Why it exists

"Chat with your PDFs" products all share a catch: your documents get uploaded to someone's server, where they become training-adjacent data or breach material. Dot's answer is blunt: run the whole loop — embedding, retrieval, generation, even speech — on your own machine, with a small local LLM and a local vector store. The result works offline, costs nothing, and the only place your spreadsheet has ever been is your computer.

Main features

• Fully local RAG — point Dot at folders (PDFs, spreadsheets, text), ask questions, get answers with the relevant chunks retrieved from your files; everything runs on-device, offline.
• Small local models by default — Phi-3.5 ships as the default (the choice of LLM is yours: Mistral and friends work too via llama.cpp); no API keys, no accounts, no egress.
• "Big Dot" mode — flip a switch to use the latest cloud Mistral models via API when the question is harmless and you want more horsepower.
• Voice in and out — speech-to-text and text-to-speech built in, so Dot works as a local, talking assistant.
• Broad document support — PDF, DOCX, PPTX, XLSX and more; load folders, ask questions, switch contexts — the demo on the site covers component spreadsheets (prices, BOMs), drafting emails and follow-up questions across files.
• Compact and hackable — an Electron app small enough to understand (FAISS + LangChain + llama.cpp under the hood) — a great codebase to learn local RAG from.

Who it's for

• Anyone with documents they legally or morally can't upload: client NDAs, medical records, unreleased research.
• Mac users on Apple Silicon (or Windows) with 8–16 GB RAM who want to try local AI without Docker or terminal fu.
• Developers wanting a readable, GPL reference implementation of local RAG + TTS + STT.
• Journalists, lawyers and doctors in demo mode: the two-sentence pitch writes itself.

How it compares

• Jan — the bigger, actively-developed cross-platform local assistant with a model hub; Dot is the focused "chat with my files" tool and a simpler install.
• LM Studio + documents — LM Studio is a polished closed-source local model runner; Dot bakes the RAG layer in for you.
• ChatGPT file uploads — vastly stronger model, zero privacy; Dot's entire reason to exist is that trade.
• PrivateGPT / localGPT — similar open-source RAG projects; Dot ships as a real macOS app with voice, not a Python script.

Notes

• Development has been quiet since late 2024 (the repo isn't archived, but don't expect fast fixes — the maintainer is a student) — the trade-off for a small, finished codebase.
• Phi-3.5-class models are 2024-class: excellent for summarizing and locating facts in your files, not a GPT-5-class reasoner; Big Dot mode is the escape hatch.
• Needs a reasonably modern machine (Apple Silicon or Windows); indexing speed and answer quality scale with available memory.
• GPL-3.0, no telemetry, no account — the privacy story is the architecture, not a policy.

Useful links

• Official site: https://dotapp.uk/
• Source code: https://github.com/alexpinel/Dot
• How to use: https://dotapp.uk/how-to-use
• Releases (beta downloads): https://github.com/alexpinel/Dot/releases