We have built Jarbiz as a Retrieval Augmented Generation (RAG) application to reduce hallucinations to near zero, and improve the accuracy and relevancy of Jarbiz's responses. In simple terms, the RAG app uses a foundational large-language model (LLM – for example GPT-4o) to understand the intent of the user’s query and plan how to answer it. Jarbiz will then find relevant data to ‘ground’ its response from high-quality datasets including premium financial and firmographic data and proprietary data such as strategic priorities and management intent. The LLM will then synthesize the data into a response for the user.

Relating to the proprietary and third-party datasets that are used to ‘ground’ Jarbiz's responses, Jarbiz has measures in place to ensure high data quality when ingesting and processing data. For datasets such as company strategic priorities or financials we are able to trace their origins back to authoritative sources such as company investor documents, earnings transcripts or company filings. Key datasets are refreshed daily. Jarbiz and its third-party data suppliers also have automated data checks to verify the accuracy and recency of the data used to inform Jarbiz's responses. Jarbiz also has its own program of quality assurance checks by human analysts to provide additional reassurance.

This approach is different from conversational AI apps such as ChatGPT or Perplexity that rely on general web searches to answer some queries, especially those that relate to recent developments after the LLM was trained and so the answers won’t be found in the LLM’s training data. For example, for queries relating to company strategic priorities these conversational AI apps will look at sources beyond the company’s official documents (for example, media reports or SWOT analyses by MBA students), thus increasing the chance of inaccurate data being included in a response.