A significant challenge in standard Retrieval-Augmented Generation is its heavy reliance on the relevance and accuracy of retrieved documents. A low-quality retriever is prone to introducing irrelevant information, which can impede the generator from acquiring accurate knowledge and potentially lead to increased hallucinations.
Conventional RAG approaches often indiscriminately incorporate these documents, regardless of whether they are truly relevant to the query. To address this lack of risk tolerance, researchers have developed Corrective Retrieval-Augmented Generation (CRAG) to improve the overall robustness of the generation process.
The core of the CRAG framework is a lightweight retrieval evaluator designed to assess the quality of documents retrieved for a specific query. This evaluator returns a confidence degree that triggers one of three distinct actions: Correct, Incorrect, or Ambiguous. If a retrieval is judged as Correct, the system refines the documents into precise knowledge strips, filtering out non-essential text to focus on the most relevant information. If the retrieval is identified as Incorrect, the documents are discarded entirely. In this scenario, the system instead resorts to large-scale web searches to find complementary knowledge sources for correction.
Furthermore, CRAG utilises a decompose-then-recompose algorithm to ensure that redundant contexts are eliminated and key insights are optimised. This algorithm segments documents into smaller units to ensure each contains an independent piece of information.
The Ambiguous action serves as a soft, moderating strategy for cases where the quality of the retrieval is hard to distinguish. When the evaluator is not confident in its judgment, CRAG combines both the refined internal documents and external web search results to complement each other. This approach strengthens the system’s resilience and makes it more adaptable to dynamic information landscapes. Furthermore, CRAG utilises a decompose-then-recompose algorithm to ensure that redundant contexts are eliminated and key insights are optimised. This algorithm segments documents into smaller units to ensure each contains an independent piece of information.
Testing across various datasets, including short-form and long-form generation tasks, shows that CRAG significantly improves the performance of standard RAG-based approaches. It has demonstrated margins of improvement as high as 36.6 per cent in accuracy for certain true-or-false question tasks. One of the most important aspects of CRAG is its plug-and-play nature, allowing it to be coupled with almost any generative model. Because it does not require models to learn specific reflection tokens, it offers greater flexibility than other advanced RAG frameworks. Ultimately, CRAG empowers AI to recognise when its existing knowledge is insufficient and autonomously seek better evidence from the wider internet.
The team at Academii are always happy to discuss all your training and education needs, help your organisation attract and train new talent, and build a resilient workforce. Please drop us a line here to know more.
