Details (original message): Add the vector RAG pipeline: paragraph chunker, OpenAI-compatible embeddings and rerank clients, dual-tier vector store with cosine search, and RRF hybrid retrieval
