CASE STUDY 02 / RETRIEVAL-AUGMENTED GENERATION
University AI Assistant.
A conversational interface for exploring documents about AI adoption in universities, powered by a PDF-to-RAG pipeline.
Applied AI development
Public repository · Academic project
Documents · LangChain chunks
Local embeddings · Chroma
Question + retrieved context
Streamlit interface · Documents on AI in universities
The problem
Documents about adopting AI in higher education contain useful guidance, but finding a relevant passage can take time. The project explores a natural-language interface for asking questions about that material.
Its subject is university AI adoption; it is not an admissions service or a student-record system.
Context & role
I built an applied AI project that brings together document ingestion, retrieval, language-model generation, and a Streamlit interface. A separate university assessment feature supports the document assistant.
The solution
PDF text is split into chunks of 1,000 characters with a 200-character overlap. Ollama runs the nomic-embed-text embedding model locally, and Chroma stores the resulting vectors.
When a user asks a question, the pipeline retrieves the three most relevant chunks and passes that context to a Groq-hosted language model through LangChain. Streamlit provides the interface.
An optional Random Forest component assesses a university profile. It is supporting functionality, separate from the document retrieval path.
Technical decisions
Overlapping chunks retain some context across segment boundaries. Vector retrieval narrows the material passed to the language model rather than sending every document for every question.
Embeddings run locally through Ollama, while answer generation uses Groq. This is a hybrid architecture: the complete assistant is not fully offline.
Evidence
The public repository contains the application implementation in PROJET/app.py and project documentation in PROJET/README.md. The illustration on this page explains the architecture; it is not an application screenshot.
The source code makes the document ingestion, retrieval, and answer-generation stages available for inspection.
Limitations
Answers depend on the source PDFs, the retrieved passages, and the language model. The interface does not currently provide implemented answer citations or conversational memory.
A deployed version would need an evaluation set and additional work on grounding, error handling, and document updates. Generated answers should be checked against the underlying documents.
Explore the project
The source link opens the University AI Assistant repository, including the Streamlit application and its setup documentation.