The UIUC AI Teaching Assistant runs 11 parallel models for text, image retrieval, generation, moderation, and ranking, achieving a median 2-second response time. It leverages proprietary data from textbooks, lecture videos, and student Q&A forums, none publicly available due to licensing restrictions. A novel RLHF approach uses semantic search retrieval, built from a dataset created by hiring 5 Electrical Engineering students; this dataset is freely available on Hugging Face.

The assistant focuses on UIUC’s ECE 120 course. Evaluation is performed by generating answers for in‑house QA pairs from expert engineers and having GPT‑3 judge their quality againstらground truth. Though GPT‑3 tends to favor its own outputs, iterative evaluation remains crucialիջ for improving features. Evaluation results and code are publicly shared.

Usage is open source (excluding commercial textbook content). Users can plug in their own Pinecone database for retrieval‑augmented generation and run the Gradio web app. Installation requires Python 3.8 and the provided requirements; API keys and a document store are needed.

The project includes scripts for PDF and video transcription to Pinecone, main aggregator logic, Gradio UX, prompting, and evaluation modules. Run the app with `bash run_ta_gradio.sh`.

The full evaluation results and code are available on the project’s GitHub and Hugging Face pages.