import os
import shutil
import gradio as gr
from gradio_client import Client, handle_file
from huggingface_hub import hf_hub_download

# 1. SECRETS & BACKEND LINK
HF_TOKEN = os.environ.get("HF_TOKEN")
PRIVATE_SPACE = "st192011/ASL-VLS-Private" 

# 2. DEFINITIVE SUPPORTED VOCABULARY LIST
SUPPORTED_VIDEOS = [
    ("00944", "ADAPT"), ("00963", "ADD"), ("01064", "ADJECTIVE"), ("00335", "ABDOMEN"),
    ("00689", "ACCOUNTANT"), ("00899", "ACTOR"), ("00584", "ACCENT"), ("00632", "ACCIDENT"),
    ("00586", "ACCENT"), ("00585", "ACCENT"), ("00626", "ACCIDENT"), ("00623", "ACCIDENT"),
    ("00846", "ACT"), ("00890", "ACTIVITY"), ("00898", "ACTOR"), ("01011", "ADDRESS"),
    ("00834", "ACROSS"), ("00624", "ACCIDENT"), ("00593", "ACCEPT"), ("00415", "ABOUT"),
    ("00961", "ADD"), ("00962", "ADD"), ("00594", "ACCEPT"), ("00964", "ADD"), 
    ("00666", "ACCOMPLISH"), ("01065", "ADJECTIVE"), ("00628", "ACCIDENT"), ("00868", "ACTIVE"), 
    ("00836", "ACROSS"), ("00430", "ABOVE"), ("00835", "ACROSS"), ("00946", "ADAPT"), 
    ("00943", "ADAPT"), ("00414", "ABOUT"), ("00376", "ABLE"), ("00832", "ACROSS"), 
    ("00627", "ACCIDENT"), ("00592", "ACCEPT"), ("00625", "ACCIDENT"), ("01012", "ADDRESS"), 
    ("00849", "ACT"), ("00663", "ACCOMPLISH"), ("00853", "ACTION"), ("00967", "ADD"), 
    ("00692", "ACCOUNTANT"), ("00583", "ACCENT"), ("00341", "ACROSS"), ("00378", "ADDRESS"),
    ("00433", "ADJECTIVE"), ("00384", "ACTOR"), ("00381", "ACTOR"), ("00377", "ACCIDENT"),
    ("00382", "ACTOR"), ("00378", "ADDRESS")
]
SUPPORTED_GLOSSES_UNIQUE = sorted(list(set([g for _, g in SUPPORTED_VIDEOS])))
dataset_options = {f"{g} (Sample {vid})": f"data/data_0/{vid.zfill(5)}.mp4" for vid, g in SUPPORTED_VIDEOS}

# 3. INITIALIZE CLIENT
try:
    client = Client(PRIVATE_SPACE, token=HF_TOKEN)
except Exception as e:
    client = None

# 4. LOGIC FUNCTIONS
def update_video_display(selection):
    if not selection: return None, None
    try:
        gloss_gt = selection.split('(')[0].strip()
        hf_path = dataset_options[selection]
        cache_path = hf_hub_download(repo_id="Voxel51/WLASL", filename=hf_path, repo_type="dataset")
        local_path = os.path.join("/tmp", os.path.basename(hf_path))
        shutil.copy(cache_path, local_path)
        return local_path, f"Ground Truth: {gloss_gt}"
    except:
        return None, "Error loading sample"

def run_omnisign_vlm(video_path):
    if not video_path: return {"⚠️ No Input": 0.0}
    if not client: return {"⚠️ Engine Offline": 0.0}

    try:
        # Request from Private Space
        result = client.predict(
            handle_file(video_path),
            api_name="/predict_sign"
        )
        
        # --- PARSING LOGIC ---
        if isinstance(result, dict) and "confidences" in result:
            return {item['label']: item['confidence'] for item in result["confidences"]}
        
        return result

    except Exception as e:
        return {f"❌ Neural Analysis Error": 0.0}

# 5. UI DESIGN (PROFESSIONAL PITCH)
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown(f"""
    # 🧠 OmniSign VLM: Neural Universal SL Protocol
    ### **Powered by Multimodal Temporal Reasoning**

    This demonstration showcases a revolutionary **Structural Protocol** for sign language interpretation, powered by a Large Vision-Language Model (VLM) core. Our protocol focuses on extracting pure **kinetic semantics** from video streams.

    **The OmniSign Protocol's Unique Advantages:**

    1.  **Motion-Oriented Core:** The system is designed to analyze the physics and trajectory of movement, rendering the prediction robust against variations in the signer, lighting, or background environment.
    2.  **Lexical Agnosticism:** The underlying VLM protocol is language-independent. It can be instantly updated to recognize new signs or expanded to include any sign language (e.g., ASL, BSL, LSF) with unparalleled efficiency.
    3.  **Future-Proof Scalability:** New vocabulary can be integrated into the system's lexicon instantly, bypassing traditional, time-consuming retraining cycles.

    ---
    This engine is currently not yet **fully optimized for predictive accuracy** and operates on a limited vocabulary. Its sole purpose is to demonstrate a **highly versatile and scalable VLM protocol**. 
    """)
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### 🎦 1. Input Source")
            video_comp = gr.Video(label="Video Buffer", autoplay=True)
            dataset_drop = gr.Dropdown(choices=[""] + sorted(list(dataset_options.keys())), label="Browse WLASL Archive")
            gt_output = gr.Textbox(label="Ground Truth Reference", interactive=False)
            run_btn = gr.Button("🚀 Execute Neural Analysis", variant="primary")
            
        with gr.Column():
            gr.Markdown("### 📊 2. VLM Perception Result")
            output_label = gr.Label(num_top_classes=3, label="Confidence Score")
            
            with gr.Accordion("🔍 Supported Vocabulary List", open=True):
                gr.Markdown(", ".join(SUPPORTED_GLOSSES_UNIQUE))

    dataset_drop.change(fn=update_video_display, inputs=dataset_drop, outputs=[video_comp, gt_output])
    run_btn.click(fn=run_omnisign_vlm, inputs=video_comp, outputs=output_label)

if __name__ == "__main__":
    demo.launch(ssr_mode=False)