I went to an awesome workshop at our local hackspace about Google Gemini and function calling. I can only say thanks to Merchi for giving it; I learned a ton of things regarding Gemini models and AI in general.
During the workshop, we tried to see how to chain functions when multiple items need to pass through the function. We used a prompt like:
"Do you have stock of iPhone 8 or Google Pixel 9?"
The function defines the way to get the stock for a given product.
tools = [
{
"type": "function",
"function": {
"name": "retrieve_stock_for_item",
"description": "Get the stock information for the given product",
"parameters": {
"type": "object",
"properties": {
"product_name": {
"type": "string",
"description": "The product name to check the stock."
}
},
"required": ["product_name"]
}
}
}
]
A normal user, who reasons logically, will check the function for two items. However, Gemini only returned one function call for one product. I tried with LLaMA 3.1:8b, and it detected both products and returned both function calls without much issue (See the tool_calls field).
{
"id": "chatcmpl-514",
"object": "chat.completion",
"created": 1726234612,
"model": "llama3.1:8b",
"system_fingerprint": "fp_ollama",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_e0f4geqf",
"type": "function",
"function": {
"name": "retrieve_stock_for_item",
"arguments": "{\"product_name\":\"Iphone 8\"}"
}
},
{
"id": "call_cbnhz6qu",
"type": "function",
"function": {
"name": "retrieve_stock_for_item",
"arguments": "{\"product_name\":\"Google Pixel 9\"}"
}
}
]
},
"finish_reason": "tool_calls"
}
],
"usage": {
"prompt_tokens": 172,
"completion_tokens": 46,
"total_tokens": 218
}
}
And the full code if you want to try it:
import requests
import json
model="llama3.1:8b"
tools = [
{
"type": "function",
"function": {
"name": "retrieve_stock_for_item",
"description": "Get the stock information for the given product",
"parameters": {
"type": "object",
"properties": {
"product_name": {
"type": "string",
"description": "The product name to check the stock."
}
},
"required": ["product_name"]
}
}
}
]
messages = [
{
"content": "Do you have stock of Iphone 8 or Google Pixel 9?",
"role": "user"
}
]
data={
"messages": messages,
"tools": tools,
"model": model,
"temperature": 0
}
response = requests.post('http://ollama.acalustra.local:11434/v1/chat/completions',
headers={"Content-Type": "application/json"},
json=data)
print(json.dumps(response.json(), indent=4))
It's fascinating that this can be achieved with just 40 lines of code. Truly fascinating!
Thanks Merchi for the workshop!