Open WebUI: An Open, Extensible, and Usable Interface for AI Interaction

TL;DR

Open WebUI addresses the need for an open, extensible, and usable interface to interact with multiple local and hosted LLMs. It introduces a two-pronged plugin architecture that enables Tools, Pipes, Filters, and Actions, plus a community platform for sharing model presets and prompts. The evaluation uses in-the-wild engagement, plugin activity, and a user survey to demonstrate social and practical value, including openness, multi-model support, and local-data privacy. The work informs HCI design for future LLM toolkits, highlighting onboarding, community-driven extensibility, and governance considerations for locally deployed AI systems.

Abstract

While LLMs enable a range of AI applications, interacting with multiple models and customizing workflows can be challenging, and existing LLM interfaces offer limited support for collaborative extension or real-world evaluation. In this work, we present an interface toolkit for LLMs designed to be open (open-source and local), extensible (plugin support and users can interact with multiple models), and usable. The extensibility is enabled through a two-pronged plugin architecture and a community platform for sharing, importing, and adapting extensions. To evaluate the system, we analyzed organic engagement through social platforms, conducted a user survey, and provided notable examples of the toolkit in the wild. Through studying how users engage with and extend the toolkit, we show how extensible, open LLM interfaces provide both functional and social value, and highlight opportunities for future HCI work on designing LLM toolkit platforms and shaping local LLM-user interaction.

Figures (7)

• Figure 1: Many Model Interaction: Users simultaneously interact with multiple language models, such as Llama 3.1, GPT 4, and Mistral. The user submits a query ("Why is the sky blue?"), which is processed by selected models, generating concurrent responses (R1, R2, R3). Users review and compare these responses, selecting the most appropriate one to continue the conversation. This method enables users to harness the strengths of each model, as they can select the best response for their needs and, in the process, generate preference data over models.
• Figure 2: Open WebUI's admin settings facilitate easy AI model management through its graphical interface, enhancing accessibility by reducing dependency on command-line tools. The "Pull a model" feature enables users to download models simply by typing their names and clicking a button, with a progress bar displaying the download status. Additionally, the menu includes straightforward options for deleting models and uploading raw GGUF files, streamlining model management.
• Figure 3: Open WebUI's Prompt Preset feature, divided into four parts. (A) illustrates creating or importing Prompt Presets, where variables within square brackets are auto-selected for easy replacement using the tab key. Users also have the option to import community-shared presets. (B) demonstrates loading these presets into the chat input area using a forward slash command, enhancing user workflow. (C) highlights the efficient replacement of auto-selected variables, with the first variable selected by default and subsequent variables easily selectable with the tab key. (D) depicts submitting the customized prompt in Open WebUI, showcasing the feature's utility in streamlining repetitive tasks.
• Figure 4: A heatmap summarizing the topics in user-generated content from Github. The rows show the top three words in each topic and the columns show topics. Each cell shows how frequently a word (row) appeared in posts matching the corresponding topic (column).
• Figure 5: A heatmap summarizing the topics in user-generated content from platforms other than GitHub (YouTube, Reddit, LinkedIn, Hackernews, Twitter, and Medium). The rows show the top three words in each topic and the columns show topics. Each cell shows how frequently a word (row) appeared in posts matching the corresponding topic (column).