AutoRAG

A RAG system is only as good as its configuration — chunking, embeddings, retrieval settings, language handling. Those choices sit far outside what most teams can make, so they end up stuck between generic chatbots that answer from nowhere and developer infrastructure they can't operate. AutoRAG's bet: capture that expertise once, and let people configure it by describing their situation in plain language.

UI/UX design and frontend development inside the graduation-project team. I owned the product surfaces end to end — the guided setup, chat, document workspaces and team dashboard — designed them in Figma, then built the web frontend you can use in the live demo. Teammates focused on the retrieval engine behind it.

A fixed academic timeline (04/2025 – 05/2026), a small team, and one hard requirement: answers must stay explainable. We validated the setup flow the way the product would be used — by putting non-technical peers in front of it and asking them to configure an engine for their own documents, in Arabic or English, without help.

I started from the full configuration space and worked backwards: which decisions actually change retrieval quality, and which can be inferred? That mapping collapsed a settings panel into eight plain-language questions. From there: flows and wireframes for the four surfaces, a Figma system in the product's cobalt language, then the frontend build — where the chat and citation patterns went through the heaviest iteration.

• A questionnaire, not a settings panel. Progressive questions with defaults beat exposing thirty parameters — the system holds the expertise, the user holds the context.
• Sourced answers by default. Every response cites its documents, and every file shows how often it's used in answers — trust is the product.
• Documents as a first-class workspace. Upload, indexing state and usage live in one place instead of hiding behind the chat.
• Bilingual from the first question. Arabic, English and mixed corpora are a setup choice, not an afterthought.

1. Guided setup. A plain-language questionnaire configures the whole engine — no settings panel.
2. Ask anything. Natural-language chat that answers with structured, sourced summaries.
3. Document workspaces. Upload, index and manage knowledge sources; see how often each file gets used.
4. Team & insights. Members, roles and usage analytics for shared workspaces.

Delivered and defended as our Computer Information Systems graduation project at King Saud University in 2026, with the working prototype — not slides — carrying the demo. What it taught me: designing for trust in AI answers is interface work as much as model work; progressive disclosure is a kindness, not a limitation; and building what you design changes what you design. Made with Figma and a hand-built web frontend, deployed on Netlify.