The first step in creating a Model type plugin is to initialize the plugin project and create the model provider file, followed by integrating specific predefined/custom models. If you only want to add a new model to an existing model provider, please refer to Quick Integration of a New Model.

​
Prerequisites

  • Dify plugin scaffolding tool
  • Python environment, version ≥ 3.12
For detailed instructions on preparing the plugin development scaffolding tool, please refer to Initializing Development Tools. Before you begin, it’s recommended that you understand the basic concepts and structure of Model Plugins.

​
Create New Project

In the scaffolding command-line tool path, create a new dify plugin project. 
./dify-plugin-darwin-arm64 plugin init
If you have renamed the binary file to dify and copied it to the /usr/local/bin path, you can run the following command to create a new plugin project: 
dify plugin init

​
Choose Model Plugin Template

All templates in the scaffolding tool provide complete code projects. Choose the LLM type plugin template. Plugin type: llm

​
Configure Plugin Permissions

Configure the following permissions for this LLM plugin:
  • Models
  • LLM
  • Storage
Model Plugin Permission

​
Model Configuration Methods Explanation

Model providers support the following two model configuration methods:
  • predefined-model Predefined Models Common large model types that only require unified provider credentials to use the predefined models under the provider. For example, the OpenAI model provider offers a series of predefined models such as gpt-3.5-turbo-0125 and gpt-4o-2024-05-13. For detailed development instructions, please refer to Integrating Predefined Models.
  • customizable-model Custom Models Requires manually adding credential configurations for each model. For example, Xinference supports both LLM and Text Embedding, but each model has a unique model_uid. If you want to integrate both, you need to configure a model_uid for each model. For detailed development instructions, please refer to Integrating Custom Models.
These two configuration methods can coexist, meaning a provider can support combinations like predefined-model + customizable-model or predefined-model. This means that with configured unified provider credentials, you can use predefined models and models fetched from remote sources, and if you add new models, you can additionally use custom models on top of this foundation.

​
Adding a New Model Provider

Adding a new model provider mainly includes the following steps:
  1. Create Model Provider Configuration YAML File Add a YAML file in the provider directory to describe the provider’s basic information and parameter configuration. Write content according to ProviderSchema requirements to ensure consistency with system specifications.
  2. Write Model Provider Code Create provider class code, implementing a Python class that meets system interface requirements for connecting with the provider’s API and implementing core functionality.
Here are the complete operation details for each step.

​
1. Create Model Provider Configuration File

Manifest is a YAML format file that declares the model provider’s basic information, supported model types, configuration methods, and credential rules. The plugin project template will automatically generate configuration files under the /providers path. Here’s an example of the anthropic.yaml configuration file for Anthropic: 
provider: anthropic
label:
  en_US: Anthropic
description:
  en_US: Anthropic's powerful models, such as Claude 3.
  zh_Hans: Anthropic 的强大模型，例如 Claude 3。
icon_small:
  en_US: icon_s_en.svg
icon_large:
  en_US: icon_l_en.svg
background: "#F0F0EB"
help:
  title:
    en_US: Get your API Key from Anthropic
    zh_Hans: 从 Anthropic 获取 API Key
  url:
    en_US: https://console.anthropic.com/account/keys
supported_model_types:
  - llm
configurate_methods:
  - predefined-model
provider_credential_schema:
  credential_form_schemas:
    - variable: anthropic_api_key
      label:
        en_US: API Key
      type: secret-input
      required: true
      placeholder:
        zh_Hans: 在此输入您的 API Key
        en_US: Enter your API Key
    - variable: anthropic_api_url
      label:
        en_US: API URL
      type: text-input
      required: false
      placeholder:
        zh_Hans: 在此输入您的 API URL
        en_US: Enter your API URL
models:
  llm:
    predefined:
      - "models/llm/*.yaml"
    position: "models/llm/_position.yaml"
extra:
  python:
    provider_source: provider/anthropic.py
    model_sources:
      - "models/llm/llm.py"
If the provider you’re integrating offers custom models, such as OpenAI providing fine-tuned models, you need to add the model_credential_schema field. Here’s a sample code for the OpenAI family of models: 
model_credential_schema:
  model: # Fine-tuned model name
    label:
      en_US: Model Name
      zh_Hans: 模型名称
    placeholder:
      en_US: Enter your model name
      zh_Hans: 输入模型名称
  credential_form_schemas:
  - variable: openai_api_key
    label:
      en_US: API Key
    type: secret-input
    required: true
    placeholder:
      zh_Hans: 在此输入您的 API Key
      en_US: Enter your API Key
  - variable: openai_organization
    label:
        zh_Hans: 组织 ID
        en_US: Organization
    type: text-input
    required: false
    placeholder:
      zh_Hans: 在此输入您的组织 ID
      en_US: Enter your Organization ID
  - variable: openai_api_base
    label:
      zh_Hans: API Base
      en_US: API Base
    type: text-input
    required: false
    placeholder:
      zh_Hans: 在此输入您的 API Base
      en_US: Enter your API Base
For more complete model provider YAML specifications, please refer to the Model Schema documentation.

​
2. Write Model Provider Code

Create a python file with the same name, e.g., anthropic.py, in the /providers folder and implement a class that inherits from the __base.provider.Provider base class, e.g., AnthropicProvider. Here’s an example code for Anthropic: 
import logging
from dify_plugin.entities.model import ModelType
from dify_plugin.errors.model import CredentialsValidateFailedError
from dify_plugin import ModelProvider

logger = logging.getLogger(__name__)


class AnthropicProvider(ModelProvider):
    def validate_provider_credentials(self, credentials: dict) -> None:
        """
        Validate provider credentials

        if validate failed, raise exception

        :param credentials: provider credentials, credentials form defined in `provider_credential_schema`.
        """
        try:
            model_instance = self.get_model_instance(ModelType.LLM)
            model_instance.validate_credentials(model="claude-3-opus-20240229", credentials=credentials)
        except CredentialsValidateFailedError as ex:
            raise ex
        except Exception as ex:
            logger.exception(f"{self.get_provider_schema().provider} credentials validate failed")
            raise ex
Providers need to inherit the __base.model_provider.ModelProvider base class and implement the validate_provider_credentials method for validating unified provider credentials. 
def validate_provider_credentials(self, credentials: dict) -> None:
    """
    Validate provider credentials
    You can choose any validate_credentials method of model type or implement validate method by yourself,
    such as: get model list api

    if validate failed, raise exception

    :param credentials: provider credentials, credentials form defined in `provider_credential_schema`.
    """
Of course, you can also initially reserve the validate_provider_credentials implementation, and reuse it directly after the model credential verification method is implemented.

​
Custom Model Providers

For other types of model providers, please refer to the following configuration methods. For custom model providers like Xinference, you can skip the full implementation step. Simply create an empty class called XinferenceProvider and implement an empty validate_provider_credentials method in it. Detailed Explanation: • XinferenceProvider is a placeholder class used to identify custom model providers. • While the validate_provider_credentials method won’t be actually called, it must exist because its parent class is abstract and requires all child classes to implement this method. By providing an empty implementation, we can avoid instantiation errors that would occur from not implementing the abstract method. 
class XinferenceProvider(Provider):
    def validate_provider_credentials(self, credentials: dict) -> None:
        pass
After initializing the model provider, the next step is to integrate specific llm models provided by the provider. For detailed instructions, please refer to:

​
Reference Resources

Edit this page | Report an issue