app.py

import json, time, csv, os
import gradio as gr
from transformers import pipeline


# ————————————————
# Load taxonomies
# ————————————————
with open("coarse_labels.json") as f:
    coarse_labels = json.load(f)
with open("fine_labels.json") as f:
    fine_map = json.load(f)

# ————————————————
# Model choices (5 only)
# ————————————————
MODEL_CHOICES = [
    "facebook/bart-large-mnli",
    "roberta-large-mnli",
    "joeddav/xlm-roberta-large-xnli",
    "valhalla/distilbart-mnli-12-4",
    "MoritzLaurer/mDeBERTa-v3-base-xnli-multilingual-nli-2mil7" ,
    "MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli"# placeholder — replace with real phantom model
]

PIPELINES = {}

def get_pipeline(name):
    if name not in PIPELINES:
        PIPELINES[name] = pipeline("zero-shot-classification", model=name)
    return PIPELINES[name]

# ————————————————
# Ensure log files exist
# ————————————————
LOG_FILE      = "logs.csv"
FEEDBACK_FILE = "feedback.csv"
for fn, hdr in [
    (LOG_FILE,      ["timestamp","model","question","chosen_subject","top3_topics","duration"]),
    (FEEDBACK_FILE, ["timestamp","question","subject_feedback","topic_feedback"])
]:
    if not os.path.exists(fn):
        with open(fn, "w", newline="") as f:
            csv.writer(f).writerow(hdr)

# ————————————————
# Inference functions
# ————————————————
def run_stage1(question, model_name):
    if not question or not question.strip():
        return {}, gr.update(choices=[]), ""
    start = time.time()
    clf = get_pipeline(model_name)
    out = clf(question, candidate_labels=coarse_labels)
    labels, scores = out["labels"][:3], out["scores"][:3]
    duration = round(time.time() - start, 3)

    # Prepare outputs
    subject_dict = {lbl: round(score,3) for lbl,score in zip(labels, scores)}
    radio_update = gr.update(choices=labels, value=labels[0])
    time_str    = f"⏱ {duration}s"

    return subject_dict, radio_update, time_str


def run_stage2(question, model_name, subject):
    # 1) Validate inputs
    if not question or not question.strip():
        return {}, "No question provided", ""
    fine_labels = fine_map.get(subject, [])
    if not fine_labels:
        return {}, f"No topics found for '{subject}'", ""

    # 2) Inference (fast, using preloaded pipeline)
    start = time.time()
    clf = get_pipeline(model_name)
    out = clf(question, candidate_labels=fine_labels)
    labels, scores = out["labels"][:3], out["scores"][:3]
    duration = round(time.time() - start, 3)

    # 3) Logging
    with open(LOG_FILE, "a", newline="") as f:
        csv.writer(f).writerow([
            time.strftime("%Y-%m-%d %H:%M:%S"),
            model_name,
            question.replace("\n"," "),
            subject,
            ";".join(labels),
            duration
        ])

    # 4) Return topics + time
    topic_dict = {lbl: round(score,3) for lbl,score in zip(labels, scores)}
    return topic_dict, f"⏱ {duration}s"


def submit_feedback(question, subject_fb, topic_fb):
    with open(FEEDBACK_FILE, "a", newline="") as f:
        csv.writer(f).writerow([
            time.strftime("%Y-%m-%d %H:%M:%S"),
            question.replace("\n"," "),
            subject_fb,
            topic_fb
        ])
    return "✅ Feedback recorded!"

# ————————————————
# Build Gradio UI
# ————————————————
with gr.Blocks() as demo:
    gr.Markdown("## Hierarchical Zero-Shot Tagger with Subject Toggle & Feedback")

    with gr.Row():
        question_input = gr.Textbox(lines=3, label="Enter your question")
        model_input    = gr.Dropdown(choices=MODEL_CHOICES, value=MODEL_CHOICES[0], label="Choose model")
        go_button      = gr.Button("Run Stage 1")

    subject_out = gr.Label(num_top_classes=3, label="Top-3 Subjects")
    subj_radio  = gr.Radio(choices=[], label="Select Subject for Stage 2")
    stage1_time = gr.Textbox(label="Stage 1 Time")
    
    go_button.click(
        fn=run_stage1,
        inputs=[question_input, model_input],
        outputs=[subject_out, subj_radio, stage1_time]
    )

    # Stage 2 UI
    go2_button  = gr.Button("Run Stage 2")
    topics_out  = gr.Label(label="Top-3 Topics")
    stage2_time = gr.Textbox(label="Stage 2 Time")

    go2_button.click(
        fn=run_stage2,
        inputs=[question_input, model_input, subj_radio],
        outputs=[topics_out, stage2_time]
    )

    gr.Markdown("---")
    gr.Markdown("### Feedback / Correction")

    subject_fb = gr.Textbox(label="Correct Subject")
    topic_fb   = gr.Textbox(label="Correct Topic(s)")
    fb_button  = gr.Button("Submit Feedback")
    fb_status  = gr.Textbox(label="")

    fb_button.click(
        fn=submit_feedback,
        inputs=[question_input, subject_fb, topic_fb],
        outputs=[fb_status]
    )

demo.launch(share=True)