1. Given a query, instruct a generative model (ChatGPT) to write a passage to answer the question. The passage may contain factual errors, but it looks like a good answer!

2. The generated passage is passed through an Encoder (Contriever) to get the embedding of the passage. The encoder acts like a lossy compressor, where the extra (hallucinated) details are filtered out from the embedding.

3. A vector to search is performed against the corpus embeddings. The most similar real documents are retrieved and returned.

Paper: https://arxiv.org/abs/2212.10496