I use LibreOffice, and I wanted AI help inside it without sending my documents anywhere. So I built it as an extension, with no fork of LibreOffice itself.
How it works
Three parts work together:
- An extension inside LibreOffice, written in Python-UNO, which reads the selection and applies changes.
- A local service built with FastAPI, which talks to the model through Ollama.
- A side panel with chat, rewrite, structure, formula and generate tools. You see a word-by-word comparison before you apply a change, and you can undo it.
Narrow actions for a small model
The model is small because it has to fit in 4 GB of VRAM. So each action is narrow, and its output is checked before it reaches the document. I benchmarked two models on grammar, formulas and document questions. llama3.2 changed the meaning of a sentence while fixing it, so I kept Qwen2.5-3B, which fixed only the grammar.
What was hard
LibreOffice draws its whole interface into one window, so its panels cannot host a normal web view. I made the panel a child window placed exactly over the sidebar, with a secured bridge so the page can read the selection and apply changes.
