American Institute of Physics – Conversational search

The rapidly evolving field of Generative AI prompted the American Institute of Physics to delve into potential applications for their content catalogue. Their goal was to enhance the accessibility and consumption of relevant research through AI. To translate this into a real world experience for users, our approach involved developing a conversational search tool. The aim was to foster a more intuitive content discovery process, boosting both subscription and single article purchases.

Teaming up with Matheson.ai, we developed a chat-based solution to semantically query the AIP content catalogue. The objective was to provide search results accompanied by generative summaries, underpinned by rigorous citation evidence. This innovative approach involved harnessing the capabilities of a Language Model (LLM) for generating embeddings, utilising a vector database for nearest neighbour matching and leveraging OpenAI's completions API to construct a Retrieve-Augmented Generation framework. The user interface, crafted with Next.js, boasts an elegant design that supports conversation histories, sharing features and feedback mechanisms. These components contribute to training the application for enhanced accuracy.