理解OpenAI API限速：避免Assistants API中隐藏的请求陷阱

import pandas as pd
import time
from openai import OpenAI

# ... (API客户端和助手初始化代码) ...

files = ["file1.txt", "file2.txt", ...] # 假设有10个文件

jacket_classifications = pd.DataFrame(columns = ["jacket", "is_nomination"])

for file in files:
    # 1. 创建文件上传请求
    gpt_file = client.files.create(file=open(file, "rb"), purpose='assistants')

    # 2. 创建消息请求
    message = client.beta.threads.messages.create(
        thread_id=thread.id, role="user", content="...", file_ids=[gpt_file.id]
    )

    # 3. 创建Run请求
    run = client.beta.threads.runs.create(
        thread_id=thread.id, assistant_id=assistant.id
    )

    # 4. 轮询Run状态
    while run.status != "completed":
        run = client.beta.threads.runs.retrieve( # ⚠️ 此处是关键！
            thread_id=thread.id, run_id=run.id
        )
        print(run.status)
        if run.status == "failed":
            print(run.last_error)
            exit()

    # ... (处理结果代码) ...

    print("Sleeping 20 seconds to ensure API call rate limit not surpassed")
    time.sleep(20) # 循环外部的延迟

import pandas as pd
import time
from openai import OpenAI

# ... (API客户端和助手初始化代码) ...

files = ["file1.txt", "file2.txt", ...]

jacket_classifications = pd.DataFrame(columns = ["jacket", "is_nomination"])

for file in files:
    gpt_file = client.files.create(file=open(file, "rb"), purpose='assistants')
    message = client.beta.threads.messages.create(
        thread_id=thread.id, role="user", content="...", file_ids=[gpt_file.id]
    )
    run = client.beta.threads.runs.create(
        thread_id=thread.id, assistant_id=assistant.id
    )

    # 轮询Run状态，并在每次轮询后增加延迟
    while run.status != "completed":
        run = client.beta.threads.runs.retrieve(
            thread_id=thread.id, run_id=run.id
        )
        print(run.status)
        if run.status == "failed":
            print(run.last_error)
            exit()

        # ⚠️ 在轮询请求后增加延迟
        # 假设Run通常在几十秒内完成，每次轮询间隔40秒可以有效控制请求频率
        time.sleep(40) 

    # ... (处理结果代码) ...

    # 外部循环的延迟可以根据总请求量和限速进一步调整，甚至可以移除
    # print("Sleeping 20 seconds to ensure API call rate limit not surpassed")
    # time.sleep(20)

import time
from tenacity import retry, wait_exponential, stop_after_attempt, RetriableError
from openai import OpenAI

# ... (API客户端和助手初始化代码) ...

# 定义一个带有指数退避的重试函数
@retry(wait=wait_exponential(multiplier=1, min=4, max=60), stop=stop_after_attempt(10))
def call_openai_api_with_retry(api_call_func, *args, **kwargs):
    try:
        return api_call_func(*args, **kwargs)
    except Exception as e: # 捕获OpenAI API可能抛出的限速或其他错误
        print(f"API call failed, retrying... Error: {e}")
        raise RetriableError(e) # 抛出可重试错误，让tenacity捕获

# 在轮询Run状态时使用重试机制
def get_run_status_with_backoff(thread_id, run_id):
    while True:
        try:
            run = call_openai_api_with_retry(client.beta.threads.runs.retrieve, thread_id=thread_id, run_id=run_id)
            if run.status != "completed":
                print(f"Run status: {run.status}. Waiting before next check...")
                # 在轮询之间仍然可以有基础的延迟，防止过于频繁的重试
                time.sleep(5) 
            else:
                return run
        except RetriableError:
            # tenacity 会处理重试逻辑，这里可以记录日志
            print("Encountered retriable error, tenacity will handle backoff.")
            time.sleep(1) # 短暂等待，避免无限循环的日志输出
        except Exception as e:
            print(f"An unrecoverable error occurred: {e}")
            break

# ... (在主循环中使用) ...
# run = get_run_status_with_backoff(thread.id, run.id)