The build-up to the 2024 election was already chaotic when then-Vice President Kamala Harris jumped into the U.S. presidential race on July 21, 2024. Her entry transformed the dynamics of the campaign season, sparking changes in coverage and new strategies in newsrooms across the country.

For the Hearst DevHub — a team of editorial engineers, AI developers, digital designers, and content strategists focused on collaboration with local newsrooms — that meant tearing up part of the roadmap. It committed to quickly building a tool for the San Francisco Chronicle that could help slice through the mis- and disinformation swirling around Harris, a Bay Area native and the first Black and Asian American vice president.

The Kamala Harris News Assistant was published six weeks later. It harnessed SFChronicle’s three decades of reporting — spanning thousands of stories on the former San Francisco district attorney, California attorney general, and U.S. senator — to cut through the noise and provide easier access to the decades of highly edited and vetted political coverage for readers from coast to coast.

Elevating the archive

The assistant’s primary goal was to leverage AI to deliver accurate, in-depth responses to questions about Harris and her campaign — information grounded in the high-quality work of SFChronicle reporters. This project aimed to make the newsroom’s extensive archive of political reporting more accessible and discoverable to core readers, while increasing engagement from non-subscribers.

To achieve this, the Kamala Harris News Assistant integrates these features:

• Comprehensive guardrails: Before the assistant responds to queries from readers, submitted questions go through a series of guardrail checks in parallel: looking for harmful content such as self-harm, hate speech or violence; assessing whether a question pertained specifically to Harris, her presidential campaign, or the 2024 U.S. presidential election; and checking whether the question is about voting, in which case the assistant redirected readers to an election information source.
• Answers grounded in SFChronicle reporting: After queries pass our guardrails, they are directed into a retrieval augmented generation (RAG) pipeline. This means the query is posed to a vector database containing SFChronicle’s nearly 30 years of reporting, where it searches for relevant chunks of information to answer questions. These chunks of information are then passed into a staff-written prompt that is sent to OpenAI’s GPT-4o Mini model.

Ethical AI practices are at the core of the Kamala Harris News Assistant. Users are provided with transparent answers, and the bot’s development process ensures human oversight to verify the accuracy and relevance of its responses.

By integrating AI to complement — not replace — human creativity, the news assistant represents Hearst’s commitment to responsible innovation in journalism.

How it worked

Most importantly, the news assistant’s responses were 100% accurate.

Readers were provided with trustworthy information rooted in vetted reporting, which helped them better understand a complex news cycle and the personal history of a leader who rose to great heights from her Bay Area beginnings.

The integration of historical and real-time data broadened the scope of political reporting. Readers gained easy access to context-rich insights spanning three decades, while journalists benefited from seeing the gaps between daily reporting assignments and readers’ top questions during the campaign.

With a goal of answering 50% of reader questions — including the expected spam and troll queries — the tool surpassed expectations, achieving a response rate of more than 75%, which proved many of the prompt-engineering decisions were correct.

In the end, the Kamala Harris News Assistant provided thousands of readers with accurate, fact-based information about a candidate and shone a light on how local news is the foundation for a high-quality information ecosystem.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .