Learner-focused course · arXiv Paper Curator
Search first.
Vectors second.
Agents last.
Build a production RAG system one layer a week — a research assistant that fetches arXiv papers, understands them, and answers questions. Unlike tutorials that jump straight to vector search, the course masters keyword search foundations first, then enhances with vectors for hybrid retrieval.
“Solid search foundations enhanced with AI, not AI-first approaches that ignore search fundamentals.”— the README’s own case for the order
Seven layers, laid one a week
newest on topweek7.0Agentic RAG & Telegram botLangGraph, document grading, query rewriting, guardrails
week6.0Monitoring & cachingLangfuse tracing, Redis caching
week5.0Complete RAG pipelineLocal LLM, streaming responses, Gradio interface
week4.0Chunking & hybrid searchKeywords combined with semantic understanding
week3.0BM25 keyword searchFiltering and relevance scoring — the layer tutorials skip
week2.0Data ingestionFetching and parsing arXiv papers automatically
week1.0InfrastructureDocker, FastAPI, PostgreSQL, OpenSearch, Airflow
Order and contents are the course’s own weekly plan. Every band is a tagged release you can check out on its own.
Every layer is a checkout
Each week ships as a git tag, so you can stand on any layer without the ones above it.
# replace with week1.0 … week7.0
git clone --branch week3.0 https://github.com/jamwithai/arxiv-paper-curator
uv sync
docker compose up --build -d
Docker DesktopPython 3.12+uv8 GB+ RAM20 GB+ disk
What the top layer adds
Decision-makingAgents evaluate and adapt retrieval strategies.
Document gradingAutomatic relevance assessment.
Query rewritingRefines the query when results are insufficient.
GuardrailsOut-of-domain detection to prevent hallucination.
Mobile accessA Telegram bot for any device.
TransparencyFull reasoning-step tracking for debugging.