What is data storytelling with LLMs?
Why are LLMs revolutionizing data storytelling?
- Semantic processing: LLMs understand context, recognizing key metrics, cause-and-effect relationships, and business nuances.
- Natural language generation: They can write paragraphs, create slides, and even produce video dialogues with consistency.
- Integrated interpretability: Projects like Motional and MIT have developed models that provide real-time explanations, making the process transparent.
- Integration with modern infrastructure: NVIDIA’s acquisition of Hugging Face for $12.93 billion and Perplexity’s new GPU embedding stack (Ivy, Tulip, ROSE) are making models faster and more cost-effective to use at scale.
How to build a data storytelling pipeline with LLMs
1. Data preparation
Upload your dataset in a structured format (parquet, CSV) and create a schema description. Use optimized embeddings, such as those from Perplexity, to index metadata and speed up semantic searches.
2. Automated data analysis
Use a prompt to ask the LLM to summarize key statistics, identify anomalies, and suggest appropriate visualizations.
You are an expert analyst. DataFrame with columns: date, sales, region.
Generate a 3-point summary of the main trends, highlight any anomalous spikes, and suggest the best chart for each trend.
Return only the text, no extra markup.3. Writing the narrative
Transform the analytical summary into a story tailored for non-technical stakeholders. The prompt should include the audience, tone, and desired outcome.
Write a storytelling paragraph for a business executive audience (formal tone) explaining why sales in Q2 increased by 18% in the EMEA region, highlighting key factors and future implications.4. Creating multimedia content
Ask the LLM to generate slides, chart captions, or video scripts based on the story. Integrating with image or video generation models can further enhance communication.
5. Review and quality control
Use a second LLM to check for factual consistency, appropriate tone, and data completeness. The explainable approach (like Motional/MIT) can highlight which data supports each claim.
Practical example: from dataset to story
Imagine you have a dataset on an e-commerce company’s performance (orders, revenue, conversion rate per campaign). Following the pipeline above, a user could obtain:
- A 3-point analytical summary.
- A 2-paragraph story for the CMO.
- A PowerPoint slide in markdown.
- A 30-second script for a monthly update video.
All of this can be achieved in less than 5 minutes, with the ability to iterate based on feedback.
Key takeaway
Practical tips to get started right away
- Start with a small dataset: A CSV of 100 rows is enough to test the workflow.
- Use structured prompts: Clearly define the audience, tone, and output format.
- Iterate with feedback: Ask the LLM to rephrase the story if it’s too technical or too simplistic.
- Monitor explainability: Verify that each claim is supported by a data reference, leveraging advances in explainable AI like those from Motional/MIT.
- Leverage embedding optimization: Use stacks like Ivy/Tulip/ROSE to reduce the cost of semantic search on larger datasets.