Speaker: Miguel Ángel Martínez Pay.
Abstract:
This seminar outlines the process of creating a guardrail for a banking transactions assistant. The guardrail acts as a security system that filters user queries, determining which can be processed by the assistant and which cannot, ensuring the overall safety of the system.
To build it, a strategy was designed to generate synthetic datasets using ChatGPT, organized and segmented by specific banking-related contexts. The resulting dataset was used to train a sequence classification model based on RoBERTa, capable of identifying whether a query falls within the allowed domain.
Additionally, multi-topic phrases were generated through automated combinations using ChatGPT, enabling the training of a topic shift detector. This detector is based on a sliding window classification technique applied to the embeddings obtained from RoBERTa’s encoder, allowing the system to detect thematic deviations within a single query or ongoing conversation.