ai-autograding-feedback (R Version)

Overview

This is the R-based implementation of the AI autograding feedback system. It generates structured LLM-powered feedback on student assignment submissions across code, text, and image scopes. The system uses markdown prompts and templates, and supports multiple LLM backends including OpenAI and Claude (via API keys).

This version exposes a direct main() function instead of using a CLI, making it easier to call programmatically or from within R scripts, notebooks, or services.

Features

• Supports image, text, and code scopes

• Function-based interface (no CLI)

• Uses .md prompt templates and markdown output formats

• Modular file structure

• LLM support via OpenAI API (Claude and remote extensible)

Main Function Parameters

Parameter	Description	Required
prompt	Path to prompt file	❌ **
prompt_text	Inline string prompt to concatenate or use in place of file	❌ **
prompt_custom	Direct prompt content as string	❌ **
scope	Processing scope (code, text, image)	✅
submission	Submission file path	✅
solution	Solution file path	❌
model	Model name: openai, claude, remote	✅
remote_model	Optional remote model string (used with remote models)	❌
output	Markdown output filepath	❌
submission_image	Path to image (image scope)	❌
solution_image	Path to reference image	❌
output_template	Path to output template file	❌
system_prompt	Path to system prompt file	❌
question	Exact question/subquestion to test	❌
marking_instructions	Marking instructions/rubric file path	❌
json_schema	JSON schema for structured response format	❌

Note: You must provide one of prompt, prompt_text, or prompt_custom.

---

Scope

The behavior of main() adapts based on the selected scope:

Code

• Input: student + solution code files (e.g., .R, .py)
• Use case: error analysis, hint generation, annotated feedback

Text

• Input: student + solution PDFs or text files
• Use case: compare student writing to rubric

Image

• Input: submission image + optional reference image (supports QMD/RMD with automatic plot extraction)
• Use case: plot evaluation, visual structure checking

---

Example Usage

In R (script or console):

library(aifeedbackr)

main(
  prompt = "tests/fixtures/code_example/instructions.md",
  scope = "code",
  model = "openai",
  submission = "tests/fixtures/code_example/fail_submission/fail_submission.R",
  solution = "tests/fixtures/code_example/solution.R",
  output = "output/q1.md"
)

---

Prompts

Prompts may include placeholders such as:

• {context}
• {file_contents}
• {file_references}
• {submission_image}
• {solution_image}
• {marking_instructions}

---

Output Templates

Output templates can be provided as file paths via the output_template parameter. They support the following placeholders:

• {model} – Model name used
• {request} – Final constructed prompt
• {response} – Model response
• {timestamp} – Generation time
• {submission} – Path to submission

Default behavior: If no template is provided, outputs only the {response} content.

---

File Structure

aifeedbackr/
├── R/
│   └── code_processing.R
│   └── text_processing.R
│   └── image_processing.R
│   └── main.R 
│   └── template_utils.R
│   └── constants.R
│   └── ClaudeModel.R
│   └── OpenAIModel.R
│   └── RemoteModel.R
├── man/
├── inst/
│   └── markus_test_scripts/
├── tests/
│   ├── fixtures/
│   ├── testthat/
│   └── testthat.R

---

Setup Instructions

1. Install Dependencies

Imports:

• base64enc
• callr
• dotenv
• ggplot2
• httr
• jsonlite
• magick
• optparse
• pdftools
• R6
• readr
• stringr
• tools
• withr

Suggests:

• testthat (>= 3.0.0)
• here

2. Set API Key

Create a .env file in the project root:

OPENAI_API_KEY=your_openai_key_here
CLAUDE_API_KEY=your_claude_key_here
REMOTE_API_KEY=your_remote_key_here

---

License

This project is derived from the original Python version and follows the same academic fair use and research-oriented licensing assumptions.