benchmark-e2e.py

import time
import requests
from pydantic import BaseModel


class Query(BaseModel):
    query: str

def send_query(query_text: str):
    headers = {
        "Host": "www.preretrieval.leonliang.lu"
    }
    query = Query(query=query_text)

    url = "http://localhost:9000/query"
    start_total = time.time()

    for i in range(10):
        try:
            start = time.time()
            response = requests.post(url, json=query.dict(), headers=headers)
            response.raise_for_status()
            duration = time.time() - start  
            total_duration = time.time() - start_total 
            print(f"Request succeeded on attempt {i+1} in {duration:.3f}s, total elapsed {total_duration:.3f}s")
            return response.json(), total_duration
        except requests.exceptions.RequestException as e:
            print(f"[Attempt {i+1}] Request failed: {e}")
            time.sleep(0.5)

    total_duration = time.time() - start_total
    raise Exception(f"Query failed after 10 retries. Total elapsed: {total_duration:.3f}s")

response, duration = send_query("What's the default batch size?")
print("Response:", response)
print(f"Duration: {duration:.3f} seconds")