Google AI Studio: temperature, System Instructions, and Function Calling

1The API engine room: governing hyperparameters and declaring functions

The API engine room allows teams to govern hyperparameters and declare functions to orchestrate reliable operational pipelines. Google AI Studio reveals the raw capabilities of Gemini models unencumbered by consumer application guardrails.

On the standard consumer interface, Google applies hidden default system instructions. Inside Google AI Studio, users select the specific model checkpoint (Gemini 2.5 Flash for low latency or Pro for deep reasoning), calibrate temperature (from 0.0 for mathematical fidelity to 1.0 for creative exploration), and leverage two complementary structural engineering mechanisms:

• Structured Outputs: You supply a target JSON Schema. Gemini masks its decoding logits to mathematically guarantee that its final response strictly adopts 100% of the requested schema, eliminating conversational padding.
• Function Calling: You declare a catalog of available tools (function name, description, and parameter JSON Schema). When a user query requires real-time facts, Gemini halts text generation and emits a structured functionCall JSON payload. Your backend application executes the query and returns a functionResponse, which Gemini ingests to synthesize its final conclusion.

Tap to enlarge

2Declaring an inventory lookup function in JSON Schema inside Google AI Studio

Declaring an inventory lookup function in JSON Schema inside Google AI Studio illustrates how to connect Gemini to enterprise databases with zero factual hallucinations.

A logistics manager wants Gemini to answer warehouse stock inquiries using authoritative internal records rather than guessing quantities.

1. Tool declaration (Tools > Function Calling in AI Studio).

In AI Studio's lateral panel, enable Function Calling and specify the following JSON Schema definition:

{
  "name": "check_product_inventory",
  "description": "Queries internal warehouse databases for real-time item availability and scheduled incoming purchase orders.",
  "parameters": {
    "type": "OBJECT",
    "properties": {
      "sku_code": {
        "type": "STRING",
        "description": "Unique product SKU identifier (e.g., SKU-4820 or SKU-9102)"
      },
      "include_incoming": {
        "type": "BOOLEAN",
        "description": "Whether to check pending supplier deliveries scheduled for the current week"
      }
    },
    "required": ["sku_code"]
  }
}

2. User prompt.

Hello, do we have remaining inventory for item SKU-4820 and is any replenishment shipment scheduled for this week?

3. Automated tool invocation by Gemini (functionCall output).

Rather than generating fabricated numbers, Gemini detects the declared tool's relevance and emits this structured payload:

{
  "functionCall": {
    "name": "check_product_inventory",
    "args": {
      "sku_code": "SKU-4820",
      "include_incoming": true
    }
  }
}

4. Backend execution (functionResponse) and final answer.

Your backend application executes the database lookup and injects the ground truth payload: {"status": "in_stock", "available_units": 18, "next_delivery": "2026-09-21", "replenishment_units": 50}.

Gemini then composes the final verified answer: 'Item SKU-4820 currently has 18 available units in the warehouse. An additional replenishment shipment of 50 units is confirmed for September 21, 2026.'

The difference. The model never guessed: it acted as an authoritative dispatcher, parsing natural language entities and grounding its output entirely on verifiable enterprise data.

3Declare a custom business function and evaluate tool invocation inside AI Studio

Declare a custom business function and evaluate tool invocation inside AI Studio to validate agentic workflows before writing integration code, following Model Context Protocol (MCP) tool standards.

To practise without handling internal booking systems, here is the specification of a room reservation tool for fictional company NovaTech Ltd:

[Business Tool Specification - NovaTech Ltd (Sandbox Environment)]
Tool: reserve_conference_room
Description: Registers a meeting room booking across NovaTech Ltd regional hubs.
Expected parameters:
- room_name (string, required): name of the room ("London", "Manchester", "Bristol", "Edinburgh")
- booking_date (string, required): date formatted as YYYY-MM-DD
- headcount (integer, required): expected attendees (must be > 0 and <= 20)
- purpose (string, optional): meeting topic or business purpose

The composer's "Load exercise" button prepopulates the objective with the weak starting prompt:

"Configure the room booking function with required arguments and expected format."

Open aistudio.google.com with Gemini 2.5 Flash (temperature 0.1). In the Tools section, activate Function Calling and declare the reserve_conference_room tool adhering to the JSON schema above (Toolformer architecture, arXiv:2302.04761).

Then submit this user prompt:

"Book the London boardroom for 8 people on September 24, 2026 to conduct our quarterly executive leadership review for NovaTech Ltd."

Self-evaluation rubric: (a) Gemini avoids conversational guessing and emits a structured functionCall object; (b) all three mandatory arguments room_name, booking_date, and headcount are extracted accurately; (c) the structure strictly conforms to the MCP Tool Use specification.

Open the prompt composer