Devpost Submission: QuestMap.AI

Inspiration

Learning a complex new subject is often overwhelming. You’re met with a "wall of data"—disorganized PDFs, long YouTube videos, and generic syllabi that don't care about what you already know. We wanted to build something that acts as a GPS for Knowledge. Inspired by the need for personalized education, we built QuestMap to turn static documents into dynamic, 3D learning journeys that adapt to the student in real-time.

What it does

QuestMap is an AI-powered learning path generator that uses Retrieval-Augmented Generation (RAG) to create hyper-personalized curriculums.

Upload Materials: Users upload textbooks, notes, or exam results.
Visual Learning Path: The AI generates an interactive learning path with Key concepts tags.
Strict Grounding: Every practice scenario and recommended resource is strictly grounded in the RAG approach—eliminating AI hallucinations.
RAG Relevance Guard: Our custom retrieval logic ensured that even if you switch topics, your learning path stays isolated and focused, preventing "context pollution" from previous uploads.

How we built it

The Brain: Powered by Google Gemini 2.5 Flash, we used advanced prompt engineering for structured JSON generation and strict RAG grounding.
Vector Search: We used Pinecone to store and retrieve document embeddings in real-time.
Frontend: A high-performance React (Vite) UI with Framer Motion for liquid animations and Lucide for a premium aesthetic.
Backend: A Node.js/Express server that parallelizes RAG retrieval and AI generation to keep the experience snappy.
Persistence: MongoDB Atlas stores user profiles and historical learning trajectories.

Challenges we ran into

Context Pollution: Early on, documents from previous sessions (like research papers) were bleeding into new, unrelated topics. We solved this by implementing a Relevance Guard with a high semantic threshold (0.65) and "Domain-Aware" AI instructions.
Hallucinated URLs: AI often makes up YouTube links. We mitigated this by generating specific search queries and resolving them via APIs to ensure every resource link actually works.
UI Lag: Generating a map, quizzes, and resources simultaneously takes time. We moved to a parallelized architecture on the frontend to fetch data in chunks, keeping the user interface responsive.

Accomplishments that we're proud of

True Grounding: Building a system where the AI tutor actually says "Based on your uploaded context..." and refuses to make up jargon that isn't in your text.
The Aesthetics: Creating a dashboard that feels less like a spreadsheet and more like a high-end command center for learning.
Semantic Separation: Successfully engineering the RAG pipeline to distinguish between broad foundational topics and narrow specialized research.

What we learned

Prompt Reliability: We learned that responseMimeType: 'application/json' is a lifesaver for building stable AI-driven applications.
Semantic Nuance: We discovered that "related" doesn't always mean "relevant"—and learned how to tune vector search to respect strict domain boundaries.

What's next for QuestMap

Collaborative Quests: Shared learning maps for study groups.
Mobile Companion: AR-powered learning maps you can explore on your phone.
Advanced Assessment: Deep-link quizzes that identify precisely which paragraph of your textbook you need to re-read.