Perplexity lets a small model search a larger model’s document index
Perplexity has released two open-weight models for finding information in text, images and visual documents. Their shared representation lets the smaller model search a collection prepared with the larger one. Developers can separate document indexing from live queries and search page images without first extracting their text. Both models are available on Hugging Face; retaining many vectors per document creates a storage trade-off.
Artificial Intelligence··Midday
Two search models share one representation
Perplexity released a pair of open-weight search models on October 7. The models handle text, images and visual pages and share the same representation space across their different sizes. They are publicly available through Hugging Face, a platform for distributing machine-learning models.[1], [2]
The smaller version has 0.6 billion parameters and the larger has 9 billion. A document collection indexed by the larger model can be queried with the smaller one. That separates the model preparing documents for search from the model responding to a live request.[1]
Individual tokens contribute to the match score
The models retain a separate vector for each token, a small unit of text or visual input. Each query token is compared with document tokens, and its strongest match contributes to the final score. This late-interaction approach preserves multiple representations for a document instead of compressing its entire contents into one vector.[1]
Page images can be searched before extracting text
The release covers PDFs and slide decks as well as ordinary text and images. Visual pages can be searched without first applying optical character recognition or parsing their text. Keeping many token vectors increases the storage needed for a collection. Developers can inspect, run and adapt the released weights on their own infrastructure.[1]
