app.py

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

# ————————————————
# 1) 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)

# ————————————————
# 2) Available zero-shot models
# ————————————————
MODEL_CHOICES = [
    "facebook/bart-large-mnli",
    "roberta-large-mnli",
    "joeddav/xlm-roberta-large-xnli",
    "mistralai/Mistral-7B-Instruct",
    "huggyllama/llama-2-7b-chat",
    "google/flan-t5-large",
    "google/flan-ul2",
    "clare-ai/llama-2-13b-instruct",
    "allenai/longformer-base-4096",
    "facebook/bart-large-mnli",  # duplicate to test allow_custom_value 
    "valhalla/t5-base-qa-qg-hl",
    "EleutherAI/gpt-neox-20b",
    "EleutherAI/gpt-j-6b",
    "bigscience/bloom-1b1",
    "bigscience/bloom-560m",
    "bigscience/bloom-3b",
    "Salesforce/codegen-2B-multi",
    "Salesforce/codegen-350M-multi",
    "madlag/llama2-7b-finetuned-qa",
    "tiiuae/falcon-7b-instruct",
    "tiiuae/falcon-40b-instruct",
    "milvus/milvus-embed-english",
    "sentence-transformers/all-MiniLM-L6-v2",
    "YOUR-OWN-CUSTOM-MODEL"
]

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

# ————————————————
# 3) Build the interface logic
# ————————————————

def hierarchical_tag(question, model_name):
    start = time.time()
    # 3.1 Instantiate classifier per-run (to change models dynamically)
    clf = pipeline("zero-shot-classification", model=model_name)

    # 3.2 Stage 1: coarse label
    coarse_out = clf(question, candidate_labels=coarse_labels)
    subject = coarse_out["labels"][0]

    # 3.3 Stage 2: fine labels within chosen subject
    fine_labels = fine_map.get(subject, [])
    fine_out = clf(question, candidate_labels=fine_labels)
    top3 = fine_out["labels"][:3]

    duration = round(time.time() - start, 3)

    # 3.4 Log the run
    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(top3),
            duration
        ])

    # 3.5 Return for display
    return subject, {lbl: round(score,3) 
                     for lbl,score in zip(fine_out["labels"][:3],
                                         fine_out["scores"][:3]
                                        )}, f"⏱ {duration}s"

def submit_feedback(question, subject, topics, corrected):
    ts = time.strftime("%Y-%m-%d %H:%M:%S")
    with open(FEEDBACK_FILE, "a", newline="") as f:
        csv.writer(f).writerow([ts, question.replace("\n"," "), subject, ";".join(topics), corrected])
    return "Thank you for your feedback!"

# ————————————————
# 4) Define the Gradio UI
# ————————————————
with gr.Blocks() as demo:
    gr.Markdown("## Hierarchical Zero-Shot Tagger with Model Selection & Logging")

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

    run_button = gr.Button("Tag Question")

    subject_out = gr.Textbox(label="Predicted Subject")
    topics_out  = gr.Label(label="Top-3 Topics")
    time_out    = gr.Textbox(label="Inference Time")

    run_button.click(
        hierarchical_tag,
        inputs=[question_input, model_input],
        outputs=[subject_out, topics_out, time_out]
    )

    gr.Markdown("---")
    gr.Markdown("### Not quite right? Submit your corrections below:")

    corrected_input = gr.Textbox(lines=1, placeholder="Correct subject;topic1;topic2;topic3")
    feedback_button = gr.Button("Submit Feedback")
    feedback_status = gr.Textbox(label="")

    feedback_button.click(
        submit_feedback,
        inputs=[question_input, subject_out, topics_out, corrected_input],
        outputs=[feedback_status]
    )

demo.launch()