Paper

RAG system for research papers. Feed it PDFs, ask it questions, get answers with source references.

Setup

pnpm install
docker compose up -d  # starts Qdrant

You'll need a .env file:

REPLICATE_API_TOKEN=your_token

Usage

# Extract markdown from PDF
pnpm --filter @paper/main dev pdf paper.pdf

# Ingest into vector store
pnpm --filter @paper/main dev ingest paper.md

# Search chunks
pnpm --filter @paper/main dev search "attention mechanism"

# Ask questions
pnpm --filter @paper/main dev ask "how does the transformer handle long sequences?"

How it works

1. PDF → Markdown via Marker on Replicate
2. Chunk with LangChain's MarkdownTextSplitter
3. Embed with Qwen3 (1024 dims) on Replicate
4. Store in Qdrant
5. Generate answers with Claude

Stack

• TypeScript
• pnpm workspaces
• Qdrant (vector store)
• Replicate (embeddings, PDF extraction)
• Biome (lint/format)
• Zod (validation)