app.py

import gradio as gr
from sentence_transformers import SentenceTransformer, util
import torch

# Load the model (Qwen3-Embedding-0.6B)
# trust_remote_code is required for some Qwen architectures
model = SentenceTransformer("Qwen/Qwen3-Embedding-0.6B", trust_remote_code=True)

def embed_text(text, is_query=True):
    """
    API 1: Text Embedding
    Qwen3 benefits from a 'query' prompt for retrieval tasks.
    """
    prompt_name = "query" if is_query else None
    embedding = model.encode(text, prompt_name=prompt_name, convert_to_tensor=True)
    return embedding.tolist()

def calculate_similarity(text_a, text_b, is_query=True):
    """
    API 2: Embedding Similarity
    Returns a float between 0 and 1 (clamped) representing the similarity.
    """
    prompt_name = "query" if is_query else None
    
    # Encode both texts
    emb_a = model.encode(text_a, prompt_name=prompt_name, convert_to_tensor=True)
    emb_b = model.encode(text_b, prompt_name=prompt_name, convert_to_tensor=True)
    
    # Compute Cosine Similarity
    similarity = util.cos_sim(emb_a, emb_b).item()
    
    # Clamp to [0, 1] for "percentage" logic
    score = max(0, min(1, similarity))
    percentage = f"{score * 100:.2f}%"
    
    return score, percentage

# Building the Gradio UI
with gr.Blocks(title="Qwen3 Embedding API") as demo:
    gr.Markdown("# Qwen3-Embedding-0.6B API & UI")
    gr.Markdown("This space provides high-quality text embeddings and similarity scores.")

    with gr.Tab("Text Embedding"):
        with gr.Row():
            with gr.Column():
                input_text = gr.Textbox(label="Input Text", placeholder="Enter text to embed...")
                is_query_toggle = gr.Checkbox(label="Is this a search query?", value=True, info="Uses Qwen's specific query prompt for better retrieval.")
                btn_embed = gr.Button("Generate Embedding", variant="primary")
            with gr.Column():
                output_vec = gr.JSON(label="Embedding Vector (Truncated in UI)")
        
        btn_embed.click(fn=embed_text, inputs=[input_text, is_query_toggle], outputs=output_vec, api_name="embed")

    with gr.Tab("Similarity Score"):
        with gr.Row():
            with gr.Column():
                text_a = gr.Textbox(label="Text A", placeholder="First sentence...")
                text_b = gr.Textbox(label="Text B", placeholder="Second sentence...")
                is_query_sim = gr.Checkbox(label="Use query prompts?", value=True)
                btn_sim = gr.Button("Compare Texts", variant="primary")
            with gr.Column():
                sim_float = gr.Number(label="Similarity Score (0-1)")
                sim_percent = gr.Label(label="Match Percentage")
        
        # API returns both the float and the label string
        btn_sim.click(fn=calculate_similarity, inputs=[text_a, text_b, is_query_sim], outputs=[sim_float, sim_percent], api_name="similarity")

    gr.Markdown("""
    ### How to use the API
    You can call these endpoints programmatically using the Gradio Python Client:
    ```python
    from gradion_client import Client
    client = Client("your-username/your-space-name")
    
    # API 1: Embedding
    result = client.predict("Hello world", True, api_name="/embed")
    
    # API 2: Similarity
    score, percent = client.predict("Text A", "Text B", True, api_name="/similarity")
    ```
    """)

if __name__ == "__main__":
    demo.launch()