The community member is asking about using Reverse HyDE (Hypothetical Document Embeddings) with LlamaIndex. Reverse HyDE generates not just a vector representation and metadata, but also a list of questions that the chunk can potentially answer. The documentation on this feature is limited.
In the comments, another community member suggests using a question/answer metadata extractor and mapping the Q/A pairs to their original chunks. Another community member mentions the DocumentSummaryIndex, which can generate and embed a document summary as well as the questions the document can answer.
There is no explicitly marked answer in the comments.
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.
