Local Zoom Assistant: Experience with NLI Model Integration

The author continues a series of articles on developing a local assistant for video conferencing. The tenth installment examines the practical experience of integrating Natural Language Inference (NLI) models to analyze meeting context. The core issue addressed is the inability of standard embeddings to distinguish between phrases with opposite meanings, such as 'budget approved' and 'budget not approved.' The article details three months of operational experience with an NLI model in the 'Charoit' project, highlighting where it proved highly effective and where it was abandoned due to technical limitations or inefficiency. The author analyzes the costs of implementing such solutions and shares insights on how to properly integrate logical inference into speech processing systems to improve the accuracy of understanding meeting participants' intentions.