Scientific RAG pipeline architecture combines offline knowledge base creation with online query processing to deliver contextually accurate responses from domain-specific sources. The offline pipeline ingests scientific sources through cleaning, chunking, and embedding stages, storing vectorized content in a vector database while fine-tuning a specialized LLM on scientific training data. The online system routes user questions through a persona manager, retrieves relevant context via RAG, and synthesizes responses using the fine-tuned model with conversation history and system prompts. This architecture solves the challenge of grounding large language models in authoritative scientific knowledge while maintaining conversational context and persona consistency. Fork this diagram on Diagrams.so to customize data sources, embedding models, or fine-tuning strategies for your domain. Consider adding monitoring and feedback loops between response quality and knowledge base updates for continuous improvement.