AI-Powered Educational Web Assistant (RAG-Based)

TEAM-IFI

Yvan Valdes TCHOUTOUO TCHEUNJIO, Heritier MAWANDU HAMBA, Kadidiatou BERTHE, Wenchel RIDORE, Landry IRUMVA, Université Nationale du Vietnam Hanoï - École Internationale

Training / education / pedagogy

In universities in Sub-Saharan Africa (Kinshasa, Abidjan, Yaoundé), students lack textbooks and stable Internet access. Our AI-based educational web assistant allows teachers to upload their course PDFs, after which students can ask questions in natural language in French or in a local language. The AI retrieves relevant passages using a RAG approach (semantic search) and generates a sourced response citing the exact page. The tool works on smartphones and does not require a stable Internet connection.

This situation raises major issues of educational equity. In Sub-Saharan Africa, the ratio is often one textbook for more than 10 students, libraries are under-resourced, and Internet access remains costly and unstable. Existing AI tools (ChatGPT, etc.) require high-speed connectivity, are primarily in English, and are not based on local educational content creating a growing digital divide in education. Our solution transforms this challenge into an opportunity: it anchors AI within real African curricula, cites exact local sources, operates in a frugal mode (models with fewer than 3B parameters, < $0.001 per query), and integrates African languages. The impact extends to learning (24/7 access to course content), equity (a smartphone is sufficient), cognition (sourced answers that encourage critical thinking), and sustainable development (eco responsible, low-cost, and replicable AI).

We are participating in this challenge because we experience the reality it addresses on a daily basis. As a team familiar with the realities of Africa, we designed this assistant not as a theoretical exercise, but as a response to a lived need: our classmates and colleagues do not have access to the same resources as students in the Global North. We aim to better understand how AI can truly be inclusive not simply translated, but designed for low-resource contexts. This challenge would allow us to examine the biases of current models, transform our prototype into a tool validated by real impact data (through university pilot programs), and build a bridge between technological innovation and African educational realities. We want to demonstrate that frugal, locally grounded AI can compete with the most expensive solutions.

The Use Case In many universities — especially in developing countries — students face a common struggle: limited access to teachers outside of class hours, combined with dense, hard-to-navigate course materials (often long PDFs with little structure). When exam season hits, students are left on their own to dig through hundreds of pages trying to find the one paragraph that answers their question. It's frustrating, time-consuming, and honestly, it shouldn't be this hard in 2026. That's the problem we identified: How can we provide every student with a personal, always-available teaching assistant who actually knows their course material? The Proposed Solution We're building an AI-powered educational web assistant that uses a technique called Retrieval-Augmented Generation (RAG). In plain terms: the teacher uploads their course PDFs, and the system turns them into a smart, searchable knowledge base. Students can then ask questions in natural language — in French, English, or even Spanish — and get precise, sourced answers pulled directly from the course documents. Here's how it works under the hood: Document Ingestion — The teacher uploads PDF manuals. The system extracts text page by page (using PyMuPDF), cleans it up, and splits it into small, meaningful chunks (~1000 characters each, with overlap to avoid cutting ideas in half). Semantic Indexing — Each chunk is converted into a numerical vector (an "embedding") using a multilingual model (intfloat/multilingual-e5-small). These vectors are stored in a ChromaDB vector database, which allows for lightning-fast semantic search — meaning the system understands meaning, not just keywords. Intelligent Q&A — When a student types a question, the system converts it into a vector, finds the most relevant chunks from the course material, and feeds them as context to a Large Language Model. The LLM (we support Groq/LLaMA, OpenAI, and Google Gemini) then generates a concise, accurate answer — always citing the source document and page number. Session Isolation — Each student gets their own isolated session, so uploaded documents and conversations are private and don't interfere with other users. Why This Approach? It actually stays on topic. The LLM is instructed to answer only from the provided course material. No hallucinations about things that aren't in the syllabus. Multilingual by design. The embedding model and the LLM prompt both handle multiple languages, so a student can ask in French and get a French answer, or switch to English — it adapts automatically. Privacy-first. Everything can run locally (with Ollama), or through cloud APIs — the choice is the institution's. No student data is sent anywhere without explicit configuration. Teacher-friendly. The teacher doesn't need to do anything special — just upload their existing PDFs. The system handles the rest. What's Next This is the first iteration. The natural next steps include adding a feedback loop (students can rate answers to improve quality over time), a conversation history feature, and eventually a teacher dashboard with analytics on what students ask most, which could genuinely help teachers improve their courses. Streamlit link : https://assistant-web-educatif.streamlit.app/