Generative AI diagrams
Whimsical
Recognizing the potential of generative AI, Whimsical organized a hackweek to explore its applications. I proposed a concept called Lingua Franca, based on a clear hypothesis:
If we could find a common language between us and the LLM, then we could generate diagrams from text.
We identified Mermaid.js as the ideal diagramming DSL after careful evaluation. Its popularity in the open-source community made it likely that GPT models were trained on its syntax. Through prompt engineering, we persuaded GPT to produce responses in this format.
Our next challenge was developing a robust parser to interpret these responses. Given Mermaid.js’s extensive feature set, we prioritized the most impactful diagram types: flowcharts, mindmaps, sticky notes, and sequence diagrams. We refined our prompts to steer the LLM towards supported features, optimizing reliability.
This parser development also unlocked a valuable secondary feature: the ability for users to import existing Mermaid diagrams for enhanced editing and styling. Who doesn’t love a two-for-one deal? ✊🏽
Finding a new distribution channel
In tandem, we built a ChatGPT plugin. This allowed users to generate diagrams within ChatGPT’s interface. As the first text-to-diagram plugin in the ChatGPT marketplace, we reached audiences unfamiliar with Whimsical.
Anticipating the occasional inaccuracies of AI-generated content, we designed a smooth transfer experience. Users can easily open the diagram in Whimsical for refinement. This involved creating our first-ever account-free sandbox environment. Here, users can edit their diagrams, then save them by creating an account.
This flow led to a dramatic uptick in our acquisition funnel. In the process, we gained valuable insight into their visualization needs.