What if you could build production-ready RAG entirely in Elixir?
summary
Claim of a full RAG stack on the BEAM using Arcana becomes concrete: embeddings run locally with Nx and Bumblebee, vectors stored in pgvector, and a Phoenix LiveView dashboard shows ~240 document chunks and retrieval traces. Pipelines include query rewrite, graph-assisted search, cross-encoder reranking and an agentic loop, with reranking boosting hit@1 by about 60%. Grounding uses a local Hallmark evaluator, an evaluation framework tracks retrieval and answer quality, and the live demo successfully answered Doctor Who queries.
AI-written summary · about
chapters (13)
- 0:00Building RAG in Elixir
- 4:52RAG necessity for LLMs
- 8:09Embeddings and vector search
- 12:52Hybrid search RRF
- 15:26Knowledge graphs in RAG
- 19:24Rerankers in retrieval
- 21:59Arcana pipelines
- 27:43Looping in RAG
- 28:48Agentic RAG with Arcana Loop
- 33:46RAG evaluation grounding
- 36:15RAG performance over time
- 37:19Elixir’s AI strengths
- 38:01Q&A
related talks
02h 47mFrom naive to advanced RAG: the complete guide
46mBlow past benchmarks using vector search and RAG with Elasticsearch
53mBuild RAG from Scratch
42mTo RAG or not to RAG
56mRAG: from dumb implementation to serious results
45mMaking semantic search & RAG real: How to build a prod-ready app
32mBuilding Performant RAG Applications for Production
41mFrom RAGs to riches: Build an AI document interrogation app in 30 mins