Complete Guide to Managing and Changing the Default Model in Open WebUI

¿How do I change the default model in Open WebUI? If you've ventured into the world of local artificial intelligence, you probably know that the pair of Open WebUI and Ollama It's simply the winning combination. While Ollama handles the heavy lifting by running the models on your hardware, Open WebUI provides a visual layer worthy of ChatGPT, allowing you to manage everything without your data ever leaving your network. It's an incredibly powerful tool, although the settings might initially leave you a little confused.

Many people rack their brains trying to understand how set a default model or how to apply parameter settings to multiple models at once without having to manually configure them one by one. The reality is that the system works in layers: you have the base models and then the presets or custom models that act as wrappers. In this article, we'll break down how to master this interface so you don't waste time and your AI behaves exactly as you want it to.

Global and Default Parameter Configuration

For those who find it counterintuitive to modify each model individually, there is an elegant solution. Administrators can define capabilities and basic parameters that affect all models in the instance. This is done by navigating to the Administration Panel, going to Settings, then to Models, and clicking the gear icon.

Within this menu we find the DEFAULT_MODEL_METADATAThis is used to establish basic functions such as web search, vision, and the code interpreter for the entire system. If a specific model has a different configuration, that will prevail over the global one, but if it has nothing defined, it will take the default value. Similarly, the DEFAULT_MODEL_PARAMS It allows you to set the temperature, Top P, and maximum tokens, preventing you from having to configure each chat from scratch.

Creation of custom models and agents

One of the gems of Open WebUI is the ability to create custom models in the workspace. It's not about creating a new model from scratch (which would require training), but about creating a preset that builds upon a base model. Imagine you're using Llama 3 but you want it to always act as a code review expert, following your company's rules; instead of writing the prompt each time, you create a "Code Reviewer" model with its own instructions.

By configuring these agents, you can link knowledge basesspecific tools and system prompts. You can even use dynamic variables such as {{ USER_NAME }} o {{ CURRENT_DATE }} so the AI ​​knows who it's talking to and what day it is. The result is a specialized agent that you can share with your team or keep private, greatly simplifying the end-user experience.

Visibility and Access Management for Models

There is a crucial technical detail that often causes errors: for a custom model to work, the user must have access to the base modelIf you create an agent based on GPT-4o but restrict access to the original model, the system will throw a "Model not found" error. The smartest solution is to leave the base model public but hide it from the selector using the three dots in the menu.

This way, users will only see the optimized, user-friendly model you created in their dropdown list, while the base model continues to work in the background. If you need some advanced users to see the raw model, the only alternative is to add the same LLM through a second connection with independent access controls.

Integration of RAG and Local Knowledge

Open WebUI It's not just a pretty face; it includes a system of Native RAG (Recovery Augmented Generation)This means you can upload your own PDFs, text files, or Word documents to the Knowledge section of the Workspace. The system breaks down the information and stores it in ChromaDB, allowing the AI ​​to consult your documents before responding.

For this to fly, it is essential to configure the embedding modelnomic-embed-text is one of the most recommended options. A trick to prevent AI overload is to adjust the fragment size; if you're using small models with reduced context windows, it's advisable to lower the fragment size to around 800 tokens. leave room for the question from the user.

Enhancing the System with Pipelines and MCP

If you want to take the tool to the next level, you should take a look at the Pipelines and FunctionsFunctions allow you to run Python code directly in the interface, whether to filter messages (such as deleting emails for privacy reasons) or to create actions that save data to Notion. Pipelines, on the other hand, are external services that can act as routers for complex models or filters.

Furthermore, the latest versions support the Model Context Protocol (MCP)This allows your local model to connect to external tools like GitHub or Slack without needing to program a pipeline from scratch. Essentially, it's the cleanest way to equip your AI with actual operational capabilities in your daily workflow.

Troubleshooting Common Connection Problems

It's very common for the models menu to appear empty after installation. This usually happens because the Docker container can't communicate with the Ollama server. On Linux systems, it's essential to add the flag –add-host=host.docker.internal:host-gateway during container execution and ensure that Ollama's URL is pointing to http://host.docker.internal:11434.

Another common problem is that Ollama is only linked to the address 127.0.0.1. To fix this, you need to configure the environment variable. OLLAMA_HOST=0.0.0.0:11434 in your machine's systemd service. Once this is resolved, the models downloaded via ollama pull They will appear instantly in the web interface without needing to restart the server.

By mastering the creation of presets in the Workspace and adjusting the global parameters of the admin panel, you can transform Open WebUI from a simple chat into a productive agent ecosystem. The key lies in leveraging the flexibility of custom models and the power of integration with Ollama to maintain privacy and full control over the on-premises AI.