Build a production-ready RAG system in seven weeks, from arXiv ingestion to a Telegram bot
This course guides you through building an AI research assistant that automatically collects the latest papers from arXiv and answers user questions by retrieving relevant content. Unlike tutorials that focus solely on vector search, it follows the industry-standard approach of establishing a foundation in keyword search before combining it with semantic search. By following the weekly code and notes, you will complete an intelligent RAG system connected to a Telegram bot.
A learning path starting with search fundamentals
The course goes beyond simply calling AI models. In week one, you set up infrastructure using Docker, FastAPI, PostgreSQL, OpenSearch, and Airflow. In week two, you build a data pipeline that fetches papers via the arXiv API and parses PDFs. In week three, you implement keyword search using the BM25 algorithm and understand why it is a core foundation for RAG systems. In week four, you introduce document chunking strategies and hybrid search to improve retrieval accuracy.
Complete RAG pipeline and operational monitoring
In week five, you learn to integrate a local LLM using Ollama to convert search results into natural language answers, including techniques for prompt optimization that improve response speed by up to six times. In week six, you learn how to monitor system performance and costs by applying tracing with Langfuse and Redis caching. You will understand a structure that can expect a 150 to 400 times speed improvement through caching.
Agent-based RAG and mobile accessibility
In the final week seven, you use LangGraph to construct an agent workflow. The agent evaluates its own search strategy and performs adaptive search, such as rewriting queries or classifying documents by grade if results are insufficient. It also implements guardrail features to detect out-of-domain questions and prevent hallucination. The completed system connects to a Telegram bot, allowing interactive AI use on mobile devices.
Prerequisites and costs
To follow this course, you need Docker Desktop, Python 3.12 or higher, and the UV package manager. Hardware requirements include at least 8GB of RAM and 20GB of free disk space. You must set up a free Jina AI API key and a Langfuse key, while a Telegram bot token is required starting from week seven. Since all services run locally, the base cost is 0 won, though using external LLM services optionally may incur costs of approximately 2 to 5 dollars. The project is released under the MIT license, allowing free learning and utilization.