Alibaba Cloud Model Studio:Assistant API

What is an assistant

An assistant is a type of AI conversational assistant characterized by:

• Multiple models: Configure your assistant with any foundational model and customize its personality and capabilities using system instructions.

• Tool calling: Use official tools like Python code interpreter or custom tools via function calling.

• Conversation management: Thread objects store message history and truncate it when exceeding the model's context length. Create a thread once, then append messages as users respond.

To experience an assistant, create an agent application in the console (no coding required). For programmatic access, see Get started with Assistant API for step-by-step integration guidance.

Differences from the text generation API

The text generation API's primary element is messages generated by models like Qwen-Plus and Qwen-Max. This lightweight API requires manual management of conversation state, tool definitions, knowledge retrieval, and code execution to build applications.

On top of the text generation API, the Assistant API introduces the following core elements:

• Message objects: Store conversation role and content (similar to text generation API messages).

• Assistant objects: Bundle a foundational model, instructions, and tools.

• Thread objects: Represent the current conversation status.

• Run objects: Execute the assistant on a thread, including text responses and tool calls.

The following sections explain how these elements work together.

Interaction method

The Assistant API uses threads to process messages sequentially, preserving conversation continuity. The process is as follows:

1. Create a message: Create a message with Message.create() and assign it to a thread to maintain context.

2. Initialize a run: Run.create() initializes the assistant runtime for message processing.

3. Wait for results: wait() pauses until the assistant completes processing and returns results.

Consider a simple drawing assistant as an example:

User input: The user inputs two messages: Input 1: "You are a painter" Input 2: "Draw a cat" Message creation: Each input is wrapped in a message and sent to the thread via Message.create(). Thread processing: The thread creates a session, receives messages, and forwards them to the assistant via Thread.create(). Assistant processing: The assistant object processes each message based on the provided context: Message 1 ("You are a painter") is identified as user input. Message 2 ("Draw a cat") is also identified as user input. Assistant response: The assistant processes user messages and generates Message 3 (assistant role) with the cat picture URL. Execution and waiting: Create and start the run with Run.create(), then wait for completion with Run.wait().

Model support

Tool support

Tool	Identifier	Description
Python code interpreter	code_interpreter	Execute Python code for programming tasks, math, and data analysis.
Image generation	text_to_image	Transform text descriptions into images to diversify response formats.
Custom plug-in	${plugin_id}	Integrate custom business interfaces to extend AI capabilities.
Function calling	function	Execute functions on local devices without requiring external network services.

Get started

To test LLMs or get started with the Assistant API:

• Playground: Test the inference capabilities of LLMs to find the most suitable one for your assistant.

• Quick Start: Basic usage and examples to help you get started.

• API references: Parameter details for resolving development challenges.