AI Fundamentals

This page explains the most common terms in Cherry Studio to help you determine whether you should change the model, adjust the prompt, or add tools.

Agents organize tasks, while the model generates content; the knowledge base, skills, and tools provide information, methods, and operational capabilities respectively.

Model services, models, and API keys

• Model service: the vendor, cloud platform, or local runtime environment that provides model access;
• Model: the capability to perform specific tasks such as text, vision, image generation, embedding, or reranking;
• API address: the service endpoint where Cherry Studio sends requests;
• API Key: the credential used by the provider to identify the account and authorize requests.

The same provider can offer multiple models. An API Key can connect to the service, but that does not mean every model is enabled, nor does it mean the interface types are necessarily compatible.

Tokens and context window

A token is the basic unit used by a model when processing text, and it is not equivalent to a Chinese character or a word. Inputs, history messages, tool definitions, knowledge base content, and outputs all consume context.

The context window is the total amount that can be handled in a single request. When approaching the limit, Cherry Studio may compress earlier content; important facts should be placed in explicit prompts, files, knowledge bases, or Agent memory, rather than relying only on a sentence from a long time ago.

Sampling parameters

Parameters	Effect	Adjustment direction
Temperature	Controls output randomness	Lower for factual tasks, and can be increased moderately for creative tasks
Top P	Limits the candidate word range	Usually adjust only one of it and Temperature as the main focus
Max Tokens	Limits the length of a single output	Too low may truncate the output; too high does not mean better content

Different models may vary slightly in their support for and interpretation of parameters. If there is no clear issue, first use the product defaults instead of copying so-called "universal parameters" from the internet.

Embedding and Rerank

• Embedding model: converts text into vectors for similarity retrieval in the knowledge base;
• Reranking model: reorders the initial retrieval results to improve the ranking of the most relevant content.

They are not responsible for generating the final answer. If the knowledge base returns no results, first check document processing and embeddings; if results exist but the ranking is not ideal, then consider reranking.

Assistant and Agent

• Assistant: centered on continuous conversation, fixed roles, and output style;
• Agent: centered on goal execution, working directory, tools, and multi-step tasks.

For simple Q&A, an assistant is more direct; use an Agent when you need files, terminal, subtasks, channels, or scheduled tasks.

Skills, tools, MCP, and knowledge base

Concept	One-sentence understanding
Tools	An action the Agent can perform
skills	Tell the Agent how to work using a fixed method
MCP	Connect external tools and resources to the Agent
Knowledge base	Restrict the materials that the Agent or assistant can retrieve
Memory	Save stable facts and experience across tasks for the same Agent

Multimodal and image generation models

Multimodal models can understand different inputs such as images and text; image generation models are specialized in generating images. The Agent's main model is responsible for understanding your request, while the built-in [Generate Image] tool calls the image generation model selected under [Settings] → [Default Model].

Being able to see images, generate images, and call tools are three different capabilities. When choosing a model, first confirm what the task requires; do not look only at the model name.

Are local models always more private?

Running the model locally can reduce sending content to the model provider, but web search, MCP, channels, and other external tools may still transmit data. Privacy depends on the entire workflow, not just the model location.