Integrate the Predefined Model

Before accessing a predefined model, make sure you have created a model provider. Accessing predefined models is roughly divided into the following steps:

1. Create Module Structures by Model Type

   Create corresponding sub-modules under the provider module based on model types (such as llm or text_embedding). Ensure each model type has its own logical layer for easy maintenance and extension.

2. Write Model Request Code

   Create a Python file with the same name as the model type (e.g., llm.py) under the corresponding model type module. Define a class that implements specific model logic and complies with the system's model interface specifications.

3. Add Predefined Model Configuration

   If the provider offers predefined models, create YAML files named after each model (e.g., claude-3.5.yaml). Write file content according to AIModelEntity specifications, describing model parameters and functionality.

4. Test Plugin

   Write unit tests and integration tests for newly added provider functionality to ensure all function modules meet expectations and operate normally.

---

Below are the access details:

1. Creation of different module structures by model type

A model provider may offer different model types, for example, OpenAI provides types such as llm or text_embedding. You need to create corresponding sub-modules under the provider module, ensuring each model type has its own logical layer for easy maintenance and extension.

Currently supported model types:

• llm: Text generation models

• text_embedding: Text Embedding models

• rerank: Rerank models

• speech2text: Speech to text

• tts: Text to speech

• moderation: Content moderation

Taking Anthropic as an example, since its model series only contains LLM type models, you only need to create an /llm folder under the /models path and add yaml files for different model versions. For detailed code structure, please refer to the Github repository.

├── models
│   └── llm
│       ├── _position.yaml
│       ├── claude-2.1.yaml
│       ├── claude-2.yaml
│       ├── claude-3-5-sonnet-20240620.yaml
│       ├── claude-3-haiku-20240307.yaml
│       ├── claude-3-opus-20240229.yaml
│       ├── claude-3-sonnet-20240229.yaml
│       ├── claude-instant-1.2.yaml
│       ├── claude-instant-1.yaml
│       └── llm.py

If the model provider contains multiple types of large models, e.g., the OpenAI family of models contains llm and text_embedding, moderation, speech2text, and tts types of models, you need to create a folder for each type under the /models path. The structure is as follows:

├── models
│   ├── common_openai.py
│   ├── llm
│   │   ├── _position.yaml
│   │   ├── chatgpt-4o-latest.yaml
│   │   ├── gpt-3.5-turbo.yaml
│   │   ├── gpt-4-0125-preview.yaml
│   │   ├── gpt-4-turbo.yaml
│   │   ├── gpt-4o.yaml
│   │   ├── llm.py
│   │   ├── o1-preview.yaml
│   │   └── text-davinci-003.yaml
│   ├── moderation
│   │   ├── moderation.py
│   │   └── text-moderation-stable.yaml
│   ├── speech2text
│   │   ├── speech2text.py
│   │   └── whisper-1.yaml
│   ├── text_embedding
│   │   ├── text-embedding-3-large.yaml
│   │   └── text_embedding.py
│   └── tts
│       ├── tts-1-hd.yaml
│       ├── tts-1.yaml
│       └── tts.py

It is recommended to prepare all model configurations before starting the model code implementation. For complete YAML rules, please refer to the Model Design Rules. For more code details, please refer to the example Github repository.

2. Writing Model Requesting Code

Next, you need to create an llm.py code file under the /models path. Taking Anthropic as an example, create an Anthropic LLM class in llm.py named AnthropicLargeLanguageModel, inheriting from the __base.large_language_model.LargeLanguageModel base class.

Here's example code for some functionality:

• LLM Request

  The core method for requesting LLM, supporting both streaming and synchronous returns.

def _invoke(self, model: str, credentials: dict,
            prompt_messages: list[PromptMessage], model_parameters: dict,
            tools: Optional[list[PromptMessageTool]] = None, stop: Optional[list[str]] = None,
            stream: bool = True, user: Optional[str] = None) \
        -> Union[LLMResult, Generator]:
    """
    Invoke large language model

    :param model: model name
    :param credentials: model credentials
    :param prompt_messages: prompt messages
    :param model_parameters: model parameters
    :param tools: tools for tool calling
    :param stop: stop words
    :param stream: is stream response
    :param user: unique user id
    :return: full response or stream response chunk generator result
    """

In the implementation, you need to be careful to use two functions to handle synchronized returns and streaming returns separately. This is because functions in Python that contain the yield keyword are recognized as generator functions, and their return type is fixed to Generator. synchronized return and streaming return need to be implemented independently in order to ensure that the logic is clear and to accommodate different return requirements.

Here's the sample code (the parameters are simplified in the example, so please follow the full parameter list in the actual implementation):

def _invoke(self, stream: bool, **kwargs) -> Union[LLMResult, Generator]:
   """Call the corresponding processing function based on return type."""
   if stream:
       return self._handle_stream_response(**kwargs)
   return self._handle_sync_response(**kwargs) 

def _handle_stream_response(self, **kwargs) -> Generator:
   """Handle streaming response logic."""
   for chunk in response: # Assume response is a streaming data iterator
       yield chunk

def _handle_sync_response(self, **kwargs) -> LLMResult:
   """Handle synchronous response logic.""" 
   return LLMResult(**response) # Assume response is a complete response dictionary

• Pre-calculated number of input tokens

If the model does not provide an interface to pre-calculate tokens, it can simply return 0, which is used to indicate that the feature is not applicable or not implemented. Example:

def get_num_tokens(self, model: str, credentials: dict, prompt_messages: list[PromptMessage],
                   tools: Optional[list[PromptMessageTool]] = None) -> int:
    """
    Get number of tokens for given prompt messages

    :param model: model name
    :param credentials: model credentials
    :param prompt_messages: prompt messages
    :param tools: tools for tool calling
    :return:
    """

• Request Exception Error Mapping Table

When a model call encounters an exception, it needs to be mapped to the InvokeError type specified by Runtime, allowing Dify to handle different errors differently.

Runtime Errors:

• InvokeConnectionError: Connection error during invocation

• InvokeServerUnavailableError: Service provider unavailable

• InvokeRateLimitError: Rate limit reached

• InvokeAuthorizationError: Authorization failure during invocation

• InvokeBadRequestError: Invalid parameters in the invocation request

@property
def _invoke_error_mapping(self) -> dict[type[InvokeError], list[type[Exception]]]:
    """
    Map model invoke error to unified error
    The key is the error type thrown to the caller
    The value is the error type thrown by the model,
    which needs to be converted into a unified error type for the caller.

    :return: Invoke error mapping
    """

See the Github code repository for full code details.

3. Add Predefined Model Configurations

If the provides predefined models, create YAML files for each model with the same name as the model name (e.g. claude-3.5.yaml). Write the contents of the file according to the AIModelEntity specification, describing the parameters and functionality of the model.

claude-3-5-sonnet-20240620 Model example code:

model: claude-3-5-sonnet-20240620
label:
  en_US: claude-3-5-sonnet-20240620
model_type: llm
features:
  - agent-thought
  - vision
  - tool-call
  - stream-tool-call
  - document
model_properties:
  mode: chat
  context_size: 200000
parameter_rules:
  - name: temperature
    use_template: temperature
  - name: top_p
    use_template: top_p
  - name: top_k
    label:
      zh_Hans: 
      en_US: Top k
    type: int
    help:
      zh_Hans: 
      en_US: Only sample from the top K options for each subsequent token.
    required: false
  - name: max_tokens
    use_template: max_tokens
    required: true
    default: 8192
    min: 1
    max: 8192
  - name: response_format
    use_template: response_format
pricing:
  input: '3.00'
  output: '15.00'
  unit: '0.000001'
  currency: USD

4. Debugging Plugins

Dify provides remote debugging method, go to "Plugin Management" page to get the debugging key and remote server address. Check here for more details:

Debug Plugin

Publishing Plugins

You can now publish your plugin by uploading it to the Dify Plugins code repository! Before uploading, make sure your plugin follows the plugin release guide. Once approved, the code will be merged into the master branch and automatically live in the Dify Marketplace.

Exploring More

Quick Start:

• Develop Extension Type Plugin
• Develop Model Type Plugin
• Bundle Type Plugin: Package Multiple Plugins

Plugins Specification Definition Documentaiton:

• Minifest
• Endpoint
• Reverse Invocation of the Dify Service
• Tools
• Models
• Extend Agent Strategy