OPEN KNOWLEDGE: AI

OpenAI Assistant API with Code Interpreter

The tutorial explains usage of Assistant API — Code Interpreter.

Photo by Safar Safarov on Unsplash

Introduction

Assistant API enables developers to create complex workflows using: functions, code interpreter and knowledge retrieval. In this tutorial, I will explain usage of them in concrete manner, but what makes them useful?

Assistant API enables GPT-4 model to decide automatically when to retrieve knowledge, call function or use code interpreter. The automatization simplifies design of Autonomous Agents workflows.

Code interpreter

I will start by importing the required libraries and retrieve the OpenAI API key.

!pip install --upgrade openai
import os
import time
from openai import OpenAI 
client = OpenAI(
    api_key=os.getenv("openaikey"))

I next define parameters in advance, which I will be using. These variables will make more sense later in this tutorial, but you can modify them for your use case.

assistant_name = "Power Point Generator"
output_file_name = "Presentation.pptx"
assistant_instruction = r"Generate {} file, always. You are subject-matter expert in the topic and professional in creating PowerPoints.. Betweem 1-5 slides. Background, colors, fonts and styling must be modern and easy to read. Make content engaging. Make the file id available to download.".format(output_file_name)
prompt_user = "Make a presentation for runner practicing for a half-marathon with an aim for a personal record. Make a presentaton with useful insights, training plan for different levels and some tips before, during and after the training period. Give insights."

I will create next the assistant.

assistant = client.beta.assistants.create(
    name=assistant_name,
    instructions=assistant_instruction,
    tools=[{"type": "retrieval"},{"type": "code_interpreter"}],
    model="gpt-4-1106-preview")

I will create then the thread within the assistant, which makes it possible to create messages within the thread.

thread = client.beta.threads.create()

message = client.beta.threads.messages.create(
    thread_id=thread.id,
    role="user",
    content=prompt_user)

I can now run the thread to generate the respond by the assistant api.

run = client.beta.threads.runs.create(
  thread_id=thread.id,
  assistant_id=assistant.id)

Once the response is available, I can retrieve it.

timeout = 180
interval_time = 5
time_taken = 0
while time_taken < timeout:
    run = client.beta.threads.runs.retrieve(
    thread_id=thread.id,
    run_id=run.id)

    if run.status == 'completed':
        break
    else:
        print(run.status)
        time.sleep(interval_time)
        time_taken += interval_time

# Alternatively solution is to run below simpler code and ensure run.status printed is 'completed', which can take up to 2 minutes:
#run = client.beta.threads.runs.retrieve(
#    thread_id=thread.id,
#    run_id=run.id)
#time.sleep(2)
# print(run.status)

I then pull the required information about the file id and filepath.

messages = client.beta.threads.messages.list(
  thread_id=thread.id)

file_path = messages.data[0].content[0].text.annotations[0].file_path.file_id

file_name = client.files.with_raw_response.retrieve_content(file_path)

I can then save the file locally.

with open(output_file_name, "wb") as file:  
    file.write(file_name.content) 

I finally delete the file and the assistant. I remind, that the Assistant API pricing is different from the other APIs, so I recommend not to leave the Assistants or files running longer periods.

client.files.delete(file_path)
client.beta.assistants.delete(assistant.id)

Conclusions

The tutorial explained usage of the Code Interpreter within the Assistant API. I created with the Code Interpreter a Powerpoint presentation by giving it a topic and I can then retrieve this file to the local file system.

Please make sure to familiarize to the Assistant API pricing before using it.

This article is in the series: Open Knowledge: AI.

The aim is to share at least 10% of my articles without the Medium paywall for free to everybody.

References

[1] Github. https://github.com/tmgthb/LLMs/tree/main. Teemu Maatta.