Hey guys, has anyone work with Reverse HyDE with LlamaIndex? Reverse Hyde (Hypothetical Document Embeddings) is when during embedding, each chunk gets generated not just a vector representation and some metadata, but also generates and stores a list of questions that it potentially can answer. Not much info in the documentation about it, just for regular HyDE (which is different from Reverse HyDE).
There is a question/answer metadata extractor yea.
But probably more useful would be taking those q/a pairs and mapping them to their original chunk, and storing each one as a node. Then swapping them out on retrieval
Not exactly what you asked for but the DocumentSummaryIndex can do something like that if you use the default prompt template. It will generate and embed a document summary as well as the questions that the document can answer.