Gladdis NPM | npm.io

Description

Obsidian exists to help you build a second brain. Gladdis exists to give that brain a pen (let it do the writing for once), ears (talking is easier than typing) and multiple split personalities (give your system prompts names). Further plans for v1.0 include: a brush (let it paint like a true artist), a voice (listening is easier than reading), an internal monologue (together we are more than our sum) and the ability to execute code (what could possibly go wrong ?).

Current Features

Simple Markdown Syntax: the conversation window is any open Markdown file, the streaming LLM response is appended.
Whisper Transcription: audio files WikiLinks are auto-transcribed, use the Obsidian "audio recorder" core plugin.
Document Loading: text files WikiLinks are auto-loaded, just drag and drop files into the conversation window.
PDF & Web Support: load PDF WikiLinks just like text files, load web pages external links as well using <URL>.
System Prompts: name your system prompts, create personalities and use FrontMatter to choose whom to talk to.

Documentation

Obsidian Commands

Gladdis provides 3 commands, map them to hotkeys in the settings or create buttons with the Commander plugin.

Chat with Gladdis: the main command to perform the full processing, including the call to the LLM model.
Process the Content: to check that all links and transcriptions are correct, and get a full token count.
Process the Prompt (or Selection): to check just the prompt or selection and get a specific token count.

Settings / Options

The settings panel let you specify the plugin data folder, your OpenAI API key and the FrontMatter defaults.

Data Root Path: Gladdis will search for config files in the configs subfolder, and will write the chat and call logs in the history subfolder.
OpenAI Secret Key: create an OpenAI API account, generate an API key and paste it in the settings panel.
FrontMatter Defaults: adjust the various defaults for the Gladdis and Whisper options.

FrontMatter Config

The defaults from the settings panel can be overridden using FrontMatter in any conversation or config file.

---
gladdis:
    label: Gladdis    # name the AI in the config
    config: Gladdis   # switch the AI you talk to
    model: gpt-3.5-turbo # switch the model at any time
    temperature: 42   # test different temperatures
    top_p_param: 100  # or various probability mass
whisper:
    config: Whisper   # switch transcription language
    model: whisper-1  # only one model at the moment
    temperature: 24   # test different temperatures
    echoOutput: true  # output transcription callout
    deleteFile: false # cleanup after transcription
---

Markdown Syntax

The same familiar Obsidian syntax from your notes is used for the conversation history (it's all just text).

Just use standard Markdown text for `User` messages and prompts.
The content of the Gladdis config file will be prepended to this.
Use triple dashes to separate messages / sections of conversation.

---

__UserName:__ labels also start a new message with custom names.

__Gladdis:__ Gladdis' label and `Assistant` are for AI messages.

__System:__ is for System Prompts (the default in config files).

---

> Quoted lines are ignored, error and token counter callouts too.

[[path/file.md|WikiLinks to file]] and ![[audio_recording.mp3]] will be parsed.
Static Web pages can be loaded using angle brackets: <https://www.example.com>.

The very last message is the prompt, the LLM response streams below.

Available Models

Only the OpenAI models are supported at the moment, however more APIs and local models are coming.

Label	Token Limit	Notes
`gpt-4-turbo-preview`	128 k	The most powerful and expensive.
`gpt-4-32k`	32 k	The old GPT-4, with large context.
`gpt-4`	8 k	The original GPT-4, from June 2023.
`gpt-3.5-turbo`	16 k	The updated GPT-3.5, cheaper than 4.
`gpt-3.5-turbo-16k`	16 k	The old GPT-3.5, with large context.
`gpt-3.5-turbo-0613`	4 k	The original GPT-3.5, from June 2023.

Future Development

Image & Audio Out: create images with DALL·E, speak with ElevenLabs or other API accessible models.
Code Execution: add Templater code to your prompts, run any code the AI generates (are you sure ?!).
AI-2-AI Prompt: let your various AI configs talk to each other, create your own "Council of Gladdises".