[Perplexity Open-Sources New Retrieval Model: 0.6B Small Model Can Directly Query Index Built by 9B Model] Beating AI Newsflash: Perplexity has open-sourced two multimodal embedding models, pplx-embed-v2-late, with sizes of 0.6B and 9B respectively. The vector representations of the two models are mutually compatible, allowing developers to use the 9B large model to build a database and the 0.6B small model for daily searches, eliminating the need to run the large model for every query. The new models support text, image, and PDF page retrieval. Traditional embedding models typically compress a piece of content into a single vector, which can result in loss of detail. The new models retain a 128-dimensional vector for each token, enabling search terms to match relevant content within documents individually. When processing PDFs, PPTs, and scanned documents, the models can directly convert page images into vectors without requiring OCR to extract text, while also preserving charts, tables, and layout information. Both models were trained using the same 18B teacher model and share a unified vector space. In the ViDoRe v3 image retrieval test, it was found that using the 0.6B model for both database creation and querying achieved a score of 62.3%. When the 9B model was used for database creation and the 0.6B model for querying, the score improved to 63.5%. Fully utilizing the 9B model resulted in a score of 65.2%. The weights for both models have been released on Hugging Face under the MIT license. [Original Link]