LibreOffice Conference 2026: AI extension for LibreOffice Draw

Duration: 23:08 1843 transcript words EN summary

Video summary

Shatami introduces his new LibreOffice Draw AI extension, a Python-based tool designed to streamline vector graphics creation and editing through natural language instructions. As a software engineer who previously developed an AI module for Writer in 2023, Shatami aims to solve the tedious process of manually clicking through menus to change colors, sizes, or positions of shapes. By allowing users to simply type commands like "create a pelican riding a bicycle" or "change the bear's ears from circles to triangles," the extension translates these requests into precise operations within LibreOffice Draw, significantly reducing the potential for human error and saving time on repetitive tasks. The technical foundation of this tool relies heavily on SVG (Scalable Vector Graphics), an XML-based format that serves as a common language between LibreOffice and AI models. Unlike raster graphics editors such as Photoshop, vector applications deal with mathematical shapes that can be scaled arbitrarily, making SVG ideal for technical charts and scalable designs. The extension communicates with Large Language Models (LLMs) to generate or modify these SVG files, which are then inserted directly into the document. Shatami highlights that while proprietary tools often require expensive subscriptions, this open-source approach allows users to leverage powerful AI capabilities without paying per token, provided they have the necessary hardware or access to local models. A critical design priority for this project is user privacy, which is addressed by supporting both local and cloud-based AI configurations. When running locally using tools like Ollama with models such as GLM-5.3 or Qwen, the extension processes data on the user's machine, eliminating the need to send sensitive information to external servers. For users without high-end hardware, a manual mode is available where requests are copied to the clipboard and pasted into a preferred AI interface, offering flexibility while maintaining control over data handling. Although current performance can be slow with limited GPU memory, Shatami notes that future updates will focus on improving undo/redo functionality, adding sidebars, integrating with his previous Writer extension, and enhancing support for complex 3D shapes.

Read the full video transcript

In the name of God. Hello, I'm Shatami and today I will present my new extension that makes it possible to use AI tools in Libra Office. This is the list of contents that I will discuss in this presentation. I'm a software engineer and I'm using Libra Office for a few years now. I developed the first AI extension for Libra office in 2023 which was devoted to Libra office writer. That extension can create text and image with the instructions from the user. It can also do some editing and translation tasks based on the input from the selected contents in writer. that extension connected to chat GP and after that other extensions have been developed extensions used AI in Cal writer impress and draw then let's see the new extension it's called draw AI it is a lira extension written in Python that lets people to create and edit shapes in Libra office draw but with natural language. Users type their in their instructions and will see the resulting draw. This means instead of many clicks, we can type a sentence and AI send operations. A famous example to examine the SVG generation functionality of the AI tools is to ask generating some cartoons. The one that you see here is generated by asking to create a pelican that rides a bicycle. This is the output of GLM5.3 from Z AI running locally with Olama. You may see better or worse outputs depending on the AI tool and model that you use. Imagine we want to change color or size of pieces of this shape. If we want to do this operation without AI, we should select shape, open color panel, select color, we should do these steps for all pieces. And for changing size or moving shapes also we should clicks and calculation a lot. It's boring. This process has a high probability of error. Draw AI extension is my solution for creating and editing shapes. It connects Libra stro with AI. Libra fistro that many of you actually use is a vector graphics editing application. It used for many operations from quick design to technical charts and three-dimensional shapes. There are similar proprietary applications like coral drawer, vio and etc. which do vector editing and this is different from Photoshop and applications like that which are photo and raster graphics editors. Also, Libra can import and export of various different vector formats including OG. It's native format, PDF and SVG. I use SVG format extensively in my extension and also Libra store has different readymade shapes. AI is integrated into proprietary vector editing applications in various levels. Use of AI in different levels from the original idea to design and shape management is something that's doable with AI. Also, AI helps do repetitive tasks which is otherwise tedious. These tools require subscription and can be costly for the user. When you are creating or editing vector graphics, you are dealing with shapes that can be modified and displayed in arbitrary sizes. On the other hand, raster graphics are suitable for different purposes like picture and things like that. Here we focus on vector editing and specifically SVG files. SVG is an XML based format that can be a common language between Libra office and AI tools. Both Libra Office and AI tools understand it well. Libra Office can open SVG files, export files to SVG and insert SVG files as drawings. This helps the extension development a lot. Here you can see a simple SVG file displayed as text. The result is the red square that you can see on the right. You also see the nice looking B near that which is also created using AI as SVG. This is a more complex example of vector graphics that's stored as SVG. Drawi functionalities are shape creation and shape editing. With user description, Draw AI can create simple or complex shapes. And also with user request, it can change shapes, color, move or delete shapes, add new shapes, and add or modify text. As you see, I asked it create a nice threedimensional bear with pink color. And after that I asked it to change eyes color from black to green, bear's color from pink to blue, create a in the hand of the bear with red color and hello text. And also I asked to change bear's ears from circle shape to triangle. In our previous extension that was working in writer, we simply sent our message which could also invite the selected text alongside the user request in return AI tool send the result in text format and we could insert that in writer in a proper format. Even color and formatting was something that was done by the extension and not AI itself. But to edit the shapes, we need a new form of communication between our software and LLM. This diagram shows different scenarios for using draw AI extension. At first user requests through user interface of the extension in libra. It's important that the AI tool is local or not. If it's local then local process answers the query and you may see instant increase in GPU usage. If it's not local then you may use other cloud models that are available in ola.com. You have to pay for them. An alternative is to use manual model. In this way request to AI is copied to clipboard. user paste it into the AI tool user interface and copies back the result into the draw AI extension. In any of these three ways, the received operations are executed and possible received SVG files are added to Drew. In that way you will edit the the document with your request in a natural human language. User request to AI includes user instruction selected shapes and their size and positions and all data is serialized as SVG and JSON format. and response should be in JSON objects format. Some of the rules are don't send explanations, codes and markdown, answer only in JSON format. Supported operations are shape addition, shape creation, SVG operations and deletion. Shape addition includes move, set size, set text, add text, and some shapes like rectangle, ellipse, and line. Shape creation includes set fill color, set line color, set position, and set size. SVG operations are insert SVG, replace with SVG, and replace selection with SVG. Deletion includes delete shape. There is another extension write agent which can be used in different libra office applications including Libra office but it is not specific to Libra office. It uses delegation mechanism to do editing which means it asks AI to use the libra stro primitives to edit shapes. This can be good for simple shapes but when it comes to more complex shapes it doesn't necessarily work well and fast. We use SVG and higher level operations which is better in two aspects. AI tools are trained extensively for generating SVG shapes and that improves the result. We use higher level commands that works better on more complex shapes. Our test were done using a local instance which is not that fast. Therefore, the difference in performance was quite visible. Implementation and execution. The extension code is implemented in Python. The Python code has different modules. ing them. I can mention draw edit which is the main class of the code. It creates with request to LLM send the request and receive the answer from LLM. Then it validates the answer and applies the instructions to the shapes. Clipboard is implemented using extra extraable service from libra to copy the request to the clipboard in manual mode. Settings are stored as config files but we may change this in the future and use libra configuration for that. We created GUI using libra API called comson.star. AWT. We have used LLM in some areas that I will describe. I have used GLM 5.3 model from Z AI for code assistance. to run the extension. I have used quen 3.88 for shape generation and it was good. And this laptop I only have 4 GB of GPU memory so I use quantry which is not that strong. Generating complex shapes takes a lot of time and [clears throat] it requires a few minutes to generate output. Here manual mode helps. Extension copies the request to the clipboard. Then you may use a web based free service get the output and paste the response. Otherwise, you need to buy LLM API subscription and pay per tokens that you use. These are some of the challenges. If you want to get good results, you should either wait for a long time, even 5 to 10 minutes, or pay for expensive API price or buy expensive AI hardware with a lot of RAM. That is a problem with the LLM itself, and I hope the situation improves in the future. Handling more complex shapes is also challenging and I think this can also improve with AI improvements. And the last challenge is privacy. I will talk about this with more details. When using cloud-based services, privacy is an important issue. Our extension is connected while caring about privacy without privacy when creating the extension and it was one of the important consideration in the design. The use of local amma is a key here as it solves most of the privacy issues. Also, it's possible to check the input to AI and output from it both in terminal and also more carefully in manual mode. One drawback is that as said one will require a good AI hardware with a lot of RAM for AI to give good and fast results. Future plans are integrating with my previous writer extension, improving undo and redo, using sidebars, supporting more internal commands, and improving 3D support. Thank you very much. I want to show you demo of this extension. This is the menu of drawer extension. Now I want to click edit with AI and I want to ask this operation create a chicken that is in the tree color in the center of the page and okay I use manual mode Then this request copy to clipboard I paste it here in AI tool and that's Z AI and I'm waiting for response. This is the response of Z AI. I copy it and paste here. Okay, this is the result. And now I want to change the color or some of pieces of this shape. I select it and once again edit the oil All right. Okay, this is the response. And as you see the color of the eyes change to green. Okay. Now I want to change the color of the chicken. I select this shape. Draw my eye. Okay. The color of the chicken change to pink. Now I want to add some shapes. Then I select this shape AI edit with AI and create a table near the tree and put a and the table. Okay, this is the result. Thank you very much.