PaperSight AI

An AI document tool where you can actually see the retrieval happening.

Upload a PDF, get a structured summary, then chat with the document — with every AI answer showing exactly which passages it pulled from, how relevant they were, and why. Not another Gemini wrapper with a file input. The RAG pipeline is the product, and it's visible.

---

What's Inside

• RAG you can see — every chat response shows expandable source citation cards with the exact chunks retrieved via pgvector cosine similarity, scored by relevance. The chat header shows a persistent "RAG · Grounded" badge; while processing, a three-stage animation telegraphs the pipeline: Retrieving context → Ranking passages → Generating response.

• Client-side PDF parsing — pdfjs-dist extracts text in the browser before anything hits the server. Batch upload up to 5 PDFs, combine them into one summary. Scanned-image PDFs get caught early with a clear error, not after a 30-second upload.

• Real streaming, not fake — summaries and chat responses stream token-by-token via SSE. The upload flow has six distinct stages (parse → upload → summarize → save → index → redirect), each with its own progress state and toast. Abort at any point.

• Guest access with teeth — no sign-up wall. Guests get 10 summaries/day via a cookie-based rate limiter with server-side validation. Summary viewing works, chat requires auth (needs user-scoped embeddings).

• Five summary styles — same document, different output. Viral, academic, executive, bullet-point, and detailed. Backed by style-specific Gemini system prompts, not just a temperature slider.

---

Tech Stack

Layer	Choice	Why
Framework	Next.js 16.1.6, React 19, TypeScript	App Router, RSCs for the summary page, Turbopack dev
AI	Gemini (@google/genai)	gemini-3-flash-preview for generation, gemini-embedding-001 (768d) for embeddings
Vector DB	Neon Postgres + pgvector	<=> cosine similarity operator, raw SQL via Prisma for vector columns
ORM	Prisma 7 + @prisma/adapter-neon	Driver adapter for serverless Neon, Unsupported("vector(768)") for the embedding column
Auth	Clerk	Middleware-based, guest fallback via UUID cookie
File storage	UploadThing	PDF storage for signed-in users; guests skip storage, still get summaries
PDF parsing	pdfjs-dist (client), pdf-parse (server)	Client-side extraction avoids uploading raw PDFs for text-only summarization
Data fetching	SWR	Stale-while-revalidate for summaries list and chat history
Styling	Tailwind 4 + custom design system	Editorial Tech aesthetic — Instrument Serif headings, Geist Mono labels, oklch palette
Animations	Framer Motion 12	Reduced-motion aware, editorial easing (no spring/bounce), <400ms max
UI primitives	Radix UI + shadcn/ui components	Accessible by default

---

Getting Started

git clone <repo-url>
cd papersight_ai
npm install

Create .env.local:

# Required
DATABASE_URL=            # Neon Postgres connection string (must have pgvector enabled)
GEMINI_API_KEY=          # Google AI Studio key
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY=
CLERK_SECRET_KEY=
UPLOADTHING_TOKEN=       # From uploadthing.com dashboard

# Optional
NEXT_PUBLIC_CLERK_SIGN_IN_URL=/sign-in
NEXT_PUBLIC_CLERK_SIGN_UP_URL=/sign-up

Then:

npx prisma generate   # Generate Prisma client
npx prisma db push    # Push schema to Neon
npm run dev

Gotcha: The pdf_embeddings table uses a vector(768) column. Make sure pgvector is enabled on your Neon project (CREATE EXTENSION IF NOT EXISTS vector;).

---

Project Structure

app/
  page.tsx                    # Landing — hero, features, upload CTA
  upload/page.tsx             # Multi-file upload with streaming preview
  summary/[id]/               # Summary view + PDF split panel + chat
  dashboard/                  # User's saved summaries
  api/
    chat/route.ts             # RAG chat — embed query, search chunks, stream response
    summarize-text/route.ts   # Streaming Gemini summarization via SSE
    save-summary/route.ts     # Persist summary + trigger background indexing
    uploadthing/              # File upload handler

components/
  summary/
    chat-panel.tsx            # Floating chat with source citations, RAG badge, streaming cursor
    pdf-viewer.tsx            # Resizable split-panel (summary + PDF iframe)
    summary-list.tsx          # Dashboard grid with SWR
  hero.tsx                    # Landing hero with live summary preview animation
  features.tsx                # Feature grid — Chat + RAG dominate top row

lib/
  embeddings.ts               # Chunking (500 words, 50 overlap), embed, store, vector search
  gemini.ts                   # Gemini client config
  pdf-parser-client.ts        # Browser-side pdfjs-dist extraction with batch support
  guest-rate-limit.ts         # Server-side cookie-based rate limiting

---

Design System

Editorial Tech — the visual language sits between a high-end digital publication and a developer tool.

• Typography: Instrument Serif (400 italic) for headings — dramatic, authoritative. Geist for body. Geist Mono for labels, badges, and metadata — always uppercase, tracked wide.
• Color: oklch throughout. Deep navy-charcoal background (0.13 0.01 260). Warm amber accent (0.78 0.16 55) — used only for CTAs, active states, RAG indicators, and links. One accent color, no rainbow.
• Surfaces: surface-raised (card bg + 1px border + layered shadows) and surface-sunken (inputs, recessed areas). No glassmorphism, no heavy shadows.
• Animation: Max 400ms. ease-out or cubic-bezier(0.4, 0, 0.2, 1). No spring, no bounce. Entry = fade + translateY(8px). Exit = fast fade only.

Full spec in DESIGN_SYSTEM.md.

---

Screenshots

Screenshots should be added here: landing page, upload flow with streaming preview, summary view with PDF split panel, chat with source citations visible.

---

What I'd Do Differently

• Embeddings should be generated at upload time, not on first chat. Right now, if a user opens chat on a document that hasn't been indexed yet, there's a noticeable cold-start delay while chunks are embedded. A background job triggered right after save would fix this — the fire-and-forget fetch in use-pdf-upload.ts is a workaround, not a solution.

• The vector search needs a similarity threshold, not just top-K. Currently it returns the 5 most similar chunks regardless of score. If a user asks something completely unrelated to the document, it still retrieves low-relevance chunks and the AI tries to answer from them. A WHERE similarity > 0.3 filter would fix the hallucination edge case.

• The chat should be a full-page layout on desktop, not a floating widget. The 420px floating panel works but wastes screen real estate. A proper three-column layout (nav | summary+PDF | chat) on the summary page would make this feel like an actual research tool. The floating panel is still fine for mobile.

---

Author

Fazlul Karim