app.py

import torch
import gradio as gr
from datasets import load_dataset
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# -----------------------------
# LOAD SEMANTIC DATASET
# -----------------------------
DATASET_ID = "Talip7/scikit-learn-issues-embeddings-mpnet"

train_ds = load_dataset(DATASET_ID, split="train")
train_ds = train_ds.add_faiss_index(column="embedding")

# -----------------------------
# LOAD EMBEDDING MODEL
# -----------------------------
EMBEDDING_MODEL = "sentence-transformers/all-mpnet-base-v2"

encoder = SentenceTransformer(
    EMBEDDING_MODEL,
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# -----------------------------
# LOAD MULTILABEL CLASSIFIER
# -----------------------------
CLASSIFIER_ID = "Talip7/scikit-learn-multilabel-classifier"

tokenizer = AutoTokenizer.from_pretrained(CLASSIFIER_ID)
clf_model = AutoModelForSequenceClassification.from_pretrained(
    CLASSIFIER_ID,
    problem_type="multi_label_classification"
).to("cuda" if torch.cuda.is_available() else "cpu")

clf_model.eval()

# -----------------------------
# UTILS
# -----------------------------
def predict_labels(text, threshold=0.5):
    inputs = tokenizer(
        text,
        truncation=True,
        padding=True,
        max_length=512,
        return_tensors="pt"
    )
    inputs = {k: v.to(clf_model.device) for k, v in inputs.items()}

    with torch.no_grad():
        logits = clf_model(**inputs).logits

    probs = torch.sigmoid(logits)[0].cpu().numpy()

    labels = []
    for i, p in enumerate(probs):
        if p >= threshold:
            labels.append(clf_model.config.id2label[i])

    return labels


def semantic_search(query, k=15):
    query_emb = encoder.encode(query, convert_to_numpy=True)
    scores, samples = train_ds.get_nearest_examples(
        "embedding",
        query_emb,
        k=k
    )
    return scores, samples


def hybrid_search(query, alpha=0.7, beta=0.3, max_results=5):
    sem_scores, sem_results = semantic_search(query)
    predicted_labels = set(predict_labels(query))

    seen = set()
    results = []

    for i in range(len(sem_scores)):
        issue_id = sem_results["issue_number"][i]
        if issue_id in seen:
            continue
        seen.add(issue_id)

        issue_labels = set(sem_results["labels"][i])
        overlap = (
            len(issue_labels & predicted_labels) / len(issue_labels)
            if issue_labels else 0.0
        )

        final_score = alpha * float(sem_scores[i]) + beta * overlap

        results.append({
            "Issue": f"#{issue_id}",
            "Final score": round(final_score, 3),
            "Semantic": round(float(sem_scores[i]), 3),
            "Label overlap": round(overlap, 2),
            "Labels": ", ".join(issue_labels),
            "URL": sem_results["html_url"][i],
        })

        if len(results) >= max_results:
            break

    return list(predicted_labels), results


def run_search(query):
    if not query.strip():
        return "Please enter a query.", ""

    labels, results = hybrid_search(query)
    label_text = ", ".join(labels) if labels else "No label confidently predicted"

    html = """
    <div style="width:100%; overflow-x:auto;">
    <table style="width:100%; border-collapse:collapse;">
        <thead>
            <tr>
                <th style="border:1px solid #555; padding:8px;">Issue</th>
                <th style="border:1px solid #555; padding:8px; text-align:center;">Final</th>
                <th style="border:1px solid #555; padding:8px; text-align:center;">Semantic</th>
                <th style="border:1px solid #555; padding:8px; text-align:center;">Label overlap</th>
                <th style="border:1px solid #555; padding:8px;">Labels</th>
                <th style="border:1px solid #555; padding:8px;">GitHub URL</th>
            </tr>
        </thead>
        <tbody>
    """

    for r in results:
        html += f"""
        <tr>
            <td style="border:1px solid #555; padding:8px;">{r['Issue']}</td>
            <td style="border:1px solid #555; padding:8px; text-align:center;">{r['Final score']}</td>
            <td style="border:1px solid #555; padding:8px; text-align:center;">{r['Semantic']}</td>
            <td style="border:1px solid #555; padding:8px; text-align:center;">{r['Label overlap']}</td>
            <td style="border:1px solid #555; padding:8px;">{r['Labels']}</td>
            <td style="border:1px solid #555; padding:8px;">
                <a href="{r['URL']}" target="_blank">{r['URL']}</a>
            </td>
        </tr>
        """

    html += "</tbody></table></div>"

    return label_text, html

# -----------------------------
# GRADIO UI
# -----------------------------
with gr.Blocks(
    title="GitHub Issue Hybrid Search",
    css="""
    /* Sayfanın ana container'ını genişlet */
    .gradio-container {
        max-width: 100% !important;
        padding-left: 40px;
        padding-right: 40px;
    }

    /* Markdown tablo tam genişlik */
    .gr-markdown table {
        width: 100%;
        table-layout: fixed;
    }

    .gr-markdown th, .gr-markdown td {
        text-align: left;
        padding: 8px;
        word-wrap: break-word;
    }

    /* Sayısal kolonlar ortalansın */
    .gr-markdown td:nth-child(2),
    .gr-markdown td:nth-child(3),
    .gr-markdown td:nth-child(4) {
        text-align: center;
    }
    """
    
) as demo:

    gr.Markdown(
        """
        # 🤗 GitHub Issue Hybrid Search & Auto-Label Assistant

        **Semantic Search (MPNet) + Multilabel Classification (DistilBERT)**  
        Precision-first hybrid ranking on real scikit-learn issues.
        """
    )

    query = gr.Textbox(
        label="Describe the issue",
        placeholder="e.g. RandomForestClassifier crashes when sample_weight is None"
    )

    btn = gr.Button("Search")

    predicted = gr.Textbox(label="Predicted labels", interactive=False)
    results_html = gr.HTML()


    btn.click(
        fn=run_search,
        inputs=query,
        outputs=[predicted, results_html]
    )

demo.launch()