URL: chat.ai.college.ucsb.edu
Sign in: UCSB NetID
Click the model name at the top of the chat window to open the model picker. Four models are available:
| Model | When to use it |
|---|---|
| gemma‑4‑31b | General‑purpose work that benefits from a very large context window (up to 128 k tokens): long‑form writing, summarisation, brainstorming, email drafting, multi‑turn conversations. |
| qwen3‑coder‑next | Code‑centric tasks – generation, debugging, code‑review, technical explanations – and fast inline autocomplete (FIM support). |
| gpt‑oss‑120b | Deep, multi‑step reasoning on long documents, research synthesis, or any scenario where you need maximal model capacity and nuanced analysis. |
| granite‑4.1‑30b | Robust on-prem alternative model from IBM - a balanced 30B parameter model that offers strong language understanding and lower latency than the larger 120B model. |
You can switch models at any point. Each conversation keeps its own model selection.
A note on model size: These are capable open-weight models, but they are smaller than frontier commercial models like GPT-4o or Claude Sonnet. See Getting Better Results below for tips on how to get the most out of them.
Web search is enabled by default. When active, Open WebUI uses an external search engine to retrieve current information and include it with your query.
Click the search icon (🔍) in the message input bar to toggle web search for the current conversation. A highlighted icon means search is active.
> ⚠️ Data reminder: When web search is enabled, your query is sent to an external search engine. Avoid sensitive or restricted information in search-enabled conversations.
A system prompt is a persistent instruction that shapes every response in a conversation. It's sent to the model before any user message — think of it as setting the model's role, tone, and constraints for the session.
Click the settings/sliders icon in the conversation toolbar and select System Prompt. Type your instructions and save. The prompt applies to the current conversation.
To reuse a prompt, copy it and paste it into new conversations as needed, or save it as a Preset (see below).
Good system prompts are specific, concise, and task-oriented. They work best when they define:
Examples:
You are a Python code reviewer. When given code, identify bugs, suggest improvements,
and explain your reasoning. Always include a corrected version of the code.
You are a research assistant for a marine biology lab. Summarize papers clearly
for a non-specialist audience. Use plain language. Avoid jargon unless defined.
You are a writing editor. Review text for clarity, conciseness, and tone.
Return a revised version followed by a brief explanation of the changes made.
Keep responses under 500 words unless the input is very long.
Open WebUI allows you to save system prompts as named presets for reuse. Go to Settings → Presets to create and manage them. Presets can be loaded at the start of any new conversation.
| Focused Retrieval (default) | Full Context | |
|---|---|---|
| How it works | RAG — finds and injects the most relevant chunks based on your query | Injects the entire file into every message, no chunking |
| Best for | Large documents, knowledge bases, multi-file collections | Short files, style guides, files where all content matters |
| Context usage | Low — only top-k chunks | High — entire file every message |
| Risk | May miss content not ranked relevant | Can exceed context window on large files |
If you don't see the toggle, your Open WebUI version may label it differently or place it under a file options menu.
Admin Panel → Settings → Documents → enable Full Context Mode
⚠️ This disables RAG and hybrid search for all uploads, including knowledge bases. Knowledge collections will inject all documents, not just relevant ones. Use with caution in multi-user environments.
Full Context mode injects the entire file on every message. Before switching:
ContextWindowExceededErrorFor files that exceed any model's context window, use a map-reduce approach: summarize in chunks, then combine the partial summaries.
If your model has native function calling enabled, attached knowledge is not auto-injected — the model must call retrieval tools to access it. If the model ignores attached files:
list_knowledge and query_knowledge_files