architecture_draft.mmd

graph TD
    %% Styling
    classDef default fill:#1A1A1A,stroke:#444,stroke-width:2px,color:#FFF,rx:8px,ry:8px,font-family:arial;
    classDef mobile fill:#003366,stroke:#0055AA,stroke-width:2px,color:#FFF;
    classDef preproc fill:#333333,stroke:#555,stroke-width:2px,color:#FFF;
    classDef model fill:#4B0082,stroke:#8A2BE2,stroke-width:2px,color:#FFF;
    classDef condition fill:#2B2B2B,stroke:#F4A460,stroke-width:2px,color:#FFF,shape:rhombus;
    classDef external fill:#004d00,stroke:#009900,stroke-width:2px,color:#FFF;
    classDef final fill:#660000,stroke:#CC0000,stroke-width:2px,color:#FFF;

    %% Nodes
    UserApp[📱 Mobile App]:::mobile
    
    ImgUpload[🖼️ Image]:::preproc
    Question[⌨️ Question Text]:::preproc
    
    PIL[🐍 PIL Preprocessing<br/>RGB conversion]:::preproc
    
    CLIP[👁️ OpenAI CLIP ViT-B/32<br/>Image Features 512-dim]:::model
    GPT2[🤗 DistilGPT-2<br/>Tokenized Question]:::model
    
    Route1{Question<br/>spatial?}:::condition
    
    Spatial[📐 Spatial VQA Model<br/>8-head attention]:::model
    Base[🧠 Base VQA Model<br/>General VQA]:::model
    
    Decoder[🤗 GPT-2 Decoder<br/>vocab decode]:::model
    NeuralAns[💬 Neural Answer]:::final
    
    Route2{Knowledge<br/>question?}:::condition
    
    ObjDet[👁️ CLIP Object Detector<br/>Top-3 objects]:::model
    Wikidata[🌍 Wikidata SPARQL<br/>P31, P186, P366]:::external
    GroqV[⚡ Groq Llama-3.3<br/>Verbalizer]:::external
    KGAns[🧩 KG Enhancement]:::final
    
    FastAPI[🚀 FastAPI]:::preproc
    GroqA[⚡ Groq Llama-3.3<br/>Accessibility]:::external
    Audio[🔊 2-sentence description]:::final

    %% Edges
    UserApp -- "Image uploaded" --> ImgUpload
    UserApp -- "Question typed" --> Question
    
    ImgUpload --> PIL
    PIL --> CLIP
    Question --> GPT2
    
    CLIP & GPT2 --> Route1
    
    Route1 -- "YES" --> Spatial
    Route1 -- "NO" --> Base
    
    Spatial & Base -- "Beam search (width=5)" --> Decoder
    Decoder --> NeuralAns
    
    CLIP -- "Anchor similarity" --> Route2
    
    Route2 -- "YES" --> ObjDet
    ObjDet -- "Detected objects" --> Wikidata
    Wikidata -- "Structured facts" --> GroqV
    GroqV --> KGAns
    
    FastAPI -- "Narration request" --> GroqA
    GroqA --> Audio
    
    NeuralAns & KGAns & Audio -- "JSON output" --> FastAPI
    FastAPI --> UserApp