I'm uploading version 1.0 beta 7 to AI Party for the last time..
Now it's carrying the tag of 3.0 beta 2.
In 3.0, the functionality has improved significantly with the introduction of RAG (FTS5 + Vector) context.
It's one of the apps I've created and frequently use together, so I keep improving it!
Here are the highlights from recent changelogs.
🔍 Introduction of FTS5 + Vector DB
Beyond the limitations of basic history or context window, it continuously manages context and background information very efficiently.
You need to install an embedding model in the server management panel and enable this feature in the chat window settings to use it.
🧠 Introduction of Forgetting Function
Old fragmentary memories are automatically forgotten based on scores.
The purpose is to remove unnecessary conversation records from the database, and you can set the duration until forgetting for each user account in the server management panel.
1. User Experience and Feature Improvements
Introduction of User Profile Facts (MCP Tool): Systematically manages user-related memories to provide more personalized responses.
Enhanced LM Studio Mode: Supports display of LLM model load status and unload button functionality.
TTS Playback Optimization: Text-to-speech playback performance has been improved.
iOS PWA Screen Lock Prevention: Changed to the latest code to prevent screen lock. However.. it seems the screen doesn't turn off at all...
iOS PWA Microphone Indicator Display Bug: After using the microphone, the microphone in-use indicator doesn't turn off. This appears to be an iOS PWA app bug.
2. Performance and Resource Optimization (Web Client Diet)
Real-time Synchronization Method Change: Removed full chat window real-time delta synchronization that had significant performance overhead.
Server Signal System: Improved so the server centrally emits signals for tasks being processed.
Code and DB Cleanup: Removed dead code, performed SQLite optimization, and dropped unnecessary tables (automatically cleaned up when server runs).
3. LLM Settings and Parameter Changes
Max Tokens Policy Change: Users cannot specify this directly; it follows the default values of the LLM runtime being used.
Temperature Setting Flexibility: Configured not to send values so the model's default values are used as-is.
RAG (FTS5 + Vector) Context Improvement: Even in mode, detailed content from 1 turn and abbreviated history from 4 turns are sent together to strengthen answer context.
Tool Use Control: If OpenAI-compatible mode or MCP usage is disabled, the model is prevented from attempting tool use.
4. System and Storage
Structure Change: The settings and file storage structure have been newly changed.
Data Migration: When running the updated version for the first time, if the existing structure is incompatible with the new system, a migration guidance window is automatically displayed.
https://github.com/DINKIssTyle/DINKIssTyle-Local-LLM-Gateway-with-TTS/releases
Though it's incomplete..
I hope it will be helpful to those researching AI and those who want to see the structure.
