LLM FlatChat

LLM FlatChat is a small chat client built with Flet. It connects to a KoboldCPP server and stores your settings locally in storage/settings.json. The app lets you swap between a chat view and a settings view via the navigation drawer.

Requirements

• Python 3.11+
• Flet (installed via pip install flet[all])

Running the App

After installing the requirements, you can run the app by executing the following command in the project directory:

flet run

Start your KoboldCPP server first. Open the settings view to enter your name above the KoboldCPP URL and adjust the API endpoint if needed. The URL is persisted in storage/settings.json and diary entries are stored in storage/diary.db.

Building Distributables

Flet can package the application for multiple platforms:

flet build apk -v       # Android
flet build ipa -v       # iOS
flet build macos -v     # macOS
flet build linux -v     # Linux
flet build windows -v   # Windows

See the Flet publish guide for signing and distribution details.

Features

• Chat with KoboldCPP via a clean Flet UI
• Chat history context awareness
• Persistent settings (KoboldCPP URL and LLM parameters)
• Simple diary stored in an SQLite database
• Diary view bottom bar with command and save buttons
• Self-reflection question command powered by KoboldCPP
• Previous entry summary command powered by KoboldCPP
• Diary questions stream live as they are generated
• Configurable self-reflection prompts and parameters
• View saved diary entries sorted by date
• Tap a saved entry to reopen and edit it
• Delete saved entries with a long press and confirmation popup
• User data is stored under the storage/ folder

Code Structure

The application code lives under the src folder and is split into two main packages:

• backend/ – Contains logic unrelated to the UI. backend.py stores chat history and talks to the KoboldCPP API, models.py defines the Message dataclass, while settings_manager.py loads and saves the persistent AppSettings.
• frontend/ – All Flet UI components. app.py wires everything together, chat_view.py displays the message list, chat_message.py renders individual messages, and settings_view.py hosts the settings form.

Execution begins in main.py which creates the backend and loads settings before handing control to Flet.

Development

Development guidelines and testing instructions are described in AGENTS.md. To run the unit tests yourself install the project with the optional development dependencies and execute:

pip install -e .[dev]
python -m pytest

Documentation

Offline documentation for Flet lives under docs/flet-docs and can be consulted without internet access. Langchain documentation resides under docs/langchain-docs and currently includes the LangMem module.

TODO

• Settings menu

  • Settings menu to set LLM URL (KoboldCPP URL)
  • Settings for LLM system prompt
  • Settings for LLM temperature
  • Settings for LLM max tokens
  • Diary question prompt and parameters

• Chat

  • Bring in last messages as context
  • Make past messages of LLM or user editable
  • Let LLM regenerate last message
  • Save chat
  • Long term memory via Langchain (LangMem)
  • fix streaming messages, despite "stream": True in backend.py it does not work
  • Read and print which model is available (curl http://localhost:5001/v1/models gives a list of available models)
  • Markdown rendering

• Diary functionality, to let users write and save diary entries (Can be used later on as context for the LLM through Langchain)

  • Create diary button in navigation drawer which leads into a diary view
  • Give diary view a text editor style to write diary entries
  • Command popup to trigger LLM self-reflection questions

• Frontend general

  • Improve styling

• User functionality

  • User name stored in settings

Bugs users found

• Chat
  • Chat messages are not correctly displayed if being long. We need to wrap them at the end of the display
• Diary
  • Inputed text is not temporarily saved when switching to chat view (annoying when you forget to save and switch to chat view)
  • Self-reflection question popup does show if the text is very long
• Settings menu
  • Temperature slider are not labelled
  • Not automatically saving settings when changing and not pressing "Save" (annoying when you forget to save and switch to chat view)