How to Build a RAG Pipeline for Live SaaS Data

Building a Retrieval-Augmented Generation (RAG) pipeline for live SaaS data presents distinct challenges compared to static document sources due to the dynamic nature of transactional data and the complexity of permissions. The architecture involves event-driven ingestion, selective re-embedding, and a hybrid approach that combines indexed retrieval with real-time API reads for fields like deal stages and ticket statuses, which change frequently. The process uses Unified for ingestion and change detection, while chunking, embedding, and retrieval logic are managed separately. Metadata models play a crucial role in ensuring targeted updates and tenant isolation, with fields like `is_latest` aiding in maintaining current data integrity. Permission handling varies by category, with live reads recommended for transactional fields where correctness is critical to prevent incorrect actions by agents. The architecture requires a balance between indexing textual content that changes infrequently and live reading of fields that are crucial for decision-making processes.