Create app.py

app.py

@@ -0,0 +1,135 @@
+from typing import Any
+
+import gradio as gr
+import spaces
+from sentence_transformers import CrossEncoder
+
+
+MODEL_ID = "ramitha2002/genieai-product-reranker"
+
+# ZeroGPU requires CUDA placement at module level.
+model = CrossEncoder(
+    MODEL_ID,
+    device="cuda",
+    max_length=384,
+)
+
+
+def format_value(value: Any) -> str:
+    if isinstance(value, list):
+        return ", ".join(str(item) for item in value)
+    return str(value)
+
+
+def build_product_text(product: dict[str, Any]) -> str:
+    fields = [
+        ("Title", product.get("title") or product.get("name")),
+        ("Description", product.get("description") or product.get("summary")),
+        ("Features", product.get("features")),
+        ("Brand", product.get("brand")),
+        ("Color", product.get("color")),
+        ("Category", product.get("category")),
+    ]
+
+    return "\n".join(
+        f"{label}: {format_value(value)}"
+        for label, value in fields
+        if value is not None and value != ""
+    )
+
+
+@spaces.GPU(duration=30)
+def rerank(
+    query: str,
+    products: list[dict[str, Any]],
+    top_n: int,
+) -> dict[str, Any]:
+    query = query.strip()
+
+    if not query:
+        raise gr.Error("Query is required.")
+
+    if not isinstance(products, list) or not products:
+        raise gr.Error("Products must be a non-empty JSON array.")
+
+    if len(products) > 30:
+        raise gr.Error("Maximum 30 products per request.")
+
+    pairs = [
+        (query, build_product_text(product))
+        for product in products
+    ]
+
+    scores = model.predict(
+        pairs,
+        batch_size=min(16, len(pairs)),
+        show_progress_bar=False,
+    )
+
+    ranked = sorted(
+        [
+            {
+                **product,
+                "rerankerScore": float(score),
+            }
+            for product, score in zip(products, scores)
+        ],
+        key=lambda product: product["rerankerScore"],
+        reverse=True,
+    )
+
+    return {
+        "results": ranked[:max(1, min(int(top_n), len(ranked)))]
+    }
+
+
+sample_products = [
+    {
+        "id": "flowers-1",
+        "name": "Pink Rose Bouquet",
+        "description": "Fresh roses arranged for birthdays",
+        "brand": "Bloom House",
+        "color": "Pink"
+    },
+    {
+        "id": "mouse-1",
+        "name": "Wireless Gaming Mouse",
+        "description": "RGB computer mouse",
+        "brand": "GamePoint",
+        "color": "Black"
+    }
+]
+
+
+with gr.Blocks(title="GenieAI Product Reranker") as demo:
+    gr.Markdown("# GenieAI Product Reranker")
+
+    query_input = gr.Textbox(
+        label="Search query",
+        value="birthday flowers for mother",
+    )
+
+    products_input = gr.JSON(
+        label="Products",
+        value=sample_products,
+    )
+
+    top_n_input = gr.Slider(
+        minimum=1,
+        maximum=30,
+        value=4,
+        step=1,
+        label="Number of results",
+    )
+
+    rerank_button = gr.Button("Rerank", variant="primary")
+    output = gr.JSON(label="Ranked products")
+
+    rerank_button.click(
+        fn=rerank,
+        inputs=[query_input, products_input, top_n_input],
+        outputs=output,
+        api_name="rerank",
+    )
+
+demo.queue(default_concurrency_limit=2).launch()