curr.higher()} (fee: {fee})"

2. Outline your JSON schema 

currency_tool = { 

    "sort": "operate",
    "operate": {
        "identify": "convert_currency",
        "description": "Convert an quantity between two currencies at reside charges.",
        "parameters": {
            "sort": "object",
            "properties": {
                "quantity":    {"sort": "quantity", "description": "Quantity to transform"},
                "from_curr": {"sort": "string", "description": "Supply foreign money, e.g. USD"}, 
                "to_curr":   {"sort": "string", "description": "Goal foreign money, e.g. EUR"}
            },
            "required": ["amount", "from_curr", "to_curr"]
        } 
    }
} 

3. Check your resolution utilizing a pure language question

response = call_ollama({
    "mannequin": "gemma4:e2b",
    "messages": [{"role": "user", "content": "How much is 5000 INR in USD today?"}],
    "instruments": [currency_tool],
    "stream": False
}) 

Gemma 4 will course of the pure language question and format a correct API name primarily based on quantity = 5000, from = ‘INR’, to = ‘USD’. The ensuing API name will then be processed by the identical ‘Suggestions’ technique described in Process 01.

Output

Gemma 4 excels at this process. You may provide the mannequin a number of instruments concurrently and submit a compound question. The mannequin coordinates all of the required calls in a single go; handbook chaining is pointless.

1. Add the timezone device

def get_current_time(metropolis: str) -> str: 

    url = f"
    with urllib.request.urlopen(url) as r:
        knowledge = json.masses(r.learn())
    return f"Present time in {metropolis}: {knowledge['time']}, {knowledge['dayOfWeek']} {knowledge['date']}"
time_tool = {
    "sort": "operate",
    "operate": {
        "identify": "get_current_time",
        "description": "Get the present native time in a metropolis.",
        "parameters": {
            "sort": "object",
            "properties": {
                "metropolis": {"sort": "string", "description": "Metropolis identify for timezone, e.g. Tokyo"}
            },
            "required": ["city"]
        }
    } 

2. Construct the multi-tool agent loop

TOOL_FUNCTIONS = { "get_current_weather": get_current_weather, "convert_currency": convert_currency, "get_current_time": get_current_time, } 

def run_agent(user_query: str): all_tools = [weather_tool, currency_tool, time_tool] messages = [{"role": "user", "content": user_query}] 

response = call_ollama({"mannequin": "gemma4:e2b", "messages": messages, "instruments": all_tools, "stream": False}) 
msg = response["message"] 
messages.append(msg) 
 
if "tool_calls" in msg: 
    for tc in msg["tool_calls"]: 
        fn     = tc["function"]["name"] 
        args   = tc["function"]["arguments"] 
        end result = TOOL_FUNCTIONS[fn](**args) 
        messages.append({"function": "device]]]", "content material": end result, "identify": fn}) 
 
    remaining = call_ollama({"mannequin": "gemma4:e2b", "messages": messages, "instruments": all_tools, "stream": False}) 
    return remaining["message"]["content"]
return msg.get("content material", "")

3. Execute a compound/multi-intent question

print(run_agent(
    "I am flying to Tokyo tomorrow. What is the present time there, "
    "the climate, and the way a lot is 10000 INR in JPY?"
))e

Output

Right here, we described three distinct features with three separate APIs in real-time by pure language processing utilizing one widespread idea. It consists of all native execution with out cloud options from the Gemma 4 occasion; none of those parts make the most of any distant assets or cloud.

What Makes Gemma 4 Completely different for Agentic AI?

Different open weight fashions can name instruments, but they don’t carry out reliably, and that is what differentiates them from Gemma 4. The mannequin persistently offers legitimate JSON arguments, processes elective parameters accurately, and determines when to return data and never name a device. As you retain utilizing it, have in mind the next:

• Schema high quality is critically essential. In case your description area is obscure, you’ll have a tough time figuring out arguments to your device. Be particular with models, codecs, and examples.
• The required array is validated by Gemma 4. Gemma 4 respects the wanted/elective distinction.
• As soon as the device returns a end result, that end result turns into a context for any of the “function”: “device” messages you ship throughout your remaining cross. The richer the end result from the device, the richer the response will probably be.
• A standard mistake is to return the device end result as “function”: “consumer” as a substitute of “function”: “device”, because the mannequin won’t attribute it accurately and can try to re-request the decision.

Additionally learn: High 10 Gemma 4 Tasks That Will Blow Your Thoughts

Conclusion

You will have created an actual AI agent that makes use of the Gemma 4 function-calling function, and it’s working fully domestically. The agent-based system makes use of all of the parts of the structure in manufacturing. Potential subsequent steps can embody:

• including a file system device that can permit for studying and writing native information on demand;
• utilizing a SQL database as a way for making pure language knowledge queries;
• making a reminiscence device that can create session summaries and write them to disk, thus offering the agent with the power to recall previous conversations

The open-weight AI agent ecosystem is evolving shortly. The flexibility for Gemma 4 to natively help structured operate calling affords substantial autonomous performance to you with none reliance on the cloud. Begin small, create a working system, and the constructing blocks to your subsequent initiatives will probably be prepared so that you can chain collectively.