Introduction to AI Search and Assistants
Google researchers have introduced a new method to improve AI search and assistants by enhancing Retrieval-Augmented Generation (RAG) models. This method helps RAG models recognize when retrieved information lacks sufficient context to answer a query, which can lead to more reliable and accurate AI-generated responses.
The Problem with Current RAG Models
Current RAG models, such as Gemini and GPT, often attempt to answer questions even when the retrieved data contains insufficient context. This can result in hallucinations, or incorrect answers, instead of abstaining from answering. The researchers found that these models can provide correct answers when given sufficient context, but they also answer correctly 35-65% of the time even when the context is insufficient.
Defining Sufficient Context
The researchers define sufficient context as meaning that the retrieved information contains all the necessary details to derive a correct answer. This classification does not require the answer to be verified, but rather assesses whether the retrieved information provides a reasonable foundation for answering the query. Insufficient context, on the other hand, means that the retrieved information is incomplete, misleading, or missing critical details needed to construct an answer.
Sufficient Context Autorater
The Sufficient Context Autorater is an LLM-based system that classifies query-context pairs as having sufficient or insufficient context. The best performing autorater model, Gemini 1.5 Pro (1-shot), achieved a 93% accuracy rate, outperforming other models and methods.
Reducing Hallucinations with Selective Generation
The researchers discovered that RAG-based LLM responses were able to correctly answer questions 35-62% of the time when the retrieved data had insufficient context. They used this discovery to create a Selective Generation method that uses confidence scores and sufficient context signals to decide when to generate an answer and when to abstain. This achieves a balance between allowing the LLM to answer a question when there’s a strong certainty it is correct and abstaining when there’s insufficient context.
How Selective Generation Works
The researchers describe how Selective Generation works: "…we use these signals to train a simple linear model to predict hallucinations, and then use it to set coverage-accuracy trade-off thresholds. This mechanism differs from other strategies for improving abstention in two key ways. First, because it operates independently from generation, it mitigates unintended downstream effects…Second, it offers a controllable mechanism for tuning abstention, which allows for different operating settings in differing applications, such as strict accuracy compliance in medical domains or maximal coverage on creative generation tasks."
Takeaways
The research paper does not state that AI will always prioritize well-structured pages, but rather that context sufficiency is one factor that influences AI-generated responses. Confidence scores also play a role in intervening with abstention decisions. Pages with complete and well-structured information are more likely to contain sufficient context, but other factors such as how well the AI selects and ranks relevant information also play a role.
Characteristics of Pages with Insufficient Context
Pages with insufficient context may be lacking enough details to answer a query, misleading, incomplete, contradictory, or require prior knowledge. The necessary information to make the answer complete may be scattered across different sections instead of presented in a unified response.
Relation to Google’s Quality Raters Guidelines
Google’s Quality Raters Guidelines (QRG) has concepts that are similar to context sufficiency. For example, the QRG defines low-quality pages as those that don’t achieve their purpose well because they fail to provide necessary background, details, or relevant information for the topic. The guidelines also describe low-quality pages as those with a large amount of off-topic and unhelpful content, or those with a large amount of "filler" or meaningless content.
Conclusion
The research paper introduces a new method to improve AI search and assistants by enhancing RAG models’ ability to recognize when retrieved information lacks sufficient context. This method can lead to more reliable and accurate AI-generated responses. While the paper does not state that AI will always prioritize well-structured pages, it highlights the importance of context sufficiency in AI-generated responses. By understanding the characteristics of pages with insufficient context and the relation to Google’s Quality Raters Guidelines, publishers can create content that is more useful for AI-generated answers.