Introduction to UVA RC GenAI | RC Learning Portal

University Policies

Mon, 20 Apr 2026 00:00:00 +0000

Using AI at UVA

UVA RC GenAI aligns with the University’s GenAI Use Guidelines. These can be found at Using AI at UVA: Guidelines & Terms of Use.

Research Data Usage

Mon, 20 Apr 2026 00:00:00 +0000

What Type of Research Data Can I Submit?

Approved: Any data suitable for Standard Security Zone (SSZ) (Rivanna/Afton), including sensitive University research data.
Prohibited: Highly sensitive data, public health information (PHI), or other controlled access data types.

Unauthorized use of UVA RC GenAI may result in disciplinary action under the University’s GenAI usage guidelines.

When in doubt: Contact us for support before submitting

What Data is Saved?

Prompts and responses are not saved
Metadata may be retained to better understand number of users, number of requests, demand for service, tokens generated and other metrics that will aid in ensuring a quality delivery of the service.
Logs are accessible only to admins

Access and Getting Started

Mon, 20 Apr 2026 00:00:00 +0000

UVA RC Gen AI is available exclusively to users with active Research Computing accounts. It’s accessible via both a chat-like browser interface and API within standard security zone systems (Afton/Rivanna).

Where can I Access UVA RC GenAI from?

On grounds: Direct browser access is available. API access must run from a standard security zone HPC node. See here for HPC access methods.
Off grounds: VPN is required for browser access and access to HPC nodes.

Additional details can be found in our user guide, and support resources can be found here on our website.

Connecting to UVA RC GenAI

Mon, 20 Apr 2026 00:00:00 +0000

Below are the first steps needed for all users regardless of access method:

Note: Accessing UVA RC GenAI requires an active UVA computing ID, research computing account and EServices password for Netbadge authentication. Duo two-factor authentication is required for browser access.

If you don’t have an RC user account please see how to get access to HPC Services
Once you have an RC account, you are ready to use UVA RC GenAI
- Ensure VPN connection if off grounds
Login with Computing ID and EServices password
Complete Duo 2FA

Terminology

Mon, 20 Apr 2026 00:00:00 +0000

The following are common terms used when working with UVA RC GenAI and LLMs in general.

Core Units

Token - The basic unit of text LLMs process. Can be a whole word, part of a word, or a single character. Roughly 100 tokens ≈ 75 words. All usage, limits, and costs are measured in tokens.
Prompt - The input text you send to the model. Prompts can be a question, instruction, or conversation history. A prompt is measured in tokens.
System Prompt - A hidden prompt that sets the model’s behavior rules for the entire conversation—such as tone, role, or safety constraints. It frames how the model interprets all subsequent user prompts.

Capacity Limits

Context Window - The maximum number of tokens (input + output combined) a model can consider at one time.
Batched Token Length - The total tokens across multiple requests sent together as a single batch. Batching can improve efficiency; this is the sum of all tokens in that group.

Generation Parameters (API access only)

Temperature - Controls output randomness. Lower values (0.1-0.3) produce more focused, deterministic responses; higher values (0.8-1.0) produce more varied and creative outputs.
Seed - An integer that initializes the random number generator. Using the same seed with identical prompts and settings produces reproducible outputs.

Usage and Operations

Metadata - Non-content information about an API exchange: timestamp, model version, token counts, request ID, cost.
Rate Limiting - Restrictions on tokens or requests per time period to prevent overload and ensure fair access.

Browser Access

Mon, 20 Apr 2026 00:00:00 +0000

After signing in, you should have browser access to UVARC GenAI through the OpenWebUI interface.

Here, you can chat through the conversational interface, adjust integrations (e.g., web search), or even upload and attach content to the chat session.

Files can be loaded into the web interface – supported extensions include: pdf, docx, txt, md, csv, png, jpeg, jpg, pptx, xls, xlsx, json, sh, html, htm, xhtml, js, and py.

Note: Chats are not saved. Conversation history disappears when you close the browser tab, sign out, or if the session expires.

More on data management will be discussed later.

API Access

Mon, 20 Apr 2026 00:00:00 +0000

API access lets you send requests to UVA RC GenAI programmatically from your code, instead of typing into a chat window. This enables automation and integration with existing workflows like bulk processing, large-scale data analysis, and connection to other tools.

Use Cases

Browser Access:

Asking a single research question or a quick explanation
Brainstorming ideas for a project or paper
Drafting or polishing a single email or paragraph
Experimenting with prompts

API Access:

Processing hundreds or research papers to extract summaries
Building a pipeline that automatically categorizes and tags new data as it arrives
Integrating the LLM into a custom internal tool

Overall, use the browser when you’re interacting directly, and use the API when you want the LLM to work inside your existing workflows to handle volume without you having to type each prompt.

Getting your API Key

Sign in to UVA RC GenAI
Click profile (top right) → Settings → Account
Select “Show” next to “API Keys”

You will have the option to view, copy or create a new API key.

Securely Storing Your Key

Use environment variables to safely store your key (e.g., export UVARC_GenAI_API="your-key-here"). Make sure to never commit keys to code repositories, and regenerate keys in UVA RC GenAI browser if compromised.

Note: You need to be on a compute node to run your code.

HPC compute nodes can be accessed either via OpenOn Demand (JupyterLab, Desktop, etc) or through an ijob from the command line.

Code Examples

Each of the following examples prompts “Hello” to UVA RC GenAI via API.

Curl Example

Contents of uva-rc-genai-curl.sh

export UVARC_GenAI_API="<yourAPIkey>" 
curl -X POST "https://open-webui.rc.virginia.edu/api/chat/completions" -H "Authorization: Bearer $UVARC_GenAI_API" -H "Content-Type: application/json" -d '{"model": "Kimi K2.5", "messages": [{"role": "user", "content": "Hello"}]}' | grep '^data: ' | grep -oP '"content":"\K[^"]*' | tr -d '\n' && echo

Python with OpenAI Library

Contents of uva-rc-genai-oai.py

import os
import openai

client = openai.OpenAI(
    base_url="https://open-webui.rc.virginia.edu/api/",
    api_key=os.environ.get("UVARC_GenAI_API")
)

response = client.chat.completions.create(
    model="Kimi K2.5",
    messages=[{"role": "user", "content": "Hello"}],
    stream=True
)

# Handle streaming response
full_text = ""
for chunk in response:
    if chunk.choices[0].delta.content:
        full_text += chunk.choices[0].delta.content

print(full_text)

Saving Outputs

Outputs can be saved programmatically by capturing the response from your HTTP client (like requests or openai in Python, axios in Node.js, or curl in shell) using standard I/O methods to save the results to your storage.

For example, pipe curl output directly to a file with > output.txt, or in Python use

with open("output.txt", "w") as f:
    f.write(response_text)

Token Tracking

It’s important to monitor token usage to stay within limits and ensure requests don’t exceed the context window. Token counts are included in every API response. For example, in the OpenAI library:

print(response.usage.prompt_tokens)      # Input tokens
print(response.usage.completion_tokens)  # Output tokens
print(response.usage.total_tokens)       # Combined

Similar fields exist in the Anthropic SDK and other tools.

Claude Code

Mon, 20 Apr 2026 00:00:00 +0000

Claude Code on HPC

Claude Code is an AI-assistant for software engineering tasks that runs directly in the terminal to write, edit and analyze code. You can link it to UVA RC GenAI.

Installing Claude Code

First, you’ll need to install claude code into your home account on a login node with the following:

curl -fsSL https://claude.ai/install.sh | bash

once the above command is finished running, run the following to add ~/.local/bin to your path:

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc && source ~/.bashrc

Claude Code with UVA RC GenAI

Set these environment variables:

export API_KEY="<your_api_key>"
export ANTHROPIC_BASE_URL="https://open-webui.rc.virginia.edu/api"
export ANTHROPIC_AUTH_TOKEN=$API_KEY
export ANTHROPIC_API_KEY=""
export CLAUDE_CODE_MAX_OUTPUT_TOKENS=4096

Then launch with: claude --model 'Kimi K2.5'

claude will run in whatever directory it’s launched in.

Jupyter Notebooks

Mon, 20 Apr 2026 00:00:00 +0000

Jupyter Notebook Workflow

UVA RC GenAI can be accessed from a Jupyter notebook using the python API client to integrate LLM capabilites for data analysis, code generation, and automated research workflows.

Download this ZIP file for a Jupyter notebook reference.

Research Use Cases

Data Management

Mon, 20 Apr 2026 00:00:00 +0000

Methods for Saving Chats

Methods for saving chats via API were discussed earlier, but you also have the option to export or save chats through the OpenWebUI interface. Clicking the three dots next to your icon in the upper-right section of the interface displays the options to either download or copy the chat session.

The chat can be downloaded in three formats (.txt, .pdf, .json). Saving as a .txt file stores only your prompts and the LLM’s responses/reasonings; however, saving as a .json includes metadata (e.g., timestamps, model, version, message IDs, etc), which may be useful in certain cases.

The “Copy” option copies only prompts, responses and reasonings, similar to what’s stored in the .txt file.

Storage Options

Downloaded browser chats will be saved to your local workstation but can be easily copied to HPC storage. If you’re saving chats on HPC storage it’s important to remember the limitations and use cases for each of RCs storage options. See here for details on our complementary storage options and here for leased options.

Reproducibility Considerations

Reproducibility controls are important since LLM outputs can vary across different runs. This makes it impossible to verify results, replicate research, or trace back errors. The following are some important considerations regarding output reproducibility.

Temperature and Seed - Set temperature=0 and a fixed seed when consistency matters. Higher temperature introduces randomness.
Prompt Versioning - Track prompts in versioned files since even small wording changes can affect outputs.
Saving Outputs - Save raw API responses with metadata (temperature, tokens, request ID) rather than just extracted text.
Context Window - Document the full context window, including system prompts, prior conversation history, and any retrieved content; exceeding the window silently drops content.

Current Limitations

Mon, 20 Apr 2026 00:00:00 +0000

Server Limitations

Context window: 65536 tokens
Batched Token Length: 131072 tokens
Rate Limiting: 60 requests per minute

You may receive errors if any of these limits are reached. Errors may highlight that a rate limit has been exceeded if more than 60 requests are made within a minute, or you may receive context overflow errors if you surpass the maximum context length. Responses may cut off mid-sentence when the total token budget is used.

Workflow Challenges

If you’re working with large documents, you may need to break them into smaller pieces to fit within the context window. High-volume workflows like processing thousands of files or running automated batches can quickly hit rate limits, so you’ll need to pace your requests. Long-running conversations or agentic workflows can exhaust the context window as chat history and file constraints accumulate.

Best Practices

Track token usage - Monitor prompts and responses to avoid silent truncation.
Avoid large file uploads - Split files into more maintainable chunks (if possible) to prevent filling context.
Saving Outputs - Save full API output or OpenWebUI chat data (.json format) for debugging and reproducibility.
Restart sessions when needed - Restart session to clear context or use /clear command in Claude Code to clear context window. You can also use the /compact command in Claude Code to clear context while maintaining a summary of your chat.

FAQs

Mon, 20 Apr 2026 00:00:00 +0000

Research Computing will continuously assess enhancements and emerging LLMs to provide the UVA research community with modern, evolving AI services. Improvements may include installing different models to expanded capabilities or adding additional hardware for enhanced performance. UVA RC GenAI is an early-access deployment and support is limited. Please contact us through our support request form with any questions, concerns, or suggestions.

Introduction to UVA RC GenAI | RC Learning Portal

University Policies

Using AI at UVA

Research Data Usage

What Type of Research Data Can I Submit?

What Data is Saved?

Access and Getting Started

Where can I Access UVA RC GenAI from?

Connecting to UVA RC GenAI

How to Login

Terminology

Core Units

Capacity Limits

Generation Parameters (API access only)

Usage and Operations

Browser Access

API Access

Use Cases

Getting your API Key

Securely Storing Your Key

Code Examples

Saving Outputs

Token Tracking

Claude Code

Claude Code on HPC

Installing Claude Code

Claude Code with UVA RC GenAI

Jupyter Notebooks

Jupyter Notebook Workflow

Research Use Cases

Data Management

Methods for Saving Chats

Storage Options

Reproducibility Considerations

Current Limitations

Server Limitations

Workflow Challenges

Best Practices

FAQs