Introduction

On the [Model Configuration] -> [Model Services] page, click to create a model service. It supports local deployment engines such as ollama, vllm, and lm studio, as well as numerous cloud service providers that you can choose from.

Main Model

  1. Almost all areas without explicit specifications are driven by the main model, including but not limited to: memory extraction in the memory module, planning model and progress assessment model in deep research. Therefore, it's necessary to select [a model with tool capabilities and stable JSON format output].
  2. Temperature: Between 0 and 1, higher values make the model more creative, lower values make it more precise
  3. Top p: Between 0 and 1, higher values result in more diverse model outputs, lower values lead to more stable outputs
  4. Output length: The maximum sum of input and output tokens; the model won't run if input/output is too long. Increasing this parameter allows the model to output more content but requires more computational power for each response.
  5. Conversation rounds: Historical records exceeding the conversation round limit will be ignored and not input to the model. When set to 0, it means the conversation compression feature is disabled
  6. Additional parameters: You can add your desired model parameters here, such as top k, enable_think, and other parameters supported by certain models

Reasoning Model

  1. For first-time use, you can add new providers in the model service interface. After adding a provider, please return to this page and select [a model with tool capabilities and stable JSON format output] to continue.
  2. When the reasoning model is enabled, it will perform reasoning before the main model responds. The reasoning model's response will be truncated, and the reasoning part will be concatenated to the main model's context, enabling the main model to acquire reasoning capabilities. This allows [models with tool capabilities and stable JSON format output] or [models with special functions after fine-tuning] to gain reasoning abilities, achieving combined model output

Vision Model

  1. For first-time use, you can add new providers in the model service interface. After adding a provider, please return to this page and select [Vision Model] to continue. Vision model names typically include keywords like "vision", "v", "o".
  2. When the vision model is enabled, sent images will be separated for vision model analysis before being passed to the main model, allowing the main model to gain visual capabilities. For repeatedly sent images, caching based on image hash values prevents unnecessary computational power consumption when sending conversation history each time.

Text-to-Image Model

  1. By default, it uses free pollinations to let the large model generate images.
  2. If you need to use OpenAI format text-to-image API, for first-time use, you can add new providers in the model service interface. After adding a provider, please return to this page and select [Text-to-Image Model] to continue. Text-to-image model names typically include keywords like "img", "image". Currently supported providers: OpenAI, Silicon Flow. Other OpenAI format text-to-image APIs should theoretically be compatible
  3. You can provide more detailed image generation prompts in the system prompt to achieve better text-to-image results.