import sys import matplotlib.pyplot as plt from matplotlib.patches import FancyBboxPatch, Rectangle def draw_rag_pipeline(output_path): fig, ax = plt.subplots(figsize=(12, 9), dpi=100) ax.set_xlim(0, 1200) ax.set_ylim(0, 900) ax.set_aspect('equal') ax.axis('off') fig.patch.set_facecolor('#0b0f19') ax.set_facecolor('#0b0f19') # Title & Subtitle ax.text(600, 860, 'Retrieval-Augmented Generation (RAG) Architecture', color='#f8fafc', fontsize=20, fontweight='bold', ha='center', va='center') ax.text(600, 825, 'End-to-End Pipeline: Offline Ingestion & Online Real-Time Synthesis', color='#94a3b8', fontsize=13, ha='center', va='center') # 1. OFFLINE INGESTION CONTAINER (Top) ax.add_patch(FancyBboxPatch((60, 560), 1080, 230, boxstyle='round,pad=0.02,rounding_size=14', facecolor='#1e293b', edgecolor='#fbbf24', lw=2, ls='--', alpha=0.6)) ax.text(210, 790, 'OFFLINE INGESTION & INDEXING', color='#fef3c7', fontsize=12, fontweight='bold', ha='center', bbox=dict(boxstyle='round,pad=0.3', facecolor='#854d0e', edgecolor='#fbbf24', lw=1.5)) # Ingestion Blocks # Knowledge Base ax.add_patch(FancyBboxPatch((90, 600), 170, 150, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#451a03', edgecolor='#fbbf24', lw=1.5)) ax.text(175, 715, 'Knowledge Base', color='#fef3c7', fontsize=14, fontweight='bold', ha='center') ax.text(175, 650, 'PDFs, Manuals,\nWikis, Docs, DBs', color='#d1d5db', fontsize=11, ha='center') # Chunker ax.annotate('', xy=(290, 675), xytext=(260, 675), arrowprops=dict(arrowstyle='-|>', color='#fbbf24', lw=2.5, mutation_scale=15)) ax.add_patch(FancyBboxPatch((290, 600), 180, 150, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#451a03', edgecolor='#fbbf24', lw=1.5)) ax.text(380, 715, 'Document Chunker', color='#fef3c7', fontsize=14, fontweight='bold', ha='center') ax.text(380, 655, 'Recursive Splitter\n512 tok / overlap 50', color='#fef08a', fontsize=11, ha='center') # Embedding Model (Offline) ax.annotate('', xy=(500, 675), xytext=(470, 675), arrowprops=dict(arrowstyle='-|>', color='#fbbf24', lw=2.5, mutation_scale=15)) ax.add_patch(FancyBboxPatch((500, 600), 180, 150, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#451a03', edgecolor='#fbbf24', lw=1.5)) ax.text(590, 715, 'Embedding Model', color='#fef3c7', fontsize=14, fontweight='bold', ha='center') ax.text(590, 655, 'Dense Vectorizer\nDim = 1536 / 3072', color='#fef08a', fontsize=11, ha='center') # Vector DB ax.annotate('', xy=(720, 675), xytext=(680, 675), arrowprops=dict(arrowstyle='-|>', color='#fbbf24', lw=2.5, mutation_scale=15)) ax.add_patch(FancyBboxPatch((720, 580), 380, 180, boxstyle='round,pad=0.02,rounding_size=10', facecolor='#064e3b', edgecolor='#34d399', lw=2)) ax.text(910, 725, 'Vector Database', color='#ecfdf5', fontsize=16, fontweight='bold', ha='center') ax.text(910, 680, 'HNSW / IVF Index (Cosine / L2)', color='#a7f3d0', fontsize=12, ha='center') ax.text(910, 630, 'Embeddings + Metadata Index', color='#6ee7b7', fontsize=11, ha='center', bbox=dict(boxstyle='round,pad=0.3', facecolor='#022c22', edgecolor='#34d399', lw=1)) # 2. ONLINE INFERENCE CONTAINER (Bottom) ax.add_patch(FancyBboxPatch((60, 140), 1080, 390, boxstyle='round,pad=0.02,rounding_size=14', facecolor='#1e293b', edgecolor='#38bdf8', lw=2, ls='--', alpha=0.6)) ax.text(230, 530, 'ONLINE REAL-TIME INFERENCE PIPELINE', color='#e0f2fe', fontsize=12, fontweight='bold', ha='center', bbox=dict(boxstyle='round,pad=0.3', facecolor='#0369a1', edgecolor='#38bdf8', lw=1.5)) # User Query ax.add_patch(FancyBboxPatch((90, 370), 170, 130, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#172554', edgecolor='#38bdf8', lw=2)) ax.text(175, 465, 'User Query', color='#7dd3fc', fontsize=14, fontweight='bold', ha='center') ax.text(175, 415, '"How does RAG\nwork?"', color='#f8fafc', fontsize=11, style='italic', ha='center') # Query Embedder ax.annotate('', xy=(290, 435), xytext=(260, 435), arrowprops=dict(arrowstyle='-|>', color='#38bdf8', lw=3, mutation_scale=16)) ax.add_patch(FancyBboxPatch((290, 370), 180, 130, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#172554', edgecolor='#38bdf8', lw=2)) ax.text(380, 465, 'Query Embedder', color='#7dd3fc', fontsize=14, fontweight='bold', ha='center') ax.text(380, 415, r'$q_{vec} = \mathrm{embed}(query)$' + '\nSimilarity Ready', color='#38bdf8', fontsize=11, ha='center') # Query -> Vector DB search arrow ax.plot([470, 550, 550, 720], [435, 435, 590, 590], color='#38bdf8', lw=2.5) ax.annotate('', xy=(720, 590), xytext=(650, 590), arrowprops=dict(arrowstyle='-|>', color='#38bdf8', lw=2.5, mutation_scale=15)) ax.text(600, 565, 'Top-k ANN Search', color='#38bdf8', fontsize=11, fontweight='bold', ha='center') # Vector DB -> Context return arrow ax.plot([850, 850, 680], [580, 350, 350], color='#34d399', lw=2.5) ax.annotate('', xy=(680, 350), xytext=(750, 350), arrowprops=dict(arrowstyle='-|>', color='#34d399', lw=2.5, mutation_scale=15)) ax.text(780, 365, 'Top-k Chunks ($k=5$)', color='#34d399', fontsize=11, fontweight='bold', ha='center') # Retrieved Context Box ax.add_patch(FancyBboxPatch((500, 240), 180, 170, boxstyle='round,pad=0.02,rounding_size=8', facecolor='#064e3b', edgecolor='#34d399', lw=2)) ax.text(590, 380, 'Retrieved Context', color='#a7f3d0', fontsize=13, fontweight='bold', ha='center') ax.text(590, 315, '• Snippet #1 (score: 0.93)\n• Snippet #2 (score: 0.89)\n• Re-ranked Top-k', color='#6ee7b7', fontsize=10, ha='center') # Direct query pass ax.plot([175, 175, 490], [370, 270, 270], color='#38bdf8', ls='--', lw=2) ax.annotate('', xy=(490, 270), xytext=(400, 270), arrowprops=dict(arrowstyle='-|>', color='#38bdf8', lw=2, mutation_scale=14)) ax.text(320, 285, 'Direct Query Augment', color='#7dd3fc', fontsize=10) # Context + Query -> LLM ax.annotate('', xy=(740, 280), xytext=(680, 280), arrowprops=dict(arrowstyle='-|>', color='#c084fc', lw=3, mutation_scale=16)) # LLM Generator Box ax.add_patch(FancyBboxPatch((740, 180), 360, 200, boxstyle='round,pad=0.02,rounding_size=10', facecolor='#4c1d95', edgecolor='#c084fc', lw=2)) ax.text(920, 350, 'LLM Generator', color='#f3e8ff', fontsize=16, fontweight='bold', ha='center') llm_prompt = "PROMPT TEMPLATE:\nSystem: Answer using retrieved context.\nContext: [Retrieved Passages]\nUser: [Original Query]" ax.text(920, 260, llm_prompt, color='#e9d5ff', fontsize=10, ha='center', bbox=dict(boxstyle='round,pad=0.4', facecolor='#2e1065', edgecolor='#c084fc', lw=1)) # 3. FINAL ANSWER CARD (Bottom) ax.add_patch(FancyBboxPatch((60, 25), 1080, 85, boxstyle='round,pad=0.02,rounding_size=10', facecolor='#1e293b', edgecolor='#34d399', lw=2)) ax.text(80, 75, 'Synthesized Grounded Answer:', color='#34d399', fontsize=13, fontweight='bold', ha='left') ax.text(80, 48, '"RAG retrieves relevant domain documents and augments the LLM prompt to deliver accurate, cited answers."', color='#f8fafc', fontsize=11, ha='left') # LLM to Answer Arrow ax.annotate('', xy=(920, 110), xytext=(920, 180), arrowprops=dict(arrowstyle='-|>', color='#34d399', lw=3, mutation_scale=16)) plt.subplots_adjust(left=0, right=1, top=1, bottom=0) plt.savefig(output_path, dpi=100) plt.close(fig) if __name__ == '__main__': if len(sys.argv) > 1: draw_rag_pipeline(sys.argv[1]) else: draw_rag_pipeline('rag-pipeline.png')