A dense bottleneck is a constrained, fixed-size vector through which an encoder compresses text for dense retrieval. It acts as a lossy compressor: surface details that do not fit are discarded while core semantic and topical patterns remain. section 7.3
In HyDE, the bottleneck helps retrieval tolerate hallucinated details in a hypothetical document. Wrong dates or names may disappear during compression, but a wrong question framing remains and can steer retrieval toward the wrong subject area. section 7.3
HyDE retrieval pipeline
A language model generates a hypothetical passage, and an encoder maps it to one vector. The same encoder embeds corpus documents so both representations occupy the same vector space; retrieval then compares the hypothetical-document vector with corpus vectors. section 7.3





