Tool and function calling

Tool Calling

Tool calling enables large language models (LLMs) to access external tools. The model doesn't directly execute these tools; it first recommends tools to invoke. The user then independently calls the tool externally and returns the result to the model. Finally, the model integrates the result to generate a response to the user's original query. Shengsuan Cloud Router standardizes this tool-calling interface across all models and providers.

ToolCall Structure
Represents tool invocation details when an LLM needs to call external functions/plugins during generation:

Index: Call sequence in chat streams (optional)

ID: Unique identifier for the tool call

Type: Distinguishes between different tool-calling logic

Function: Described by FunctionCall with function details

Function Calling

When using text (chat) models via API, you can define functions for the model to optionally call. The model then outputs a JSON object containing necessary parameters. The text model API does not execute functions—it only outputs JSON that you can use to call functions in your own code.

While modern models can autonomously determine when to call functions and generate JSON matching function signatures, this carries inherent risks. We strongly recommend implementing user confirmation steps before executing real-world actions (e.g., sending emails, online publishing, purchases).

FunctionCall Structure
Describes called function details:

Description: Explains the call's purpose

Name: Identifies the specific interface

Parameters: Description of required arguments (typically JSON)

Arguments: Actual parameter string passed to the function

Common Use Cases

Function calling reliably extracts structured data from models. Examples include:

Building API-connected assistants

Example functions:
send_email(to: string, body: string)
get_current_weather(location: string, unit: 'celsius' | 'fahrenheit')

Converting natural language to API calls

Example conversion:
"Who are my top customers?" → get_customers(min_revenue: int, created_before: string, limit: int)

Follow with internal API call

Extracting structured data from text

Example functions:
extract_data(name: string, birthday: string)
sql_query(query: string)

Basic Function Calling Workflow

Call the model
Send user query + define available functions in functions parameter

Model responds
Model may choose to call ≥1 functions. If calling:

Outputs JSON matching your schema

⚠️ Models may hallucinate parameters

Parse & execute

Parse JSON string in your code

Call functions using provided arguments

Resubmit to model

Append function results as new messages

Let model summarize results for user

Note: Shengsuan Router normalizes variations in function-calling implementations across different model providers.

Tool Calling#

Function Calling#

Common Use Cases#

Basic Function Calling Workflow#

Tool Calling

Function Calling

Common Use Cases

Basic Function Calling Workflow