Retrieval-augmented generation (RAG) pipeline integrating web crawling, document chunking, vector embeddings, and LLM inference for context-aware responses. Data flows from websites through a web crawler into a raw document store, then through chunking and embedding services into a vector database for semantic retrieval. User queries are encoded, reranked, and combined with retrieved context via a prompt builder before passing through safety guardrails to the LLM endpoint. The evaluation pipeline continuously monitors output quality and updates the model registry with performance metrics and embedding model versions. This architecture solves the hallucination problem by grounding LLM responses in retrieved documents while maintaining safety and observability. Fork this diagram on Diagrams.so to customize data sources, embedding models, or LLM providers for your specific use case. Consider adding a feedback loop from user interactions back to the evaluation pipeline for continuous improvement.