Contextualized output reliability evaluation for language models

A computer system for evaluating reliability of a response generated by a language model comprising:

• one or more processors, and
• a non-transitory program memory coupled to the one or more processors and storing executable instructions that, when executed by the one or more processors, causes the computer system to:
• obtain a query provided to the language model and the response to the query generated by the language model;
• apply a plurality of hallucination evaluation techniques to one or more of the query and the response to generate a plurality of hallucination scores including a different hallucination score for each of the plurality of hallucination evaluation techniques, wherein the plurality of hallucination evaluation techniques include (i) at least one preventative measure that preventatively checks a hallucination risk of the query based on a probability that the query will result in a hallucination, and (ii) at least one detective measure that checks a hallucination risk of the response;
• combine the plurality of hallucination scores to obtain a confidence score;
• and
• provide an indication of the confidence score for display on a user interface, wherein the confidence score indicates the reliability of the response.

A non-transitory computer-readable medium storing executable instructions for evaluating reliability of a response generated by a language model, the instructions, when executed by one or more processors, cause the one or more processors to:

• obtain a query provided to the language model and the response to the query generated by the language model;
• apply a plurality of hallucination evaluation techniques to one or more of the query and the response to generate a plurality of hallucination scores including a different hallucination score for each of the plurality of hallucination evaluation techniques, wherein the plurality of hallucination evaluation techniques include (i) at least one preventative measure that preventatively checks a hallucination risk of the query based on a probability that the query will result in a hallucination, and (ii) at least one detective measure that checks a hallucination risk of the response;
• combine the plurality of hallucination scores to obtain a confidence score;
• and
• provide an indication of the confidence score for display on a user interface, wherein the confidence score indicates the reliability of the response.