How to Build an Agentic RAG with RubyLLM and Rails (www.panasiti.me)

A new approach to building an agentic Retrieval-Augmented Generation (RAG) application utilizing RubyLLM and Rails has been introduced, enhancing the user experience for Italian pension and tax consultants. This updated pipeline empowers the model to dynamically drive its own retrieval process, allowing it to refine searches, follow document cross-references, and iterate until it gathers sufficient information to answer user queries. This improvement addresses the limitations of the traditional single-shot RAG pipeline, where retrieval occurred only once before the model generated output. Significantly, this agentic approach transforms how information is processed and retrieved, ultimately leading to more accurate and contextually relevant responses. The technical framework includes Rails 8.1 and Ruby 3.3, alongside PostgreSQL with pgvector for vector searches and full-text search capabilities. Key features of the agent include a hybrid search mechanism that utilizes both vector similarity and full-text search, combined with a ranking method called Reciprocal Rank Fusion (RRF) to enhance the diversity of the results. By enabling multiple tool calls through RubyLLM, the system can adaptively query the knowledge base, markedly increasing the reliability and comprehensiveness of the responses generated by the model.