zenfiles/zen-news-summarization

📜 ZenNews: Generate summarized news on a schedule

In today's information age, we are bombarded with a constant stream of news and media from a variety of sources. Summarizing tasks, particularly when it comes to news sources, can be a powerful tool for the efficient consumption of information. They distill complex or lengthy content into easily digestible chunks that can be scanned and absorbed quickly, allowing us to keep up with the news without being overwhelmed. They can also help us separate the signal from the noise, highlighting the most important details and helping us identify what's worth further investigation.

This is where ZenNews come into play. It offers a tool that can automate the summarization process and save users time and effort while providing them with the information they need. This can be particularly valuable for busy professionals or anyone who wants to keep up with the news but doesn't have the time to read every article in full.

🎯 The goal of the project

The definition of the concrete use case aside, this project aims to showcase some of the advantages that ZenML brings to the table. Some major points we would like to highlight include:

• The ease of use: ZenML features a simple and clean Python SDK. As you can observe in this example, it is not only used to define your steps and pipelines but also to access/manage the resources and artifacts that you interact with along the way. This makes it significantly easier for you to build their applications around ZenML.

• The extensibility: ZenML is an extensible framework. ML projects often require custom-tailored solutions and what you get out of the box may not be what you need. This is why ZenML is using base abstractions to allow you to create your own solutions without reinventing the whole wheel. You can find great examples of this feature by taking a look at the custom materializer (ArticleMaterializer) and the custom stack component (DiscordAlerter) implemented within the context of this project.

• Stack vs Code: One of the main features of ZenML is rooted within the concept of our stacks. As you follow this example, you will see that there is a complete separation between the code and the infrastructure that the pipeline is running on. In fact, by utilizing just one flag, it is possible to switch from a local default stack to a remote deployment with scheduled pipelines.

• The scalability: This is a small PoC-like example that aims to prove that ZenML can help you streamline your workflows and accelerate your development process. However, it barely scratches the surface of how you can improve it even further. For more information, check this section.

🐍 Base installation

The ZenNews project is designed as a PyPI package that you can install through pip:

pip install zennews

The package comes equipped with the following set of key pieces:

• The pipeline: The zen_news_pipeline is the main pipeline in this workflow. In total, it features three steps, namely collect, summarize and report. The first step is responsible for collecting articles, the second step summarizes them, and the last step creates a report and posts it.

• The steps: There is a concrete implementation for each step defined above.

  • For the collect step, we have the bbc_news_source which (by default) collects the top stories from the BBC news feed and prepares Article objects.
  • For the summarize step, we have implemented bart_large_cnn_samsum step. As the name suggests, it uses the BART model to generate summaries.
  • Ultimately, for the report step, we have implemented the post_summaries step. It showcases how a generalized step can function within a ZenML pipeline and uses an alerter to share the results.

• The materializer: As mentioned above, the steps within our pipeline are using the concept of Articles to define their input and output space. Using the ArticleMaterializer, we can show how to handle the materialization of these artifacts when it comes to a data type that is not built-in.

• The custom stack component: The ultimate goal of ZenNews is to serve the use the direct outcomes of the pipeline. That is why we have used it as a chance to show the extensibility of ZenML in terms of the stack components and implemented a DiscordAlerter.

• The CLI application: The example also includes a Click CLI application. It utilizes how easily you can use our Python SDK to build your application around your ZenML workflows. In order to see it in action simply execute:

  zennews --help 
  

🕹 Test it locally right away

After installing the zennews package, you are ready to test it out locally right away. The following command will get the top five articles from the BBC news feed, summarize them and present the results to you.

Warning: This will temporarily switch your active ZenML stack to the default stack and when the pipeline runs, you will download the model to your local machine.

zennews bbc

You can also parameterize this process. In order to see the possible parameters, please use:

zennews bbc --help

🚀 Switching to scheduled pipelines with Vertex

The potential of an application like zennews can be only unlocked by scheduling summarization pipelines instead of manually triggering them yourself. In order to showcase it, we will set up a fully remote GCP stack and use the VertexOrchestrator to schedule the pipeline.

Deploy ZenML on GCP

Before you start building the stack, you need to deploy ZenML on GCP. For more information on how you can achieve do that, please check the corresponding docs page.

ZenNews Stack

Once the ZenML is deployed, we can start to build up our stack. Our stack will consist of the following components:

• GCP Container Registry
• GCS Artifact Store
• Vertex Orchestrator
• Discord Alerter (part of the zennews package)

Let's start by installing the gcp integration:

zenml integration install gcp

Container Registry

The first component is a GCP container registry. Similar to the previous component, you just need to provide a name and the URI to your container registry on GCP.

zenml container-registry register <CONTAINER_REGISTRY_NAME> \
    --flavor=gcp \
    --uri=<REGISTRY_URI>

Artifact Store

The next component on the list is a GCS artifact store. In order to register it, all you have to do is to provide the path to your GCS bucket:

zenml artifact-store register <ARTIFACT_STORE_NAME> \
    --flavor=gcp \
    --path=<PATH_TO_BUCKET> 

Orchestrator

Following the artifact store, we will register a Vertex AI orchestrator.

zenml orchestrator register <ORCHESTRATOR_NAME> \
    --flavor=vertex \
    --project=<PROJECT_ID> \
    --location=<GCP_LOCATION> \
    --workload_service_account=<EMAIL_OF_YOUR_SERVICE_ACCOUNT> \
    --service_account_path=<PATH_TO_YOUR_SERVICE_ACCOUNT_KEY>

You need to simply provide the id of your project, the name of your GCP region and the service account you would like to use.

Warning: In this version, you have to provide both the email of the service account and the path to a key.json file. This interaction will be improved in the upcoming releases.

Make sure that the service account has the proper roles for the following services: Cloud Functions, Cloud Scheduler, Service Account, Storage, and Vertex AI.

GCP Stack

With these four components, we are ready to establish and activate the base version of our GCP stack.

zenml stack register <STACK_NAME> \
    -c <CONTAINER_REGISTRY_NAME> \
    -a <ARTIFACT_STORE_NAME> \
    -o <ORCHESTRATOR_NAME> \
    --set

Alerter

The last component in our stack is a special case. As mentioned before the zennews package already comes equipped with a custom stack component implementation, namely the DiscordAlerter. In a nutshell, it uses the discord.py package to send messages via a webhook to a discord text channel. You can find the implementation right here.

The following sections show how we can register DiscordAlerter as a custom flavor, create an instance of it, and update our stack.

Registering the custom flavor

All you have to do to register such a custom flavor is to provide the corresponding source path to the flavor class.

zenml alerter flavor register zennews.alerter.discord_alerter_flavor.DiscordAlerterFlavor

ZenML will import and add that to the list of available alerter flavors.

zenml alerter flavor list

Registering the alerter

Now that the flavor is registered, you can create an alerter with the flavor discord-webhook. Through this example, you will also see how you can use secret references to handle sensitive information during the registration of stack components.

Let's start by creating the secret:

zenml secret create <SECRET_NAME> \
    --webhook_url=<ACTUAL_URL_OF_THE_WEBHOOK>

This will store the secret values in the ZenML server. Once the secret creation is complete, you can register your alerter as follows:

zenml alerter register <ALERTER_NAME> \
    --flavor discord-webhook \
    --webhook_url=<SECRET_REFERENCE>  # formatted as {{SECRET_NAME:WEBHOOK_URL}}

Updating the stack

The last step is to update our stack with our new alerter:

zenml stack update <STACK_NAME> -al <ALERTER_NAME>

Scheduling pipelines through the zennews CLI

Now the stack is set up, you can use the --schedule option when you run your zennews pipeline. There are three possible values that you can use for the schedule option: hourly, daily (every day at 9 AM), or weekly (every Monday at 9 AM).

zennews bbc --schedule daily

This will use your active stack (the GCP stack) and schedule your ZenNews pipeline.

📙 Limitations, future improvements and upcoming changes

Before we end this project, it is also important to talk about the limitations we faced, the possible future improvements, and changes that are already in motion:

• The first limitation of ZenNews is the number of supported news sources. As this project was initially designed as a PoC, the only supported news source is BBC. However, thanks to our design, it is really easy to expand this list by adding additional steps, which consume data and create Article objects.
• The ability to schedule pipelines through ZenML played a critical role within the context of this project. However, this feature has its own limitations. While you can create scheduled pipelines, once the pipeline and its schedule is created, you can not cancel or modify the behavior of this scheduled pipeline. This means that if you want to cancel it, you have to do it via the orchestrator UI or interface yourself and not from within ZenML.
• The other limitation regarding the schedules is the format. As of now, the CLI application takes the user input and converts it into a cron expression. Any orchestrator which does not support these expressions is not supported.
• As the ZenML team, we have been looking for ways to improve the interface of our base alerters. You might see some changes in upcoming releases.
• Similar to the alerters, we are working on improving the management of our secrets.

Tune in to our slack to stay updated about the upcoming changes and ask any questions you might have.