Readable

Turn PDFs, scans, recordings and Office files into notes. Everything runs on your computer.

Install · Quick start · Privacy · Changelog · LyraGen

Drop a file into your vault and Readable writes a note for it: title, tags and summary from a local model, the file embedded, and its full text with tables and figures. Plain Markdown, so Obsidian search, Omnisearch and your AI tools can find what was in the file.

Install

Desktop only. In Obsidian, open Settings → Community plugins → Browse, search for Readable and install it, or open its directory page and choose Add to Obsidian. You can also copy main.js, manifest.json and styles.css from the latest release into .obsidian/plugins/readable/.

Quick start

1. Install the engines. Enable Readable and press Install when it asks for Docling, which reads documents (about 2 GB). Then open Settings → Readable → LLM, keep Readable's own as the provider, install it and download the recommended model under Models (Gemma 4 E4B, about 5 GB, best with 16 GB of memory). Already run Ollama or LM Studio? Pick it as the provider instead.

2. Ingest a file. Right-click a file or folder → Ingest with Readable. To skip the right-click, turn on Auto-ingest new files under Automation and pick a watched folder.

3. Watch the queue. The queue on the right shows each file as it is read and summarised. Files that are done get a tick in the file list.

4. Read the note. Each file gets a note in the Readable folder: a summary, then the full text with its tables.

The note's title and tags come from what the file says, so search finds Q3 board review.pdf by "supply chain risk" as well as by its name.

Your files stay on your computer

Docling and the model server run on this machine, and text goes only to the model server address you set. Readable has no telemetry, no account and no paid tier. Links and HTML inside documents are written as text, never embedded.

Network use

Readable downloads its engines and models only when you press Install or Download in its settings, at pinned versions:

With Readable's own AI that last address is a server on this computer that only this computer can reach. If you point it at Ollama, LM Studio or another server, text goes wherever that server runs. Nothing else is contacted.

Files outside the vault

Readable keeps its engines, models and each vault's queue in its own app-data folder, never in the vault: ~/Library/Application Support/Readable on macOS, %APPDATA%\Readable on Windows, ~/.local/share/readable on Linux. Docling with its models takes about 2 GB, the speech add-on about 2 GB more, and each chat model 2 to 6 GB. Remove and Delete in settings delete them. If you set an extra models folder, Readable reads .gguf files from it and never writes there.

Programs it runs

Readable starts programs from that folder, with fixed arguments and never through a shell: a Python process that runs Docling and Whisper, ffmpeg for recordings, and the llama.cpp server for Readable's own AI. Each stops when it is not needed or when Obsidian closes.

Development

npm install
npm run dev     # watch build
npm test

Architecture and conventions: ARCHITECTURE.md.

Readable installs Docling, uv, Python, llama.cpp, Whisper and ffmpeg on request, each under its own licence; it does not bundle them. Each model's licence is on its Hugging Face page. Bundled npm packages: THIRD_PARTY_NOTICES.md.

MIT · Made by LyraGen