In this presentation, Sid Probstein, CEO of SWIRL, began by highlighting his background in search and natural language processing. He mentioned his experiences at Fast Search and Transfer, now part of Microsoft, and his founding of Attibio, now known as ServiceNow. In describing how he started his career at John Hancock Financial Services, he noted the challenges posed by data silos, and the importance of efficient technology use, and started to explore the uses of Retrieval Augmented Generation (RAG).
Probstein emphasized the need for simple answers in business intelligence, comparing it to the vision of the ‘Star Trek’ computer. He acknowledged the skepticism around generative AI, but explained why he remains optimistic about its transformative potential, especially after the release of GPT-3. He shared his experience of integrating three search systems, and the frustration that led him to create SWIRL, an open-source metasearch engine. SWIRL aggregates and re-ranks search results from various endpoints, addressing the challenge of comparing different types of data.
In explaining the utility of the technology, Probstein explained that SWIRL leverages cloud-based indices and systems, using large language models to re-rank results. This approach has evolved to enable secure conversations, with internal data, within enterprises, a crucial aspect of AI's promise. Probstein illustrated this with an example of finding a cyber insurance policy using SWIRL, demonstrating its ability to locate, summarize, and provide links to relevant documents quickly. He highlighted the idea that Swirl's current functionality allows for omni-channel conversations with data, whether through a custom UI, Microsoft Teams, or other interfaces. He also emphasized that SWIRL acts as an API engine, facilitating secure and comprehensive interactions with enterprise data. He provided an example of a business process where SWIRL helped summarize and locate insurance policy details, saving significant time compared to manual searches.
He discussed SWIRL's capability to handle both structured and unstructured data, showing examples of rendering charts from a funding database in Google BigQuery and summarizing complex PDF documents. Sid noted that SWIRL operates on full-text searches, and can integrate with various repositories without bulk data transfers or indexing, maintaining security and efficiency.
Probstein later elaborated on SWIRL's architecture, which consists of two main components: the SWIRL Copilot and SWIRL AI Connect. The Copilot transforms an LLM into a conversational agent that understands user context and search terms, while AI Connect handles the meta-search, re-ranking results, and maintaining conversation context. He explained that SWIRL uses a reader LLM to analyze text precisely without altering it, ensuring accurate and relevant search results.
The speaker concluded by demonstrating SWIRL's effectiveness in providing more accurate search results compared to traditional search engines, highlighting its ability to understand word order and context. He invited feedback and further questions, ready to delve deeper into SWIRL's capabilities and potential applications.
GAI Insights Team
