Google researchers introduced a method to improve AI search and assistants by enhancing Retrieval-Augmented Generation (RAG) models’ ability to recognize when retrieved information lacks sufficient context to answer a query. If implemented, these findings could help AI-generated responses avoid relying on incomplete information and improve answer reliability. This shift may also encourage publishers to create content with sufficient context, making their pages more useful for AI-generated answers.
Their research finds that models like Gemini and GPT often attempt to answer questions when retrieved data contains insufficient context, leading to hallucinations instead of abstaining. To address this, they developed a system to reduce hallucinations by helping LLMs determine when retrieved content contains enough information to support an answer.
Retrieval-Augmented Generation (RAG) systems augment LLMs with external context to improve question-answering accuracy, but hallucinations still occur. It wasn’t clearly understood whether these hallucinations stemmed from LLM misinterpretation or from insufficient retrieved context. The research paper introduces the concept of sufficient context and describes a method for determining when enough information is available to answer a question.
Their analysis found that proprietary models like Gemini, GPT, and Claude tend to provide correct answers when given sufficient context. However, when context is insufficient, they sometimes hallucinate instead of abstaining, but they also answer correctly 35–65% of the time. That last discovery adds another challenge: knowing when to intervene to force abstention (to not answer) and when to trust the model to get it right.
Defining Sufficient Context
The researchers define sufficient context as meaning that the retrieved information (from RAG) contains all the necessary details to derive a correct answer. The classification that something contains sufficient context doesn’t require it to be a verified answer. It’s only assessing whether an answer can be plausibly derived from the provided content.
This means that the classification is not verifying correctness. It’s evaluating whether the retrieved information provides a reasonable foundation for answering the query.
Insufficient context means 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 was Gemini 1.5 Pro (1-shot), achieving 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. That meant that sufficient context wasn’t always necessary for improving accuracy because the models were able to return the right answer without it 35-62% of the time.
They used their discovery about this behavior to create a Selective Generation method that uses confidence scores (self-rated probabilities that the answer might be correct) and sufficient context signals to decide when to generate an answer and when to abstain (to avoid making incorrect statements and hallucinating). This achieves a balance between allowing the LLM to answer a question when there’s a strong certainty it is correct while also allowing for abstention when there’s sufficient or insufficient context for answering a question.
The researchers describe how it 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
Before anyone starts claiming that context sufficiency is a ranking factor, it’s important to note that the research paper does not state that AI will always prioritize well-structured pages. Context sufficiency is one factor, but with this specific method, confidence scores also influence AI-generated responses by intervening with abstention decisions. The abstention thresholds dynamically adjust based on these signals, which means the model may choose to not answer if confidence and sufficiency are both low.
While pages with complete and well-structured information are more likely to contain sufficient context, other factors such as how well the AI selects and ranks relevant information, the system that determines which sources are retrieved, and how the LLM is trained also play a role. You can’t isolate one factor without considering the broader system that determines how AI retrieves and generates answers.
If these methods are implemented into an AI assistant or chatbot, it could lead to AI-generated answers that increasingly rely on web pages that provide complete, well-structured information, as these are more likely to contain sufficient context to answer a query. The key is providing enough information in a single source so that the answer makes sense without requiring additional research.
What are pages with insufficient context?
- Lacking enough details to answer a query
- Misleading
- Incomplete
- Contradictory
- Incomplete information
- The content requires prior knowledge
The necessary information to make the answer complete is scattered across different sections instead of presented in a unified response.
Google’s third party 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.
Passages from the Quality Raters Guidelines:
“Low quality pages do not achieve their purpose well because they are lacking in an important dimension or have a problematic aspect”
“A page titled ‘How many centimeters are in a meter?’ with a large amount of off-topic and unhelpful content such that the very small amount of helpful information is hard to find.”
“A crafting tutorial page with instructions on how to make a basic craft and lots of unhelpful ‘filler’ at the top, such as commonly known facts about the supplies needed or other non-crafting information.”
“…a large amount of ‘filler’ or meaningless content…”
Even if Google’s Gemini or AI Overviews does not implement the inventions in this research paper, many of the concepts described in it have analogues in Google’s Quality Rater’s guidelines which themselves describe concepts about high quality web pages that SEOs and publishers that want to rank should be internalizing.
Read the research paper:
Sufficient Context: A New Lens on Retrieval Augmented Generation Systems
Featured Image by Shutterstock/Chris WM Willemsen
