ChatPDF

Students, researchers, and professionals routinely face 50–500 page PDFs — research papers, resumes, contracts, lecture notes — but reading every page to find one fact is slow and error-prone. Pasting the whole document into a generic chatbot blows token limits, hallucinates when context is missing, and cannot scope answers to *your* file. Existing PDF tools often stop at keyword search or a one-shot summary with no follow-up thread, no account persistence, and no guarantee that answers cite document content instead of the open internet.

The product gap is a **document-grounded Gen AI app**: sign in, upload once, get an automatic summary and starter questions, then chat in plain English with retrieval-augmented answers — all sessions saved per user, not lost on refresh.

Solo Full-Stack Engineer, Architect & Product Owner — I designed and shipped ChatPDF end-to-end.

**Frontend:** React 19 SPA (Vite 6, TypeScript, Tailwind CSS v4, Framer Motion) with cinematic video hero landing, glass-style UI, features/how-it-works/FAQ/testimonials marquee, responsive navbar, email + Google auth screens, drag-and-drop PDF upload (up to 15 MB), sidebar session history, chat workspace with summary header, suggested starter questions, replace-PDF flow, and always-dark theme.

**Backend:** Node.js 20+ Express server (TypeScript, npm workspaces) — JWT + bcrypt auth, Google OAuth 2.0 callback, pdf-parse text extraction, chunking + FAISS (faiss-node) vector index, multi-agent RAG pipeline (query optimization → retrieval → context analysis → answer generation), Groq LLM integration with multi-key pool and automatic failover on rate limits, input truncation safeguards, and Neon PostgreSQL persistence.

**Ops:** Split Vercel (frontend) + Render (backend — native faiss-node requires full Node host, not serverless) + Neon DB; render.yaml Blueprint; schema auto-created from backend/db/schema.sql on boot.

ChatPDF is an **npm workspaces monorepo** — frontend/ (React + Vite) and backend/ (Express + RAG).

1$   React 19 + Vite (Vercel)
2$        │  JWT
3$        ▼
4$   Express (Render)   ── native faiss-node needs a full Node host
5$        │
6$        ├─ pdf-parse ──▶ chunks ──▶ FAISS index (in-process)
7$        │
8$        ├─ multi-agent RAG
9$        │    query optimiser ─▶ retrieval ─▶ context analyser ─▶ answer
10$        │                                            │
11$        │                                            ▼
12$        │                              Groq  (multi-key failover)
13$        │
14$        └──▶ Neon PostgreSQL
15$             users · chat_sessions · pdf_documents · chat_messages

**Request flow:** 1. User signs in → JWT issued, stored client-side 2. User uploads PDF → backend extracts text via pdf-parse, chunks content, builds FAISS index, generates AI summary 3. User asks a question → multi-agent RAG retrieves relevant chunks → Groq generates grounded answer 4. Messages + session metadata persisted to Neon PostgreSQL

**Frontend** (frontend/): React 19, TypeScript, Vite 6, Tailwind v4, Framer Motion. Landing: cinematic hero video, glass cards, smooth-scroll sections, CTA to app. App shell: sidebar (+ New Chat, session list, user avatar from Google photo or generated fallback), main chat pane (PDF metadata, word/page count, summary block, message thread, composer). Dev: empty VITE_API_URL — Vite proxies to backend :3002.

**Backend** (backend/): Express REST API. Auth routes: email register/login (bcrypt + JWT), Google OAuth redirect + callback. PDF routes: multipart upload, text extraction, chunking, FAISS index build per session, summary generation. Chat routes: question → RAG pipeline → Groq completion. Groq layer supports GROQ_API_KEY + _2, _3, _4 failover when rate-limited.

**RAG pipeline (multi-agent):** Query optimizer rewrites user question for retrieval → FAISS semantic search over document chunks → context analyzer ranks/filters passages → answer generator produces final response with truncation guards for token limits.

**Database** (Neon PostgreSQL, auto-migrated on start): | Table | Purpose | |-------|--------| | users | Accounts (email + Google) | | chat_sessions | Threads per user | | pdf_documents | PDF metadata + summaries | | chat_messages | Full conversation history |

**Deployment:** Frontend → Vercel (frontend/ root, dist output, VITE_API_URL → Render backend). Backend → Render (npm install && npm run build -w backend, npm run start -w backend). Database → Neon. CORS locked via CLIENT_URL env.

ChatPDF targets the **document Q&A wedge** — students with lecture PDFs, job seekers with long resumes, researchers skimming papers, professionals reviewing contracts. Positioning: *Chat with your PDF* — upload, summarize, ask follow-ups, answers grounded in your file.

**Product-led loop:** landing hero CTA → sign up (Google one-click) → upload PDF → instant summary + suggested questions → natural chat → return via sidebar history.

**Go-to-market:** live at chat-with-pdf-ashen.vercel.app → open-source GitHub → portfolio + LinkedIn demos with resume-PDF chat screenshots (strong recruiter hook).

**Monetization paths (not shipped):** free tier with daily upload/message caps, Pro for larger files + longer retention, team workspaces with shared document libraries.

**Trust signals:** JWT-scoped sessions, per-user PDF isolation, no API keys in client bundle, CORS locked to production frontend URL.

Live at chat-with-pdf-ashen.vercel.app — frontend on Vercel, Express API on Render, Neon PostgreSQL.

**Shipped (product):** Cinematic landing with glass UI · email + Google OAuth auth · drag-and-drop PDF upload (15 MB) · automatic AI summary · suggested starter questions · natural-language Q&A grounded in document · sidebar session history · replace PDF mid-thread · user avatar (Google photo or fallback) · always-dark theme.

**Shipped (platform):** Multi-agent RAG pipeline · FAISS vector search over chunks · Groq LLM with multi-key failover · pdf-parse extraction · JWT + bcrypt auth · 4-table Neon schema (users, chat_sessions, pdf_documents, chat_messages) · npm workspaces monorepo · Vite dev proxy · render.yaml Blueprint · schema auto-init on boot.

Open source at github.com/subhm2004/Chat_with_pdf with full README, architecture diagram, env templates, Google OAuth setup guide, and Vercel + Render deployment checklist.

Add streaming token delivery so long answers feel live. Move PDF file storage to S3 or persistent Render disk — local uploads are lost on backend redeploy. Redis adapter if horizontal scaling is needed. Page-level citations (click chunk → highlight in PDF viewer) would strengthen trust. Rate limiting per user tier on upload and chat endpoints. Integration tests covering upload → FAISS index → RAG → Groq answer chain.