Workflow · Production

We Built a News Site Powered by LLMs and Public Data: Here's What We Learned

The problem

As the volume and pace of data grows, it has become a challenge for data journalists to make sense of it all, and there is no scalable way to cover constantly updating datasets like economic indicators, political polls, and environmental data.

Workflow diagram · grounded in source
1
Poll data sources for updates
trigger
“we've built a pipeline to continually ask data sources, "Do you have any new data for me?" — essentially diffing the data we know about with the latest data that the publisher provides”
2
Fetch related news articles
integration
“The system also connects to a newswire to retrieve the latest articles on any topic”
3
LLM generates story text
ai_action
“we use an LLM (currently GPT-4 Turbo from OpenAI) to generate a summary of the updates — the basis for the headlines, subheadings, and descriptive text that you see on our website (Figure 2). To generate this text, we dynamically constru…”
4
LLM markup annotation
ai_action
“The resulting text is then annotated with a simple markup language (via the LLM), which allows us to highlight any references to specific data points or news articles”
5
LLM editing and error check
validation
“we again pass this text back to the LLM and ask it to edit the previously created output, looking in particular for errors that don't follow from our inputs and wording inconsistent with our style guide”
6
Visualizations and story ranking
output
“We then use the resulting text and data to create the visualizations which appear on our website. We compute a ranking of the top stories at any given time by taking into account the relative magnitude and recency of the latest data upda…”
Reported outcome

Realtime automates the creation of data-driven story analyses and visualizations using LLMs, giving readers access to up-to-date information and allowing journalists to focus on in-depth reporting rather than rote data processing.

Reported metrics
LLM pipeline performance from multi-call approachsignificant performance improvements
Reported stack
GPT-4 TurboOpenAIVegaVega-LiteDSPy
Source
https://generative-ai-newsroom.com/we-built-a-news-site-powered-by-llms-and-public-data-heres-what-we-learned-aba6c52a7ee4
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Realtime automates the creation of data-driven story analyses and visualizations using LLMs, giving readers access to up-to-date information and allowing journalists to focus on in-depth reporting rather than rote dat…

What tools did this team use?

GPT-4 Turbo, OpenAI, Vega, Vega-Lite, DSPy.

What results were reported?

LLM pipeline performance from multi-call approach: significant performance improvements (source-reported, not independently verified).

How is this workflow AI workflow structured?

Poll data sources for updates → Fetch related news articles → LLM generates story text → LLM markup annotation → LLM editing and error check → Visualizations and story ranking.