Arcana: Production-Ready RAG Entirely in Elixir

In this talk, I’ll introduce Arcana (https://github.com/georgeguimaraes/arcana), an open-source library that brings RAG to the BEAM. Using Nx and Bumblebee, Arcana runs embedding models locally, no external APIs needed. Combined with pgvector for vector storage and Phoenix LiveView for the dashboard, you get a complete RAG solution that feels native to Elixir.

We’ll cover:

• How RAG works and why it matters for AI applications
• Running embedding models locally with Nx and Bumblebee
• Efficient document chunking strategies
• Vector similarity search with pgvector
• Building agentic RAG pipelines: query expansion, re-ranking, and self-correcting answers
• Pluggable components via behaviour: swap any pipeline step with your own implementation
• Building an evaluation framework to measure retrieval quality
• You’ll leave with practical knowledge to add semantic search and agentic RAG capabilities to your Phoenix applications, keeping your AI stack simple, fast, and 100% Elixir.