Orchestrate OpenAI operations with Apache Airflow

OpenAI is an AI research and deployment company that provides an API for accessing state of the art models like GPT-4 and DALL·E 3. The OpenAI Airflow provider offers modules to easily integrate OpenAI with Airflow.

In this tutorial you'll use Airflow and the OpenAI Airflow provider to ask a question to Star Trek captains, create embeddings of the answers from each captain, and plot them in two dimensions.

Why use Airflow with OpenAI?

OpenAI offers a variety of powerful model endpoints for different tasks like text generation, vector embedding, and translation tasks. These models are used in both user-facing applications, such as chatbots, and internal applications, such as a smart search for internal knowledge base content.

Integrating OpenAI with Airflow into an end-to-end machine learning pipeline allows you to:

• Use Airflow's data-driven scheduling to run operations using OpenAI model endpoints based on upstream events in your data ecosystem, such as when new user input is ingested or a new dataset is available.
• Send several requests to a model endpoint in parallel based on upstream events in your data ecosystem or user input via Airflow params.
• Monitor the OpenAI service using Airflow alerts and protect against API rate limits and outages with Airflow retries.
• Use Airflow to orchestrate the creation of vector embeddings using OpenAI models, which is especially useful for large datasets that can't be processed automatically by vector databases.

Time to complete

This tutorial takes approximately 15 minutes to complete.

Assumed knowledge

To get the most out of this tutorial, make sure you have an understanding of:

• The basics of the OpenAI API. See OpenAI Introduction.
• The basics of vector embeddings. See the OpenAI Embeddings guide.
• Airflow fundamentals, such as writing DAGs and defining tasks. See Get started with Apache Airflow.
• Airflow operators. See Operators 101.
• Airflow hooks. See Hooks 101.

Prerequisites

• The Astro CLI.
• An OpenAI API key with at least tier 1 usage limits.

Step 1: Configure your Astro project

1. Create a new Astro project:

   $ mkdir astro-openai-tutorial && cd astro-openai-tutorial
   $ astro dev init

2. Add the following lines to your requirements.txt file to install the OpenAI Airflow provider and other supporting packages:

   apache-airflow-providers-openai==1.0.0
   openai==0.28.1
   matplotlib==3.8.1
   seaborn==0.13.0
   scikit-learn==1.3.2
   pandas==1.5.3
   numpy==1.26.2
   adjustText==0.8

3. To create an Airflow connection to OpenAI, add the following environment variables to your .env file. Make sure to replace <your-openai-api-key> with your own OpenAI API key.

   AIRFLOW_CONN_OPENAI_DEFAULT='{
       "conn_type": "openai",
       "password": "<your-openai-api-key>"
   }'

Step 2: Create your DAG

1. In your dags folder, create a file called captains_dag.py.

2. Copy the following code into the file.

   """
   ## Ask questions to Star Trek captains using OpenAI's LLMs, embed and visualize the results
   
   This DAG shows how to use the OpenAI Airflow provider to interact with the OpenAI API.
   The DAG asks a question to a list of Star Trek captains based on values you provide via 
   Airflow params, embeds the responses using the OpenAI text-embedding-ada-002 model, 
   and visualizes the embeddings in 2 dimensions using PCA, matplotlib and seaborn.
   """
   
   from airflow.decorators import dag, task
   from airflow.models.param import Param
   from airflow.models.baseoperator import chain
   from airflow.providers.openai.hooks.openai import OpenAIHook
   from airflow.providers.openai.operators.openai import OpenAIEmbeddingOperator
   from sklearn.metrics.pairwise import euclidean_distances
   from sklearn.decomposition import PCA
   from adjustText import adjust_text
   from pendulum import datetime
   import matplotlib.pyplot as plt
   import seaborn as sns
   import pandas as pd
   import numpy as np
   import openai
   
   
   OPENAI_CONN_ID = "openai_default"
   IMAGE_PATH = "include/captains_plot.png"
   
   star_trek_captains_list = [
       "James T. Kirk",
       "Jean-Luc Picard",
       "Benjamin Sisko",
       "Kathryn Janeway",
       "Jonathan Archer",
       "Christopher Pike",
       "Michael Burnham",
       "Saru",
   ]
   
   
   @dag(
       start_date=datetime(2023, 11, 1),
       schedule=None,
       catchup=False,
       params={
           "question": Param(
               "Which is your favorite ship?",
               type="string",
               title="Question to ask the captains",
               description="Enter what you would like to ask the captains.",
               min_length=1,
               max_length=500,
           ),
           "captains_to_ask": Param(
               star_trek_captains_list,
               type="array",
               description="List the captains whose answers you would like to compare. "
               + "Suggestions: "
               + ", ".join(star_trek_captains_list),
           ),
           "max_tokens_answer": Param(
               100,
               type="integer",
               description="Maximum number of tokens to generate for the answer.",
           ),
           "randomness_of_answer": Param(
               10,
               type="integer",
               description=(
                   "Enter the desired randomness of the answer on a scale"
                   + "from 0 (no randomness) to 20 (full randomness). "
                   + "This setting corresponds to 10x the temperature setting in the OpenAI API."
               ),
               min=0,
               max=20,
           ),
       },
   )
   def captains_dag():
       @task
       def get_captains_list(**context):
           "Pull the list of captains to ask from the context."
           captains_list = context["params"]["captains_to_ask"]
           return captains_list
   
       @task
       def ask_a_captain(open_ai_conn_id: str, captain_to_ask, **context):
           "Ask a captain a question using gpt-3.5-turbo."
           question = context["params"]["question"]
           max_tokens_answer = context["params"]["max_tokens_answer"]
           randomness_of_answer = context["params"]["randomness_of_answer"]
           hook = OpenAIHook(conn_id=open_ai_conn_id)
           openai.api_key = hook._get_api_key()
   
           response = openai.ChatCompletion.create(
               model="gpt-3.5-turbo",
               messages=[
                   {"role": "system", "content": f"You are captain {captain_to_ask}."},
                   {"role": "user", "content": question},
               ],
               temperature=randomness_of_answer / 10,
               max_tokens=max_tokens_answer,
           )
   
           response = response.choices[0]["message"]["content"]
   
           print(f"Your Question: {question}")
           print(f"Captain {captain_to_ask} said: {response}")
   
           return response
   
       captains_list = get_captains_list()
       captain_responses = ask_a_captain.partial(open_ai_conn_id=OPENAI_CONN_ID).expand(
           captain_to_ask=captains_list
       )
   
       get_embeddings = OpenAIEmbeddingOperator.partial(
           task_id="get_embeddings",
           conn_id=OPENAI_CONN_ID,
           model="text-embedding-ada-002",
       ).expand(input_text=captain_responses)
   
       @task
       def plot_embeddings(embeddings, text_labels, file_name="embeddings_plot.png"):
           "Plot the embeddings of the captain responses."
           pca = PCA(n_components=2)
           reduced_embeddings = pca.fit_transform(embeddings)
   
           plt.figure(figsize=(10, 8))
           df_embeddings = pd.DataFrame(reduced_embeddings, columns=["PC1", "PC2"])
           sns.scatterplot(
               df_embeddings, x="PC1", y="PC2", s=100, color="gold", edgecolor="black"
           )
   
           font_style = {"color": "black"}
           texts = []
           for i, label in enumerate(text_labels):
               texts.append(
                   plt.text(
                       reduced_embeddings[i, 0],
                       reduced_embeddings[i, 1],
                       label,
                       fontdict=font_style,
                       fontsize=15,
                   )
               )
   
           # prevent overlapping labels
           adjust_text(texts, arrowprops=dict(arrowstyle="->", color="red"))
   
           distances = euclidean_distances(reduced_embeddings)
           np.fill_diagonal(distances, np.inf)  # exclude cases where the distance is 0
   
           n = distances.shape[0]
           distances_list = [
               (distances[i, j], (i, j)) for i in range(n) for j in range(i + 1, n)
           ]
   
           distances_list.sort(reverse=True)
   
           legend_handles = []
           for dist, (i, j) in distances_list:
               (line,) = plt.plot(
                   [reduced_embeddings[i, 0], reduced_embeddings[j, 0]],
                   [reduced_embeddings[i, 1], reduced_embeddings[j, 1]],
                   "gray",
                   linestyle="--",
                   alpha=0.3,
               )
               legend_handles.append(line)
   
           legend_labels = [
               f"{text_labels[i]} - {text_labels[j]}: {dist:.2f}"
               for dist, (i, j) in distances_list
           ]
   
           for i in range(len(reduced_embeddings)):
               for j in range(i + 1, len(reduced_embeddings)):
                   plt.plot(
                       [reduced_embeddings[i, 0], reduced_embeddings[j, 0]],
                       [reduced_embeddings[i, 1], reduced_embeddings[j, 1]],
                       "gray",
                       linestyle="--",
                       alpha=0.5,
                   )
   
           plt.legend(
               legend_handles,
               legend_labels,
               title="Distances",
               loc="center left",
               bbox_to_anchor=(1, 0.5),
           )
   
           plt.tight_layout()
           plt.title(
               "2D Visualization of captain responses", fontsize=16, fontweight="bold"
           )
           plt.xlabel("PCA Component 1", fontdict=font_style)
           plt.ylabel("PCA Component 2", fontdict=font_style)
   
           plt.savefig(file_name, bbox_inches="tight")
           plt.close()
   
       chain(
           get_embeddings,
           plot_embeddings(
               get_embeddings.output,
               text_labels=captains_list,
               file_name=IMAGE_PATH,
           ),
       )
   
   
   captains_dag()

   This DAG consists of four tasks to make a simple MLOps pipeline.

   • The get_captains_list task fetches the list of Star Trek captains you want to ask your question to. You'll provide the list of captains when you run the DAG with Airflow params.
   • The ask_a_captain task uses the OpenAIHook to connect to the OpenAI API. It then uses the chat completion endpoint to generate answers to the question you provide. This task is dynamically mapped over the list of captains to generate one dynamically mapped task instance per captain.
   • The get_embeddings task is defined using the OpenAIEmbeddingOperator to generate vector embeddings of the answers generated by the upstream ask_a_captain task. This task is dynamically mapped over the list of answers to retrieve one set of embeddings per answer. This pattern allows for efficient parallelization of the vector embedding generation.
   • The plot_embeddings task takes the embeddings created by the upstream task and performs dimensionality reduction using PCA to plot the embeddings in two dimensions.

   

Step 3: Run your DAG

1. Run astro dev start in your Astro project to start Airflow, then open the Airflow UI at localhost:8080.

2. In the Airflow UI, run the captains_dag DAG by clicking the play button. Then, provide Airflow params for:

   • Question to ask the captain: The question you want to ask the captains.
   • captains_to_ask: A list of Star Trek captains you want to ask the question to. Make sure to create one line per captain and to provide at least two names.
   • max_tokens_answer: The maximum number of tokens available for the answer.
   • randomness_of_answer: The randomness of the answer. The value provided is divided by 10 and given to the temperature parameter of the chat completion endpoint. The scale for the param ranges from 0 to 20, with 0 being the most deterministic and 20 being the most random.

   

3. After the DAG run completed, go to the include folder to view the image file created by the plot_embeddings task. The image should look similar to the one below.

   

Conclusion

Congratulations! You used Airflow and OpenAI to get answers from your favorite Star Trek captains and compare them visually. You can now use Airflow to orchestrate OpenAI operations in your own machine learning pipelines. 🖖