Making RAG Chatbots Enterprise-Ready: Group Permissions and Pluggable Backends
summary
Company-deployed RAG system Maru is open-sourced and already powers legal and ops workflows, using group-based access control mapped to OS-like folders, roles (editor/admin), and a CLI for ingest and sync. Architecture splits RDB (original docs, Tortoise ORM) from vector stores (ChromaDB/Milvus), supports API or local LLMs, and adds a Marync bridge client for physical data isolation. Novel retrieval fixes include a proportional retriever driven by a group classifier and reranker; demo shows bias mitigation when news=1,000 vs other groups=100 and real queries return correct sources.
AI-written summary · about
chapters (13)
- 0:00Enterprise RAG chatbots intro
- 1:49Maru open source RAG engine
- 3:33Maru install and usage demo
- 4:39Maru’s creation journey
- 6:45File system integration for usability
- 8:13RAG pipeline flexibility
- 9:16Data security with MaruSync
- 10:23Corpus imbalance in retrieval
- 13:31Agents and MCP integration
- 14:39Corpus imbalance solutions demo
- 20:36Scalability versus customization
- 22:02Maru architecture and vision
- 22:23Why Maru for enterprise RAG
related talks
59mRAG Against the Machine
01h 36mTutorial: From Planni...
15mRAG Beyond Chatbots
40mRolling out Enterprise AI: Tools, Insights, & Team Empowerment
42mTo RAG or not to RAG
48mUnlocking LLM Potential: From Exploration to Integration
11mBuilding easily deployable RAG applications with Llamastack
31mAdvanced Enterprise RAG Systems