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 an 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

Native 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>.
Image Vision AIs: send images to vision-supporting LLMs using local WikiLinks or Markdown external web links.
System Prompts: name your system prompts, create personalities and use FrontMatter to choose whom to talk to.

Plugin Demo Vault

Download the Plugin Demo Vault in ZIP or TGZ to try it out !

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 lets 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:
    config: Gladdis   # select which AI you talk to
    model: gpt-4o     # choose the current AI model
    server: http://   # use local or custom servers
    temperature: 42   # test different temperatures
    top_p_param: 88   # or various probability mass
whisper:
    config: Whisper   # switch transcription config
    model: whisper-1  # change the transcription AI
    temperature: 24   # test different temperatures
    echoOutput: true  # write 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 underscores to separate the 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

The following OpenAI models are supported via the paying API:

Label	Tokens	Vision	Notes
`o3-mini`	200 k		OpenAI's latest reasoning model designed to excel at science, math, and coding tasks.
`o1-mini`	128 k		OpenAI's first reasoning model designed to solve hard problems across domains.
`gpt-4o`	128 k	✅	OpenAI's high-intelligence flagship model for complex, multi-step tasks.
`gpt-4o-mini`	128 k	✅	OpenAI's affordable and intelligent small model for fast, lightweight tasks.
`chatgpt-4o-latest`	128 k	✅	Dynamic model continuously updated to the current version of GPT-4o in ChatGPT.
`gpt-4-turbo`	128 k	✅	The latest GPT-4 Turbo model with vision capabilities (GPT-4o is cheaper and faster).
`gpt-3.5-turbo`	16 k		The latest GPT-3.5 Turbo model with higher accuracy (GPT-4o mini is cheaper and more capable).

Open-Source models can be used locally via Ollama or LocalAI:

Label	Tokens	Vision	Notes
`llama3.3`	128 k		Meta's new 70B model with similar performance to the Llama 3.1 405B model.
`llama3.2-vision`	128 k	✅	Meta's image reasoning generative models in both 11B and 90B model sizes.
`llama3.2`	128 k		Meta's new 1B and 3B models optimized for multilingual dialogue use cases.
`llama3.1`	128 k		Meta's 8B, 70B and 405B models with overall stronger reasoning capabilities.
`phi4`	16 k		Microsoft's new Phi 4 "Medium" (14B) open model with a 16k context window.
`phi3.5-vision`	128 k	✅	Microsoft's Phi 3 and 3.5 "Mini" (3.8B) models with vision capabilities.
`phi3.5` / `phi3`	128 k		Microsoft's new Phi 3.5 "Mini" (3.8B) and Phi 3 "Medium" (14B) open models.
`qwen2.5`	32 k		Alibaba Cloud's Qwen multilingual Chinese models ranging from 0.5B to 72B.
`gemma2`	8 k		Google's Gemma 2B, 9B and 27B models featuring a brand new architecture.
`llava`	4 k	✅	Family of Large Language-and-Vision Assistant models between 7B and 34B.

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".