Prompt Engineering

The prompt engineering is a crucial part of LLM-empowered applications, especially for the multi-agent ones. However, most API providers focus on the chatting scenario, where a user and an assistant speak alternately.

To support multi-agent applications, AgentScope builds different prompt strategies to convert a list of Msg objects to the required format.

Note

There is no one-size-fits-all solution for prompt crafting. The goal of built-in strategies is to enable beginners to smoothly invoke the model API, rather than achieve the best performance. For advanced users, we highly recommend developers to customize prompts according to their needs and model API requirements.

Using Built-in Strategy

The built-in prompt strategies are implemented in the format method of the model objects. Taking DashScope Chat API as an example:

from agentscope.models import DashScopeChatWrapper
from agentscope.message import Msg
import json


model = DashScopeChatWrapper(
    config_name="_",
    model_name="qwen-max",
)

# `Msg` objects or a list of `Msg` objects can be passed to the `format` method
prompt = model.format(
    Msg("system", "You're a helpful assistant.", "system"),
    [
        Msg("assistant", "Hi!", "assistant"),
        Msg("user", "Nice to meet you!", "user"),
    ],
)

print(json.dumps(prompt, indent=4, ensure_ascii=False))

[
    {
        "role": "system",
        "content": "You're a helpful assistant."
    },
    {
        "role": "user",
        "content": "## Conversation History\nassistant: Hi!\nuser: Nice to meet you!"
    }
]

After formatting the input messages, we can input the prompt into the model object.

response = model(prompt)

print(response.text)

Nice to meet you too! How can I assist you today?

Non-Vision Models

In the following table, we list the built-in prompt strategies, as well as the prefix of supported LLMs.

Taking the following messages as an example:

Msg("system", "You're a helpful assistant named Alice.", "system"),
Msg("Alice", "Hi!", "assistant"),
Msg("Bob", "Nice to meet you!", "user")

LLMs	model_name	Constructed Prompt
OpenAI LLMs	gpt-	[ { "role": "system", "name": "system", "content": "You're a helpful assistant named Alice." }, { "role": "user", "name": "Alice", "content": "Hi!" }, { "role": "user", "name": "Bob", "content": "Nice to meet you!" } ]
Gemini LLMs	gemini-	[ { "role": "user", "parts": [ "You're a helpful assistant named Alice.\\n## Conversation History\\nAlice: Hi!\\nBob: Nice to meet you!" ] } ]
All other LLMs (e.g. DashScope, ZhipuAI …)		[ { "role": "system", "content": "You're a helpful assistant named Alice." }, { "role": "user", "content": "## Conversation History\\nAlice: Hi!\\nBob: Nice to meet you!" } ]

Tip

Considering some API libraries can support different LLMs (such as OpenAI Python library), AgentScope uses the model_name field to distinguish different models and decides the used strategy.

Vision Models

For vision models, AgentScope currently supports OpenAI vision models and Dashscope mutli modal API. The more supported APIs will be added in the future.