Launching a High-Accuracy Chatbot Using Generative AI Solutions on AWS with Megamedia

As one of the largest broadcasters in Chile, Megamedia has websites as well as radio, television, and cable channels in the region. The company’s news website, meganoticias.cl, hosted a tool to help members of the public get information about government support and benefits, such as electricity and housing subsidies, scholarships, bonuses, automatic family subsidies, unemployment insurance, and more. People could find information gathered from various government sites—such as what benefits exist and how to apply for them—in one place on meganoticias.cl, under “Dato útil.” However, Megamedia believed it could find a more user-friendly way to deliver the information.

In early 2024, Megamedia decided to use generative AI to build a chatbot that could help people answer their questions about government benefits programs. “We were looking for a way for people to not only read articles about these benefits but also to make personalized queries about their situation,” says Claudio Arce, general editor of meganoticias.cl for Megamedia. With a chatbot, for example, an unemployed parent looking for information about both childcare and unemployment benefits could ask questions about their specific situation and get personalized responses. The company began building the application using Generative AI Application Builder on AWS, a solution from the AWS Solutions Library that facilitates the development, rapid experimentation, and deployment of generative AI applications without requiring deep experience in AI.

Megamedia also began working with ARKHO, one of the first AWS Partners in Chile, which was already using generative AI on AWS and brought its expertise to the project. Together, Megamedia and ARKHO refined the model accuracy for the chatbot and migrated the model to Megamedia’s website. The Dato Útil chatbot went live on meganoticias.cl, officially launching at the beginning of May 2024, just 4 months after the beginning of the project.

Using the Dato Útil chatbot, people can now ask questions about public programs in Chile and receive information specific to their situation in around 30 seconds. Previously, they could spend hours searching through government sites themselves. “This project is going to provide value to people who don’t have all the tools and access to get the answers otherwise,” says Nicol Campusano, marketing coordinator at ARKHO. The team at Megamedia also worked to make sure that the user experience functioned well on mobile devices and in any browser.

Because it is powered by large language models (LLMs), interacting with the Dato Útil chatbot feels like a real conversation, a much better user experience than chatbots with only static, preloaded answers. The chatbot is built using Claude 2, an LLM from Anthropic, through Amazon Bedrock, a fully managed service that offers a choice of high-performing foundation models from well-known AI companies. ARKHO provided expertise in Amazon Bedrock, which is integrated within Generative AI Application Builder on AWS. “We use Amazon Bedrock for basically all our generative AI use cases because it is very flexible,” says Nicolas Caraball, data scientist at ARKHO. “It gives us access to a lot of models, including newly released ones, and it is pay-per-use, so we can save pretty significant costs by not hosting the models ourselves.”

One of the most important requirements for the Dato Útil chatbot was that it provide accurate information. Megamedia set its bar for success at a minimum of 80 percent accuracy. “As a news site, we have an editorial weight that is different from another type of website,” says Daniela Cartagena, director of digital strategies at Megamedia. “Credibility is key, so what we present needs to be true.” Megamedia worked with ARKHO to improve the accuracy of the solution by using high-quality prompts and data sources to train the model. To integrate the data into the model, ARKHO used Amazon Kendra, an intelligent enterprise search service, to extract text from the source documents about government benefits programs and create embeddings for vector searches. The Dato Útil chatbot was then trained in the specific context of government benefits, and the responses were refined together with a panel of journalists. The model now has an over 90 percent accuracy rate.

Megamedia was excited to innovate and deliver this tool to the public. It also worked to make sure that its journalists, especially those who were skeptical about the technology, were brought on board to help shape the tool and to learn how it could also benefit them. For example, journalists can use the chatbot to search through government source information to create new articles faster and more efficiently. “There is a lot of fear about AI and how it can be bad for a pressroom, so we wanted to show the team internally that it can be a solution rather than a threat,” says Arce.