app.py

# app.py
import gradio as gr
from rag import RAGPipeline
from attack import AdversarialAttackPipeline
import time

# --- 1. Load Models and Pipelines (Singleton Pattern) ---
# This section runs only ONCE when the application starts up.
# This is the key to low latency for subsequent user requests.
print("Starting application setup...")
start_time = time.time()

# Instantiate the RAG pipeline, which will now load from the pre-computed cache
rag_pipeline = RAGPipeline(
    json_path="calebdata.json",
    defense_model_path="./defense_model",
    cache_dir="cache"
)

# Instantiate the Attack pipeline
attack_pipeline = AdversarialAttackPipeline(rag_pipeline_instance=rag_pipeline)

end_time = time.time()
print(f"Application setup complete. Model loading took {end_time - start_time:.2f} seconds.")

# --- 2. Define the Core Function for Gradio ---
# This function will be called every time a user clicks "Submit"
def run_adversarial_test(query, attack_method, perturbation_level):
    """
    Runs a single attack and returns the results formatted for the Gradio UI.
    """
    print(f"Received query: '{query}', Method: {attack_method}, Level: {perturbation_level}")

    # The attack_pipeline object is already loaded in memory
    result = attack_pipeline.run_attack(
        original_query=query,
        perturbation_method=attack_method,
        perturbation_level=perturbation_level
    )

    # Format the output for Gradio components
    return (
        result["normal_query"],
        result["perturbed_query"],
        result["normal_response"],
        result["perturbed_response"],
        f"Defense Triggered: {result['defense_triggered']} | Reason: {result['reason']}",
        f"Response Similarity: {result['cos_sim']['response_sim']}%",
        f"Adversarial Risk Index (ARI): {result['ari']}"
    )

# --- 3. Build the Gradio Interface ---
with gr.Blocks(theme=gr.themes.Soft(), title="RAG Adversarial Tester") as demo:
    gr.Markdown("# Adversarial Robustness Testing for RAG Systems")
    gr.Markdown(
        "This interface allows you to test the resilience of a RAG pipeline against various adversarial query perturbations. "
        "Enter a query, select an attack method and perturbation level, and observe the impact."
    )

    with gr.Row():
        with gr.Column(scale=1):
            query_input = gr.Textbox(label="Original Query", placeholder="e.g., Who is the Vice-Chancellor?")
            attack_method_input = gr.Dropdown(
                label="Attack Method",
                choices=["random_deletion", "synonym_replacement", "contextual_word_embedding"],
                value="random_deletion"
            )
            perturbation_level_input = gr.Slider(
                label="Perturbation Level",
                minimum=0, maximum=1,
                step=0.1, value=0.2, # Mapping low/medium/high to a slider
                # Note: The attack.py code will need a slight modification to handle a float level
                # Or, use a dropdown: choices=["low", "medium", "high"]
            )
            submit_btn = gr.Button("Run Attack", variant="primary")

        with gr.Column(scale=2):
            gr.Markdown("### Attack Results")
            original_query_output = gr.Textbox(label="Original Query (from input)", interactive=False)
            perturbed_query_output = gr.Textbox(label="Adversarial (Perturbed) Query", interactive=False)
            original_response_output = gr.Textbox(label="✅ Normal Response", interactive=False)
            perturbed_response_output = gr.Textbox(label="🔴 Perturbed Response", interactive=False)
            
            with gr.Row():
                defense_status_output = gr.Textbox(label="Defense Status", scale=2)
                response_sim_output = gr.Textbox(label="Response Similarity")
                ari_output = gr.Textbox(label="ARI Score")

    # Connect the button to the function
    submit_btn.click(
        fn=run_adversarial_test,
        inputs=[query_input, attack_method_input, perturbation_level_input],
        outputs=[
            original_query_output,
            perturbed_query_output,
            original_response_output,
            perturbed_response_output,
            defense_status_output,
            response_sim_output,
            ari_output
        ]
    )

# Launch the interface
if __name__ == "__main__":
    demo.launch()