With over 100,000 businesses relying on Checkr for monthly personnel background checks, the use of generative AI (genAI) and machine learning tools is essential to navigate through vast amounts of unstructured data.
Through an automated process, each potential job candidate undergoes a thorough background check that analyzes information from various sources to identify any criminal or other relevant issues.
About 2% of Checkr’s data is considered “messy,” requiring the adoption of genAI tools like OpenAI’s GPT-4 large language model (LLM) to handle such records efficiently.
Despite GPT-4 achieving an 88% accuracy rate it dropped to 82% when dealing with messy data, falling short of customer standards.
To address this challenge, Checkr integrated retrieval augmented generation (RAG) into its LLM system to enhance accuracy. While this improved accuracy rates for most records to 96%, it decreased significantly for more complex data sets down to just 79%.
In addition to accuracy concerns, both the standard GPT-4 model and the RAG-enhanced version faced slow response times during background checks – taking up to 15 and seven seconds respectively.
2024-10-11 03:15:03
Article from www.computerworld.com
