View Notebook on Github

​

Assistants API Overview with AgentOps

This notebook has been adapted from this OpenAI Cookbook example.

The new Assistants API is a stateful evolution of our Chat Completions API meant to simplify the creation of assistant-like experiences, and enable developer access to powerful tools like Code Interpreter and Retrieval.

​

Chat Completions API vs Assistants API

The primitives of the Chat Completions API are Messages, on which you perform a Completion with a Model (gpt-3.5-turbo, gpt-4, etc). It is lightweight and powerful, but inherently stateless, which means you have to manage conversation state, tool definitions, retrieval documents, and code execution manually.

The primitives of the Assistants API are

• Assistants, which encapsulate a base model, instructions, tools, and (context) documents,
• Threads, which represent the state of a conversation, and
• Runs, which power the execution of an Assistant on a Thread, including textual responses and multi-step tool use.

We’ll take a look at how these can be used to create powerful, stateful experiences.

​

Setup

Note The Assistants API is currently in beta so the latest Python SDK is needed (1.58.1 at time of writing) for this example.

%pip install -U openai
%pip install -U agentops
%pip install -U python-dotenv

​

Pretty Printing Helper

import json

def show_json(obj):
    display(json.loads(obj.model_dump_json()))

​

Complete Example with Assistants API

​

Assistants

The easiest way to get started with the Assistants API is through the Assistants Playground.

Let’s begin by creating an assistant! We’ll create a Math Tutor just like in our docs.

You can view Assistants you’ve created in the Assistants Dashboard.

You can also create Assistants directly through the Assistants API. But we need to have the AgentOps and OpenAI API keys first.

You can get your OpenAI API key from the OpenAI Dashboard.

To obtain the AgentOps API key, signup for an account on AgentOps and create a project. After creating the project, you can now create an API key in the Project Settings.

Next, we’ll set our API keys. There are several ways to do this, the code below is just the most foolproof way for the purposes of this notebook. It accounts for both users who use environment variables and those who just want to set the API Key here in this notebook.

1. Create an environment variable in a .env file or other method. By default, the AgentOps init() function will look for an environment variable named AGENTOPS_API_KEY. Or…

2. Replace <your_agentops_key> below and pass in the optional api_key parameter to the AgentOps init(api_key=...) function. Remember not to commit your API key to a public repo!

Now we are all set! Let’s import the necessary libraries and initialize the AgentOps and OpenAI clients.

from openai import OpenAI
import agentops
from dotenv import load_dotenv
import os

load_dotenv()
AGENTOPS_API_KEY = os.getenv("AGENTOPS_API_KEY") or "<your_agentops_key>"
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY") or "<your_openai_key>"

agentops.init(api_key=AGENTOPS_API_KEY, default_tags=["openai", "beta-assistants"])
client = OpenAI(api_key=OPENAI_API_KEY)

Next, we’ll create an Assistant which will be our Math Tutor.

assistant = client.beta.assistants.create(
    name="Math Tutor",
    instructions="You are a personal math tutor. Answer questions briefly, in a sentence or less.",
    model="gpt-4o-mini",
)
show_json(assistant)

Regardless of whether you create your Assistant through the Dashboard or with the API, you’ll want to keep track of the Assistant ID. This is how you’ll refer to your Assistant throughout Threads and Runs.

Next, we’ll create a new Thread and add a Message to it. This will hold the state of our conversation, so we don’t have re-send the entire message history each time.

​

Threads

Create a new thread:

thread = client.beta.threads.create()
show_json(thread)

Then add the Message to the thread:

message = client.beta.threads.messages.create(
    thread_id=thread.id,
    role="user",
    content="I need to solve the equation `3x + 11 = 14`. Can you help me?",
)
show_json(message)

Note Even though you’re no longer sending the entire history each time, you will still be charged for the tokens of the entire conversation history with each Run.

​

Runs

Notice how the Thread we created is not associated with the Assistant we created earlier! Threads exist independently from Assistants, which may be different from what you’d expect if you’ve used ChatGPT (where a thread is tied to a model/GPT).

To get a completion from an Assistant for a given Thread, we must create a Run. Creating a Run will indicate to an Assistant it should look at the messages in the Thread and take action: either by adding a single response, or using tools.

Note Runs are a key difference between the Assistants API and Chat Completions API. While in Chat Completions the model will only ever respond with a single message, in the Assistants API a Run may result in an Assistant using one or multiple tools, and potentially adding multiple messages to the Thread.

To get our Assistant to respond to the user, let’s create the Run. As mentioned earlier, you must specify both the Assistant and the Thread.

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

Unlike creating a completion in the Chat Completions API, creating a Run is an asynchronous operation. It will return immediately with the Run’s metadata, which includes a status that will initially be set to queued. The status will be updated as the Assistant performs operations (like using tools and adding messages).

To know when the Assistant has completed processing, we can poll the Run in a loop. (Support for streaming is coming soon!) While here we are only checking for a queued or in_progress status, in practice a Run may undergo a variety of status changes which you can choose to surface to the user. (These are called Steps, and will be covered later.)

import time

def wait_on_run(run, thread):
    while run.status == "queued" or run.status == "in_progress":
        run = client.beta.threads.runs.retrieve(
            thread_id=thread.id,
            run_id=run.id,
        )
        time.sleep(0.5)
    return run

run = wait_on_run(run, thread)
show_json(run)

​

Messages

Now that the Run has completed, we can list the Messages in the Thread to see what got added by the Assistant.

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

As you can see, Messages are ordered in reverse-chronological order – this was done so the most recent results are always on the first page (since results can be paginated). Do keep a look out for this, since this is the opposite order to messages in the Chat Completions API.

Let’s ask our Assistant to explain the result a bit further!

# Create a message to append to our thread
message = client.beta.threads.messages.create(
    thread_id=thread.id, role="user", content="Could you explain this to me?"
)

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

# Wait for completion
wait_on_run(run, thread)

# Retrieve all the messages added after our last user message
messages = client.beta.threads.messages.list(
    thread_id=thread.id, order="asc", after=message.id
)
show_json(messages)

This may feel like a lot of steps to get a response back, especially for this simple example. However, you’ll soon see how we can add very powerful functionality to our Assistant without changing much code at all!

Et voilà!

You may have noticed that this code is not actually specific to our math Assistant at all… this code will work for any new Assistant you create simply by changing the Assistant ID! That is the power of the Assistants API.

​

Tools

A key feature of the Assistants API is the ability to equip our Assistants with Tools, like Code Interpreter, Retrieval, and custom Functions. Let’s take a look at each.

​

Code Interpreter

Let’s equip our Math Tutor with the Code Interpreter tool, which we can do from the Dashboard…

…or the API, using the Assistant ID.

assistant = client.beta.assistants.update(
    MATH_ASSISTANT_ID,
    tools=[{"type": "code_interpreter"}],
)
show_json(assistant)

Now, let’s ask the Assistant to use its new tool.

thread, run = create_thread_and_run(
    "Generate the first 20 fibbonaci numbers with code."
)
run = wait_on_run(run, thread)
pretty_print(get_response(thread))

And that’s it! The Assistant used Code Interpreter in the background, and gave us a final response.

For some use cases this may be enough – however, if we want more details on what precisely an Assistant is doing we can take a look at a Run’s Steps.

​

Steps

A Run is composed of one or more Steps. Like a Run, each Step has a status that you can query. This is useful for surfacing the progress of a Step to a user (e.g. a spinner while the Assistant is writing code or performing retrieval).

run_steps = client.beta.threads.runs.steps.list(
    thread_id=thread.id, run_id=run.id, order="asc"
)

Let’s take a look at each Step’s step_details.

for step in run_steps.data:
    step_details = step.step_details
    print(json.dumps(show_json(step_details), indent=4))

We can see the step_details for two Steps:

1. tool_calls (plural, since it could be more than one in a single Step)
2. message_creation

The first Step is a tool_calls, specifically using the code_interpreter which contains:

• input, which was the Python code generated before the tool was called, and
• output, which was the result of running the Code Interpreter.

The second Step is a message_creation, which contains the message that was added to the Thread to communicate the results to the user.

​

Retrieval

Another powerful tool in the Assistants API is Retrieval: the ability to upload files that the Assistant will use as a knowledge base when answering questions. This can also be enabled from the Dashboard or the API, where we can upload files we want to be used.

# Upload the file
file = client.files.create(
    file=open(
        "language_models_are_unsupervised_multitask_learners.pdf",
        "rb",
    ),
    purpose="assistants",
)
# Update Assistant
assistant = client.beta.assistants.update(
    MATH_ASSISTANT_ID,
    tools=[{"type": "code_interpreter"}],
    tool_resources={"code_interpreter": {"file_ids": [file.id]}},
)
show_json(assistant)

thread, run = create_thread_and_run(
    "What are some cool math concepts behind this ML paper pdf? Explain in two sentences."
)
run = wait_on_run(run, thread)
pretty_print(get_response(thread))

Note There are more intricacies in Retrieval, like Annotations, which may be covered in another cookbook.

​

Functions

As a final powerful tool for your Assistant, you can specify custom Functions (much like the Function Calling in the Chat Completions API). During a Run, the Assistant can then indicate it wants to call one or more functions you specified. You are then responsible for calling the Function, and providing the output back to the Assistant.

Let’s take a look at an example by defining a display_quiz() Function for our Math Tutor.

This function will take a title and an array of questions, display the quiz, and get input from the user for each:

• title
• questions
  • question_text
  • question_type: [MULTIPLE_CHOICE, FREE_RESPONSE]
  • choices: [“choice 1”, “choice 2”, …]

Unfortunately I don’t know how to get user input within a Python Notebook, so I’ll be mocking out responses with get_mock_response.... This is where you’d get the user’s actual input.

def get_mock_response_from_user_multiple_choice():
    return "a"


def get_mock_response_from_user_free_response():
    return "I don't know."


def display_quiz(title, questions):
    print("Quiz:", title)
    print()
    responses = []

    for q in questions:
        print(q["question_text"])
        response = ""

        # If multiple choice, print options
        if q["question_type"] == "MULTIPLE_CHOICE":
            for i, choice in enumerate(q["choices"]):
                print(f"{i}. {choice}")
            response = get_mock_response_from_user_multiple_choice()

        # Otherwise, just get response
        elif q["question_type"] == "FREE_RESPONSE":
            response = get_mock_response_from_user_free_response()

        responses.append(response)
        print()

    return responses

Here’s what a sample quiz would look like:

responses = display_quiz(
    "Sample Quiz",
    [
        {"question_text": "What is your name?", "question_type": "FREE_RESPONSE"},
        {
            "question_text": "What is your favorite color?",
            "question_type": "MULTIPLE_CHOICE",
            "choices": ["Red", "Blue", "Green", "Yellow"],
        },
    ],
)
print("Responses:", responses)

Now, let’s define the interface of this function in JSON format, so our Assistant can call it:

function_json = {
    "name": "display_quiz",
    "description": "Displays a quiz to the student, and returns the student's response. A single quiz can have multiple questions.",
    "parameters": {
        "type": "object",
        "properties": {
            "title": {"type": "string"},
            "questions": {
                "type": "array",
                "description": "An array of questions, each with a title and potentially options (if multiple choice).",
                "items": {
                    "type": "object",
                    "properties": {
                        "question_text": {"type": "string"},
                        "question_type": {
                            "type": "string",
                            "enum": ["MULTIPLE_CHOICE", "FREE_RESPONSE"],
                        },
                        "choices": {"type": "array", "items": {"type": "string"}},
                    },
                    "required": ["question_text"],
                },
            },
        },
        "required": ["title", "questions"],
    },
}

Once again, let’s update our Assistant either through the Dashboard or the API.

Note Pasting the function JSON into the Dashboard was a bit finicky due to indentation, etc. I just asked ChatGPT to format my function the same as one of the examples on the Dashboard :).

assistant = client.beta.assistants.update(
    MATH_ASSISTANT_ID,
    tools=[
        {"type": "code_interpreter"},
        {"type": "function", "function": function_json},
    ],
)
show_json(assistant)

And now, we ask for a quiz.

thread, run = create_thread_and_run(
    "Make a quiz with 2 questions: One open ended, one multiple choice. Then, give me feedback for the responses."
)
run = wait_on_run(run, thread)
run.status

Now, however, when we check the Run’s status we see requires_action! Let’s take a closer look.

show_json(run)

The required_action field indicates a Tool is waiting for us to run it and submit its output back to the Assistant. Specifically, the display_quiz function! Let’s start by parsing the name and arguments.

Note While in this case we know there is only one Tool call, in practice the Assistant may choose to call multiple tools.

# Extract single tool call
tool_call = run.required_action.submit_tool_outputs.tool_calls[0]
name = tool_call.function.name
arguments = json.loads(tool_call.function.arguments)

print("Function Name:", name)
print("Function Arguments:")
arguments

Now let’s actually call our display_quiz function with the arguments provided by the Assistant:

responses = display_quiz(arguments["title"], arguments["questions"])
print("Responses:", responses)

Great! (Remember these responses are the one’s we mocked earlier. In reality, we’d be getting input from the back from this function call.)

Now that we have our responses, let’s submit them back to the Assistant. We’ll need the tool_call ID, found in the tool_call we parsed out earlier. We’ll also need to encode our listof responses into a str.

tool_outputs = []
tool_calls = run.required_action.submit_tool_outputs.tool_calls

for tool_call in tool_calls:
    arguments = json.loads(tool_call.function.arguments)
    responses = display_quiz(arguments["title"], arguments["questions"])
    tool_outputs.append({
        "tool_call_id": tool_call.id,
        "output": json.dumps(responses),
    })

run = client.beta.threads.runs.submit_tool_outputs(
    thread_id=thread.id,
    run_id=run.id,
    tool_outputs=tool_outputs
)
show_json(run)

We can now wait for the Run to complete once again, and check our Thread!

run = wait_on_run(run, thread)
pretty_print(get_response(thread))

Now let’s end the AgentOps session. By default, AgentOps will end the session in the “Intedeterminate” state. You can also end the session in the “Success” or “Failure” state.

We will end the session in the “Success” state.

agentops.end_session(end_state="Success")

Woohoo 🎉

​

Conclusion

We covered the basics of the Assistants API using OpenAI’s Python SDK and AgentOps for observability.

For more information, check out the Assistants API deep deep dive guide and its documentation.