Conventional rule-based dialog systems are suitable only for a narrow set of scenarios and types of questions, and are not easy to extend to new domains and/or higher complexity questions. Traditional machine learning models require substantial data to train and retraining can affect previously released functionality. This disclosure describes an LLM augmented with an index of examples and a retrieval system for use in dialog systems. The techniques leverage the instruction-following capabilities of LLMs. Upon receipt of a user query, the conversation context and a set of examples retrieved from an index are utilized to generate a prompt for the LLM. The LLM response is an API call specification to a backend system. The response from the backend system, which can include data responsive to the user query, along with a different set of instructions is provided to the same LLM. The LLM response is provided to the user. The set of examples in the index can be updated independent of the rest of the system. The described techniques leverage LLM capabilities to implement a dialog system that can learn from examples and can easily be extended to new domains.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.