v2.4.0-beta1

🚀 BISHENG 2.4 beta1

🧩 1. Enhanced Workflow Core Capabilities

🔀 1.1 Input Node Logic Refactor & Enhancements

1. Strategy decoupling: The file-handling strategy for input nodes has been split out, allowing you to choose among:

• Parse content (extract text)
• Do not parse (keep the original file for multimodal/code nodes)
• Save to a temporary knowledge base
  This better supports a wide range of business scenarios.

2. Smart recommended questions: Supports configuration to automatically generate and display three recommended questions based on historical conversation context.

🖼️ 1.2 Report Node Rendering Optimization

Now supports recognizing Markdown headings and bold styling, and automatically converting them into corresponding Word document styles to improve the visual quality of exported documents.

🖥️ 2. Multimodal Support in Workbench Daily Mode

A new backend vision capability toggle has been added. When enabled, uploaded images can be semantically understood via a multimodal model, instead of relying on OCR text recognition.

🧠 3. Deep Integration with DeepSeek-V3.2

DeepSeek-V3.2 is the first model to incorporate “thinking” into tool usage (historically, DeepSeek’s official model APIs—including R1-0528—could not call tools in thinking mode). In this iteration, we deeply adapted DeepSeek-V3.2 and resolved tool-calling errors encountered when assistants use DeepSeek-V3.2.

🔧 4. Other Improvements & Bug Fixes

Includes multiple optimizations such as:

• Support for high-availability deployments
• Drag-and-drop file upload in the chat input box
• Ability to configure whether “灵思模式” is displayed
• Audit logs now cover daily-mode sessions
  Also fixes an issue where switching the embedding model for the knowledge base did not take effect.

💡 5. Notes

This upgrade changes certain fields returned by the workflow API. Specifically: in the Input Node → Form Input Mode → File-type form, the form key is now automatically generated by the backend.

📢 6. Skill Module Deprecation Notice

As an early application-building approach in BISHENG, the skills (langflow) module has become difficult to use due to an excessive number of components and unintuitive orchestration logic, and its capabilities have been superseded by workflows and assistants. Therefore, starting from 2.4 release, we will deprecate and remove the skills module.