Improve AI search and assistants using Retrieval-Augmedated Generation models

Google researchers introduce a method that improves AI search and assistants, improving the ability of RAG models to determine when the information obtained does not have a sufficient context for response. This helps to avoid AI-response on incomplete information and improves the reliability of answers.

The study shows that models like Gemini and GPT often try to answer questions when the data obtained contains insufficient context, which leads to hallucinations instead of abstaining. To solve this problem, they have developed a system that reduces hallucinations by helping LLM to determine when the resulting content contains enough information to maintain the answer.

A sufficient context is defined as information that contains all the necessary details to output the correct answer. Classification that something contains a sufficient context does not require that it is a verified answer. It only evaluates whether the answer can be deduced from the content provided.

• 📌 Models like Gemini, GPT, and Claude, provide the correct answers when they have a sufficient context.
• 📌 If the context is insufficient, they sometimes hallucus instead of abstaining from the answer, but they also correspond to 35-65% of the time correctly.
• 📌 Researchers have developed a system of reducing hallucinations by helping LLM when the resulting content contains enough information to maintain the answer.