Self-reflection can enhance RAG, enabling correction of poor quality retrieval or generations.
Corrective Retrieval-Augmented Generation (CRAG) is a method that works like a built-in fact-checker.
It adds both creativity and accuracy by creating text and then checking for any mistakes or made-up information. This helps make sure the final result is reliable and matches real-world facts. It's like a safety feature for AI writers, making their work more trustworthy and lowering the chances of spreading false information.
Corrective-RAG (CRAG) is a recent paper that talks about a cool way to make a self-reflective RAG.
The method givesating/scores retrieved documents based on how well they answer a question:
For Correct documents -
- If at least one document is really relevant, it moves on to creating text
- Before creating text, it cleans up the knowledge
- This breaks down the document into "knowledge strips"
- It rates each strip and gets rid of ones that don't matter
For Ambiguous or Incorrect documents -
- If all documents are not relevant enough or if it's not sure, the method looks for more information
- It uses a web search to add more details to what it found
- The diagram in the paper also show that they might change the question to get better results.