Navigator

Your AI-powered academic assistant — upload your syllabus, ask anything, study with generated material.

Features · How it works · Stack

---

The Problem

Students receive syllabi as PDFs and lose them in a folder. When exams arrive, they don't know which topics are prerequisites for others, which assessment comes first, or have personalized practice material. The information is there — but it is not accessible, not interactive, and doesn't adapt to what each student needs to review.

The Solution

Navigator transforms any academic syllabus into a complete learning system:

1. Upload your PDF — Navigator parses it, chunks it, and generates vector embeddings.
2. Ask the chat — Answers grounded in the real syllabus content, with exact citations. No hallucinations.
3. Visualize the knowledge map — An editable graph of topics and prerequisites, generated automatically.
4. Study with adaptive material — Flashcards, quizzes, summaries, and mind maps generated from your material, with configurable difficulty and anti-repetition.
5. Check your agenda — Evaluations, weekly topics, and review recommendations crossing all your courses.

A single app. A single deploy. Zero additional infrastructure.

---

✨ Features

Feature	Description
RAG Chat	Questions about the syllabus with grounded answers and real PDF citations. Schedule-aware: knows which evaluations you have this week.
Knowledge Graph	AI-generated topic and prerequisite graph. Fully editable: add, rename, connect, and save nodes. Cycle-validated.
Study Area	6 modes: flashcards (3D flip), dynamic quiz, simulation exam, review mode, mind map, and auto summary. Choose difficulty and topic.
Smart Agenda	Monthly calendar with events detected from the schedule. Per-day notes. Recommendations of what to review based on prerequisites.
Multi-course	The chat searches across all your courses. The agenda crosses evaluations. The study area focuses on your weakest topics.
Guest Access	Use the app without an account: the PDF is processed and auto-deleted in 24 h. Upon signup, data becomes permanent.
SSE Streaming	Real-time chat answers. Final event carries title, citations, provider, and model.

---

🔍 How it Works

PDF → parse (unpdf) → chunks per page → embeddings (OpenAI) → pgvector (Neon)
                                                         ↓
                                          topic graph  (structured output)
                                          schedule     (structured output)
                                          study material (multi-agent)
                                                         ↓
Chat: question → embed → cosine similarity → hybrid rerank → context + agenda → LLM → SSE

Everything lives in a single full-stack Next.js app. No separate backend, no Python, no Docker in production. The entire RAG pipeline (ingestion, embeddings, retrieval, generation) is implemented in TypeScript and runs as Next.js API routes.

Ingestion is async in 2 phases: the upload responds immediately; a durable job queue (FOR UPDATE SKIP LOCKED) processes embeddings and generates the graph/schedule in the background with retries and exponential backoff.

RAG Pipeline Layers

Layer	What it does
Ingestion (sync)	PDF parsing (unpdf), chunking by page (1200 chars / 120 overlap), magic bytes validation, SHA-256 hash. Stores text in Neon — no embeddings yet.
Async Worker	Reads pending chunks → batch embeddings (text-embedding-3-small, 1536d) → pgvector HNSW. Then: generates topic graph (structured output + DFS cycle validation) and evaluation schedule.
Multi-index	Dense (pgvector <=>) + hybrid lexical (tsvector GIN + RRF) for topic-based retrieval. Covers the full syllabus, not just the first 24 k chars.
Retrieval	Over-fetch K=24 candidates → relevance gate (cosine > 0.9 → no context injected) → vector + lexical rerank → top-8 chunks with citations (page / char offset).
Study Agents	Adaptive router (exam weight × domain mastery × schedule urgency) → TS graph orchestrator → specialized agents (flashcard, inquisitor, synth) → different-family verifier → item bank with embedding-based dedup.
Chat Generation	RAG context + agenda block (all courses) + 6-turn history → GROUNDED_SYSTEM_PROMPT (student mentor) → chatStream SSE. Final event: title, citations, provider, model.
Job Queue	Atomic claim (FOR UPDATE SKIP LOCKED), exponential backoff 2^attempts, stuck-job rescue (> 10 min), fire-and-forget on upload + backup cron.

---

🏗 Tech Stack

Layer	Technology
Framework	Next.js 14 (App Router), React 18, TypeScript strict
Database	Neon serverless Postgres + pgvector (embeddings 1536d, HNSW)
Auth	NextAuth v5, bcryptjs, RBAC roles (guest → free → pro → admin)
LLM	OpenAI SDK + OpenRouter (fallback by tier). GPT-4o-mini default.
Graph UI	@xyflow/react (editable: add / rename / delete / connect)
UI	Tailwind CSS 4, shadcn/ui (Radix / base-ui), Lucide icons, Sonner toasts
Validation	Zod (server) + react-hook-form (client)
PDF	unpdf (in-process text extraction, no native deps)
Cache / Rate limit	Upstash Redis (optional; in-memory fallback per instance)
CI	GitHub Actions (typecheck + 197 Vitest tests)
Deploy	Vercel (single app)

---

_{Built with Next.js, OpenAI, pgvector, and many late-night coffee sessions. ☕}