Arxiver

Your AI-powered research companion for exploring, saving, and understanding arXiv papers.

Arxiver helps researchers and engineers discover relevant papers, get instant AI-generated summaries, chat with papers using RAG, and receive personalized recommendations — all with actionable insights on how papers apply to your projects and company.

NOTE: Switch to shadcn-ui-migrate for latest changes.

Demo

Features

Smart Paper Discovery Feed

A personalized paper feed powered by vector similarity search. When you save papers, Arxiver learns your interests and uses your last saved papers as seed vectors for k-NN search across the arXiv catalog. Select topic categories to filter your feed. Falls back to a chronological catalog feed for new users.

Full-Text Paper Reader

Read papers directly in the app with an embedded PDF viewer. Papers are streamed through a proxy to avoid CORS issues. The detail view shows the abstract, extracted keywords, primary category, and links to the original arXiv page.

Save and Organize Papers

Build a personal library of saved papers. When you save a paper, Arxiver kicks off background processing — generating a thumbnail, extracting keywords using KeyBERT with a SciBERT model, and indexing the paper for future recommendations. Bulk delete and manage your library from a dedicated page.

AI Paper Summarization

Generate concise summaries of any saved paper. The summarization engine retrieves all ingested chunks of the paper, sorts them by page order, and produces a structured summary using an LLM. Summaries are cached per paper so they load instantly on repeat visits.

AI Chat with Papers (RAG)

Have a conversation with any paper. Once a paper is ingested, its content is chunked, embedded with Cohere multilingual embeddings, and stored in a Qdrant vector collection. When you ask a question, the system retrieves relevant chunks, optionally reranks them, and streams a grounded answer with source citations. Each paper gets its own chat sessions with full message history.

Web-Augmented Chat

Toggle web search in chat to let the AI pull in external context. Uses Tavily or LangSearch for web search and Firecrawl for deep page scraping, combining web results with paper chunks for more comprehensive answers.

AI Usability Analysis

Get a structured assessment of how a paper can be applied in practice. The usability engine analyzes each paper across 13 business domains (healthcare, finance, manufacturing, etc.), 2 reproducibility axes, 14 emerging technology areas, and produces an overall impact score — all returned as structured data and rendered as visual charts.

Hybrid Recommendation System

Every saved paper and catalog page shows related paper recommendations. The engine runs three parallel Qdrant queries:

• Semantic similarity — finds papers with similar title and abstract embeddings
• Topic-constrained similarity — same as above but filtered to the paper's primary arXiv category
• Author-based discovery — finds other papers by the same authors

Results are blended using a weighted re-ranking formula (60% similarity, 15% category affinity, 15% recency, 10% baseline) and deduplicated.

Keyword Extraction

Automatic keyword extraction for every saved paper using KeyBERT with a quantized SciBERT ONNX model. Includes a curated seed list of 180+ CS/AI terms, technical keyword boosting, MMR diversity, and embedding-based deduplication. Keywords power discovery and are displayed on paper detail pages.

ArXiv Catalog with Semantic Search

A daily-updated local mirror of arXiv metadata harvested via the OAI-PMH protocol. Papers are indexed into Qdrant with server-side embeddings, enabling semantic search across the catalog — search by meaning, not just keywords.

Bring Your Own API Keys

No server-side LLM keys required. Users provide their own API keys for Cohere, Groq, Gemini, OpenRouter, Tavily, and Firecrawl through the settings page. Keys are encrypted at rest using Fernet symmetric encryption. Supports multiple LLM providers (Gemini, Groq, Cohere, OpenRouter) with a provider/model format.

Custom Instructions

Set custom instructions that guide how the AI summarizes papers and responds in chat, tailoring outputs to your domain and use case.

Authentication

Email/password registration with bcrypt hashing and Google OAuth2 sign-in. JWT-based sessions with access/refresh token pairs. Sessions are tracked server-side for immediate revocation.

---

Tech Stack

Layer	Technologies
Frontend	React 18, TypeScript, Vite, Tailwind CSS v4, shadcn/ui, TanStack Query
Backend	FastAPI, Python 3.13, SQLAlchemy 2.0 (async), Alembic
AI/ML	LangChain, LangGraph, Cohere Embeddings, KeyBERT, SciBERT ONNX
Vector DB	Qdrant Cloud
Database	PostgreSQL (Supabase)
Storage	Supabase S3 (thumbnails)
Observability	Logfire + OpenTelemetry

---

Project Setup

Prerequisites

• Python 3.13+ with uv package manager
• Bun runtime (for frontend)
• PostgreSQL database (e.g., Supabase)
• Qdrant Cloud instance

Backend

cd backend

# Install dependencies
uv sync

# Set up environment variables
# Create a .env file with the required variables (see Environment Variables below)

# Run database migrations
uv run alembic upgrade head

# Start the server (default port 8000)
uv run python run.py

Frontend

cd frontend

# Install dependencies
bun install

# Start dev server (default port 8080)
bun run dev

Environment Variables

Backend — create backend/.env:

Variable	Description
DATABASE_URL	PostgreSQL async connection string
JWT_SECRET_KEY	Secret for signing JWT tokens
ENCRYPTION_KEY	Fernet key for encrypting user API keys
GOOGLE_CLIENT_ID	Google OAuth client ID
GOOGLE_CLIENT_SECRET	Google OAuth client secret
GOOGLE_REDIRECT_URI	Google OAuth callback URL
QDRANT_URI	Qdrant Cloud cluster URL
QDRANT_API_KEY	Qdrant Cloud API key
S3_ENDPOINT_URL	Supabase S3 endpoint
S3_ACCESS_KEY_ID	Supabase S3 access key
S3_SECRET_ACCESS_KEY	Supabase S3 secret key
S3_BUCKET_NAME	S3 bucket for thumbnails
LOGFIRE_TOKEN	Logfire observability token

Frontend — set in shell or .env:

Variable	Description
VITE_API_URL	Backend API base URL (defaults to http://localhost:8000/api/v1)

---

Project Structure

ArxiverApp/
├── backend/
│   ├── main.py                 # FastAPI app entry point
│   ├── run.py                  # Uvicorn launcher
│   ├── alembic/                # Database migrations
│   ├── src/
│   │   ├── router/             # API route handlers
│   │   ├── controller/         # Business logic
│   │   ├── model/              # SQLAlchemy ORM models
│   │   ├── schema/             # Pydantic request/response schemas
│   │   ├── core/               # AI engines (chat, summary, recommendations, keywords, catalog)
│   │   ├── lib/                # Auth, Qdrant client, middleware
│   │   ├── database/           # Async DB engine and sessions
│   │   └── config/             # LLM system prompts
│   └── pyproject.toml
├── frontend/
│   ├── src/
│   │   ├── pages/              # Page components
│   │   ├── components/         # Reusable UI components
│   │   ├── context/            # React contexts (Auth, UserData, Search, Theme)
│   │   ├── lib/                # API client, utilities
│   │   ├── hooks/              # Custom React hooks
│   │   └── types/              # TypeScript type definitions
│   └── package.json
├── start-fastapi.sh            # Backend launch script
└── start-webapp.sh             # Frontend launch script