Rag Pipeline Academic Systems Diagram

16:9 landscape professional academic systems diagram of a Retrieval-Augmented Generation (RAG) pipeline, 6 clear left-to-right sequential flow stages, clean minimalist tech illustration style, plain white background, high-contrast legible sans-serif text, soft professional color palette of muted blues, grays, and subtle green directional arrow accents, no clutter, crisp high-resolution, peer-reviewed journal-quality figure.
Stage 1 (far left): Rectangular box labeled "User query", with example text inside reading "What are common side effects of Lumafen?", small simple user silhouette icon positioned at the left edge of the box.
Stage 2: Hexagonal node labeled "Embedding encoder (BERT-style)", small caption below reading "dense vector d=768".
Stage 3: Stylized 3D database cylinder icon labeled "Vector store", with text on its surface reading "Index: 1.2M chunks", solid directional arrow from stage 2 to stage 3 labeled "kNN, k=5".
Stage 4: Stack of 5 small document thumbnail icons labeled "Retrieved passages", small caption below reading "top-k chunks + metadata".
Stage 5: Hexagonal hub node labeled "Frozen LLM", two incoming directional arrows: 1) long curved arrow originating from Stage 1 labeled "original query", 2) solid straight arrow from Stage 4 labeled "retrieved context".
Stage 6 (far right): Rectangular box labeled "Grounded answer", with inline text marker inside reading "cite: doc#47", small caption below reading "with source citations".
Additional overlays: Light gray dashed outline wrapping fully around Stages 2 and 3, with label above the outline reading "OFFLINE - built once"; separate light gray dashed outline wrapping fully around Stages 4 and 5, with label above the outline reading "ONLINE - per query".
Top center header: Large bold title reading "Retrieval-Augmented Generation pipeline", smaller italicized subtitle directly below the title reading "Lewis et al., 2020".
All text is crisp, undistorted, and fully legible, all arrows are clearly directional with no overlapping elements.