Knowledge Base

Configure Models

To use the knowledge base, you need to configure embedding models (required) and rerank models (optional). For related configuration, please refer to Model Configuration.

After configuration, you can enable the knowledge base feature in the chat box on the record page:

During the enabling process, the model availability status will be detected.

Knowledge Base Vector Calculation

The knowledge base feature is implemented based on RAG principles, storing and searching by converting Markdown files to vectors. The entire process is completed through embedding models.

When using for the first time, if Markdown files already exist in the writing page, you can first perform a full calculation (optional):

Subsequent writing does not require full calculation each time. It will automatically calculate when auto-saving during writing.

This process consumes embedding model resources; use judiciously or use free models.

Parameter Settings

If you are unfamiliar with knowledge base-related knowledge, you can ignore this and use default values.

When using the knowledge base, you may not always find the information you need completely. By adjusting parameters, you can more precisely control the retrieval effect of the knowledge base:

• Chunk Size: Maximum character count for text chunks. Larger chunks may contain more context but increase vector calculation complexity.
• Overlap Size: Overlap character count between text chunks. Larger overlap maintains context coherence.
• Retrieval Count: Number of relevant documents returned during retrieval. More documents may provide richer information but may introduce noise.
• Similarity Threshold: Minimum similarity threshold between document and query. Only documents exceeding this threshold will be returned. Value range 0.0-1.0, higher means stricter requirements.