Bulk AI Tool — Batch AI Processing Service 

Overview 

The Bulk AI Tool is an on-premises web application hosted on FAU's High Performance Computing (HPC) cluster that lets you run an AI prompt over every row of a dataset and collect structured output automatically. Instead of copying and pasting data into an AI chat tool one record at a time, you upload a spreadsheet or batch of documents, write a single prompt template, define what fields you want extracted, and the tool processes the entire dataset in parallel — returning clean, structured results you can download as a spreadsheet. 

How-To Guide & User Reference 

 

Overview 

The Bulk AI Tool is an on-premises web application hosted on FAU's High Performance Computing (HPC) cluster that lets you run an AI prompt over every row of a dataset and collect structured output automatically. Instead of copying and pasting data into an AI chat tool one record at a time, you upload a spreadsheet or batch of documents, write a single prompt template, define what fields you want extracted, and the tool processes the entire dataset in parallel — returning clean, structured results you can download as a spreadsheet. 

 

All processing runs entirely within FAU's infrastructure using the FAU HPC LiteLLM proxy. No data is sent to external cloud services. 

 

Example Use Cases 

• Extract sentiment, themes, or scores from hundreds of student course evaluations. 

• Classify or tag a large set of research abstracts by topic, methodology, or keywords. 

• Summarize or pull key fields from a batch of uploaded PDF reports or Word documents. 

• Standardize free-text survey responses into structured categories. 

• Enrich a CSV dataset with AI-generated fields such as language detection, categorization, or risk scores. 

 

What You Can Do 

• Upload a CSV, TXT, PDF, or Word file (up to 10 MB per file) as your input dataset. 

• Write a prompt template using {{Column Name}} placeholders that map to your data headers. 

• Define a structured output schema — name the fields you want and choose their data type (text, number, boolean, or list). 

• Use the built-in AI assistant to help you draft a good prompt and schema for your use case. 

• Test your prompt and schema on a single row before running the full batch. 

• Run the full batch and download results as a CSV or JSON file. 

• Track usage and run history through your session. 

 

Before You Begin 

• You have an active FAU NetID and password. 

• Your input data is in one of the supported formats: CSV, TXT, PDF, DOCX, or DOC. 

• Each file is 10 MB or smaller. 

• You are using a modern browser (Chrome, Firefox, Edge, or Safari — current versions). 

 

Note: The Bulk AI Tool uses FAU's Microsoft Entra ID for authentication. You will be redirected to the FAU single sign-on page automatically when you visit the service URL. 

 

Accessing the Service 

1. Open a browser and navigate to https://bulk.ai.hpc.fau.edu. 

2. You will be redirected to the FAU SSO login page. Sign in with your FAU email address and password. 

3. Complete multi-factor authentication (MFA) if prompted. 

4. Once authenticated, the Bulk AI Tool home page will load. 

 

Running a Bulk AI Job 

Step 1 — Upload Your Data 

1. In the Upload section, drag and drop your file into the upload area, or click to browse and select a file. 

2. The app will parse your file and display the column headers it detected. 

3. Review the headers. For CSV files, these come from the first row. For PDFs and Word documents, the app maps the document text to a set of default column names automatically. 

 

• Supported formats: CSV, TXT, PDF, DOCX, DOC. 

• Maximum file size: 10 MB per file. 

• For best results with tabular data, use a CSV with a clear header row. 

 

Note: You can also load a sample dataset from the built-in demo data to try the tool before uploading your own files. 

 

Step 2 — Write a Prompt Template 

The prompt template tells the AI what to do with each row. Use double curly braces around a column name to insert that row's value into the prompt. 

 

Example — if your CSV has columns named "Student Name" and "Comment": 

 

Analyze the following student course evaluation comment and rate the overall sentiment. 

Student: {{Student Name}} 

Comment: {{Comment}} 

• Column name placeholders are not case-sensitive and tolerate extra spaces (e.g. {{ student name }} works). 

• You can reference as many columns as needed in a single template. 

• Keep your prompt focused — the more specific the instruction, the more consistent the output. 

 

Step 3 — Define an Output Schema 

The output schema tells the AI exactly which fields to return for each row and what data type each field should be. This ensures results come back as clean, structured data rather than free-form text. 

 

• In the Schema section, click Add Field. 

• Enter a field name (letters, numbers, and underscores only — no spaces). 

• Choose a type: Text, Number, Boolean (true/false), or List. 

• Repeat for each field you want in the output. 

 

Example schema for the course evaluation use case above: 

 

 

Note: Field names must be unique. The schema builder will flag duplicates before you run the job, since duplicate names produce ambiguous output. 

 

Step 4 — (Optional) Use the AI Assistant 

If you are unsure how to write a prompt or structure your schema, the built-in AI assistant can help. Click the Assistant panel, describe what you are trying to accomplish in plain language, and the assistant will suggest a prompt template and output schema tailored to your dataset. 

 

• You can apply the suggested prompt and schema with a single click. 

• Continue the conversation to refine the suggestions before running. 

 

Step 5 — Test on a Single Row 

• Before running the full dataset, click Test Row. 

• The app will run your prompt and schema against the first row of your data and return the result immediately. 

• Review the output. If the fields are not what you expected, adjust your prompt or schema and test again. 

 

Note: Always test at least one row before running a large batch. This saves time and avoids running hundreds of rows with a poorly worded prompt. 

 

Step 6 — Run the Full Batch 

• Select the AI model you want to use from the model dropdown. The default model is Gemini 2.5 Flash, which is recommended for most tasks. 

• Click Run Bulk Job. 

• The app processes up to 20 rows at a time in parallel. A progress indicator shows how many rows have been completed. 

• When all rows are finished, the results appear in the Results section. 

 

Important: Large datasets (hundreds or thousands of rows) may take several minutes to complete. Do not close or refresh the browser tab while a job is running, as this will interrupt the batch. 

 

Downloading Results 

When a batch job completes, you can download the results in two formats: 

 

• CSV — Each row from your original dataset plus one column per output schema field. Ready to open in Excel or import into other tools. 

• JSON — The full structured output per row, including the original input values and AI-generated fields. 

 

Click the Download CSV or Download JSON button in the Results section to save the file to your computer. 

 

Note: If email delivery is enabled for your deployment, you can also choose to receive the results as email attachments when the job completes. Check with your Research Computing contact to confirm whether this feature is available. 

 

Choosing an AI Model 

The model dropdown lists the AI models available through the FAU HPC. A status indicator shows whether the proxy is currently reachable. 

 

Troubleshooting 

 

 

Data and Privacy 

• All files using local models are processed within FAU's on-premises HPC infrastructure, meaning no data is sent off FAU servers. If cloud is the selected provider, data is sent to the cloud provider but is under FAU's data privacy agreement with those providers.

• Uploaded files are held temporarily in memory during processing and are not stored on disk by the application. 

• Usage is logged per job (not per row) for service accounting purposes. Logged fields include your FAU email, the model used, the number of rows, and processing duration. Individual row data is not logged. 

• Do not upload documents containing sensitive personal information (FERPA-protected student records, HIPAA-protected health information, or other regulated data) without first confirming the appropriate data handling procedures with the FAU Office of Information Technology. 

• Questions about data classification or appropriate use: contact the FAU OIT Security team at security@fau.edu. 

 

Tips for Best Results 

• Keep prompts specific. Vague prompts produce inconsistent output. Tell the model exactly what you want: "Rate the sentiment as positive, neutral, or negative" is better than "Analyze the sentiment." 

• Use short, clear field names. Field names become JSON keys. Avoid spaces and special characters. Good examples: sentiment, confidence_score, is_actionable. 

• Test before running. Always use Test Row on at least one representative row before launching a large batch. 

• Start with a small batch. If you have thousands of rows, try a 50-row sample first to confirm the output quality meets your needs. 

• Use the AI assistant for complex tasks. If you are unsure how to structure the prompt or schema for a new use case, describe the goal to the assistant and apply its suggestion as a starting point. 

• Check the proxy status before large runs. The status indicator at the top of the page shows whether the FAU LiteLLM proxy is reachable. Start large jobs when the proxy is confirmed healthy. 

 

Contact and Support 

For issues with this service, submit a ticket through the FAU Help Desk or contact Research Computing directly.