PBS used Amazon Bedrock to incorporate generative AI into its search and viewing infrastructure, deliver a better user experience and reveal different parts of their catalogue.
AWS Bedrock generative AI application architecture
Public Broadcasting Service (PBS), the long-standing network behind favourite US programs like Sesame Street and Frontline, recently undertook a project to enhance search results returned to their viewers using the PBS App and PBS LearningMedia platforms. They wanted to upgrade the way audiences discover and consume content across their streaming platforms, and connect with viewers through a more personalized viewing experience.
Also recognising the value of using generative AI for such tasks, PBS worked with science and strategy experts from the AWS Generative AI Innovation Centre to consider ways to incorporate generative AI into its infrastructure. Together, these two teams developed and launched a new search engine that runs on AWS’ generative AI service Amazon Bedrock, which connects developers to foundation models, services to deploy and operate agents, and tools for fine-tuning, safeguarding and optimising models
Building Better Searching
Working backwards from their problem, PBS first envisioned the updated search engine, trained and tested models, and then catalogued and tagged metadata for more than 700,000 assets. The full process took less than six months, and as a result, PBS viewers can now search the entire PBS library based on their interests cutting across subgenres, topics, themes and even moods.
Metadata was the key. "Had we tagged the metadata manually for this project, it would have taken years," estimated Mikey Centrella, Director of Product at PBS, Digital Innovation. "With Amazon Bedrock, we were able to finish the job quickly, at a very low cost and within the secure AWS environment we already have. Within a few months, we had built a proof-of-concept, tested and successfully scaled it into production. By improving our search tools, we can deliver a better experience for viewers while showing them different parts of our catalogue they might not otherwise have found."
PBS’ relationship with AWS began more than a decade ago. Since then, the organization has employed AWS products and services in nearly every aspect of its operations, mainly to help it scale cost-efficiently but also to trial and take up new technology. Always seeking new innovation, Mikey and his team evaluate emerging technologies and identify opportunities where PBS can enhance or grow its services.
Experimental Approach to Modernisation
"One of the advantages of building on AWS is that we can explore ways to modernise our operations without impacting what's already in production," said Mikey. "Our assets are already in the cloud, so it's a safe way to experiment, figure out what we might want to invest in and then apply it across the business. As a nonprofit, we must be smart about our choices and use our dollars efficiently."
PBS App
In this case, he and his team wanted to create a solution with generative AI. At the same time, their intention was a result that would add value and have an immediate impact for viewers. Starting out with the AWS Generative AI Innovation Centre in late 2024, the team began to explore how best to augment their existing search engine – originally built on AWS and cloud-native tools from PBS – using generative AI.
"The essential requirements for a streaming services to be competitive, have changed because of technology," noted Mikey. "Our users wanted a more robust search engine, so that's what we created."
Innovation Project
After defining the project scope, PBS began an eight-week engagement with the Innovation Centre. The Centre’s mission is to accelerate the adoption of generative AI, accomplished through bespoke, generative AI proof-of-concepts designed in collaboration with customers, such as PBS.
Mike's team began building the new search engine framework and testing different models. Although the existing search engine implementation had basic logic functionality, and could pull title or rudimentary keyword matches, the team integrated generative AI for extensive metadata tagging. Consequently, the new search engine can produce much more specific results.
Each area of search classification encompasses hundreds of options and dozens of metadata tags that can be applied to any given project. For example, a viewer can search for 'heartwarming' in the PBS App and receive a result such as All Creatures Great and Small, a charming series about a veterinarian in 1930s Northern England.
"Generative AI is very useful for pattern recognition and summarisation. Metadata – when set up correctly – serves as a solid foundation to use it successfully," Mike said. "For analysing content, we found that the show descriptions and transcripts gave us the most value for the price. We narrowed down four areas of classification internally, then used models on Amazon Bedrock to analyze the content and apply the most appropriate metadata."
Model Performance and Selection
AWS and PBS established a sandbox environment to evaluate the model performance collaboratively using Amazon Bedrock’s services. Amazon Bedrock gives access to hundreds of foundation models from mainstream AI companies – like Anthropic, Mistral AI, AI21 Labs as well as Meta and Amazon itself – along with evaluation tools to pick the best model based on users’ performance and cost needs. In this way, an AI strategy will remain relevant as the Club’s needs evolve and new models emerge.
PBS LearningMedia platform
Ultimately, the Anthropic Claude Sonnet model was determined to be the best fit. Anthropic’s Claude models were developed to build AI agents that can analyse thousands of data sources, execute long-running tasks and write content and other complex actions. The Sonnet version can autonomously plan and execute complex, multi-step workflows, making it effective for research and collaborative, sustained reasoning.
The PBS team then ran parallel workstreams on 50 shows, while PBS colleagues manually created metadata tags based on show scripts to compare with the results created using generative AI. Looking at the outcomes, they found that the model was now capable of producing data sets as good as or better than the ones that were created manually, and very quickly.
Based on those results, PBS moved the model into operation, generating metadata tags for its entire content library of 700,000 assets and incorporating this new metadata into its search engine.
Further Generative AI Implementations
With the updated search engine launched, the PBS Innovation Team is now working again with the AWS Generative AI Innovation Centre to evaluate other ways to incorporate generative AI into its infrastructure. The most likely next step is to extend enhanced search recommendations to the network's kid's catalogue.
"PBS understands that the audience and their behaviours are changing, across all the available streaming options, gaming platforms and mobile devices. We live in a multiplatform, multimedia world, and we want to meet people where they are," said Mike. "It can be challenging to address all the new technology and distribution types that seem to emerge daily, and at the same time, linear still matters. AWS makes various ways available to us to evolve and experiment, while maintaining our core principles." aws.amazon.com
