app.py

import gradio as gr
from fastapi import FastAPI, Request
import uvicorn
from sentence_transformers import SentenceTransformer
from sentence_transformers.util import cos_sim
from sentence_transformers.quantization import quantize_embeddings
import spaces
from gradio_client import Client
import json
import os


app = FastAPI()


    
@app.post("/v1/embeddings")
async def openai_embeddings(request: Request):
    body = await request.json();
    token = request.headers.get("authorization");
    apiName = body.get("ApiName");
    
    print(body);
    
    BearerToken = None;
    if not token is None:
        parts = token.split(' ');
        BearerToken = parts[1];
        print("Using token...");
        
    SpacePath = body['model']
    
    print("Creating client...");
    SpaceClient = Client(SpacePath, token = BearerToken)
    print(" Created!");
    
    if not apiName:
        apiName = "/embed"
    
    text = body['input'];
    
    print("Predicting...");
    result = SpaceClient.predict(
            text=text,
            api_name=apiName
        )
        
    print("Loading result embeddings from json...");
    embeddings = json.loads(result);
 
    print("Returning...");
    return {
		'object': "list"
		,'data': [{
			'object': "embeddings"
			,'embedding': embeddings
			,'index':0
		}]
		,'model': SpacePath
		,'usage':{
			 'prompt_tokens': 0
			,'total_tokens': 0
		}
	}
    
SpaceHost = os.environ.get("SPACE_HOST");

if not SpaceHost:
    SpaceHost = "localhost"
    
    
with gr.Blocks() as demo:
    gr.Markdown(f"""
        This space allow you connect SQL Server 2025 with Hugging Face to generate embeddings!  
        First, create a ZeroGPU Space that export an endpoint called /embed.  
        That endpoint must accept a parameter called text.  
        Then, create the external model using T-SQL:
            
        ```sql 
            CREATE EXTERNAL MODEL HuggingFace
            WITH (
                  LOCATION = 'https://{SpaceHost}/v1/embeddings',
                  API_FORMAT = 'OpenAI',
                  MODEL_TYPE = EMBEDDINGS,
                  MODEL = 'user/space'
            );  
        ```
        
        If you prefer, just type the space name into field bellow and we generate the right T-SQL command for you!  
        
        You can duplicate this space to your own org and use token with CREDENTIAL parameter.  
        
    
    """)
    
    SpaceName = gr.Textbox(label="Space", submit_btn=True)
    EndpointName = gr.Textbox(value="/embed", label = "EndpointName");
    tsqlCommand = gr.Textbox(lines=5, label = "Generated T-SQL command");
    
    
    def UpdateTsql(space):
        return f"""
            CREATE EXTERNAL MODEL [MyHfModel] /*You can change the name here!*/
            WITH (
                  LOCATION = 'https://{SpaceHost}/v1/embeddings',
                  API_FORMAT = 'OpenAI',
                  MODEL_TYPE = EMBEDDINGS,
                  MODEL = '{space}'
            )
        """
        
    
    SpaceName.submit(UpdateTsql, [SpaceName], [tsqlCommand])
    
    


    
## hack para funcionar com ZeroGPU nesse mesmo space
#print("Demo run...");
#(app2,url,other) = demo.launch(prevent_thread_lock=True, server_name=None, server_port=8000);
# demo.close 

print("Mounting app...");  
GradioApp = gr.mount_gradio_app(app, demo, path="", ssr_mode=False); 


if __name__ == '__main__':
    print("Running uviconr...");
    uvicorn.run(GradioApp, host="0.0.0.0", port=7860)