I built a modular RAG knowledge assistant to replace slow Notion searches across SOPs, pricing, vendor catalogs and training docs. Pipeline ingests/structures content (tables/media), creates OpenAI embeddings, stores vectors in tuned self-hosted Qdrant and uses a GPT-based LLM to return contextual answers. Added role-based access with separate internal and public chat endpoints, plus real-time analytics and feedback logging for continuous improvement at low infra cost.