Chat-based data visualization solution for Mediatool

Mediatool developed a seamless experience in extracting and summarizing data through natural language queries in a chat interface that retrieves needed information and visualizes it directly in the interface. Modulai acted as a technical advisor and assessed the details of the methodology.

Background
Mediatool’s users faced challenges in efficiently locating data in the vast datasets available for their marketing campaigns and summarizing it. The previous system required users to either navigate through existing reports to find relevant summaries or create new ones manually. Mediatool sought to streamline this process, aiming to offer data visualization as immediate answers to users’ natural language queries, such as “How much do we plan to spend on social media by product category” and “How many impressions did we have last week in social media?”.

Solution
Mediatool initiated a project utilizing OpenAI’s chat models through few-shot learning to generate complex JSON structures for automatic data retrieval and visualization, based on input and feasibility study by Modulai. This allows customers to chat with an AI assistant in a user-friendly way to retrieve data for decision-making in seconds. Chat interface removes technological barriers for customers who can now use natural language to communicate with the platform.

 

Tools/Tech
They utilized OpenAI’s GPT-4 and GPT-3.5 language models and LangChain for quick prototyping. GPT-4 proved to be surprisingly good at in-context learning from a few examples. Its JSON-structured output can be passed to Mediatool’s data summarization API with minimal to no post-processing, thus streamlining the natural language to the data visualization pipeline. And since August 2023, fine-tuning of GPT-3.5 allows an even more flexible approach to tweaking the model’s input/output schema allowing it to continue improving both accuracy and latency.