Article
Perplexity open-sources a new retrieval model: 0.6B a small model can directly search 9B indexes created by large models
According to a news flash from Beating AI, Perplexity has open-sourced two multimodal embedding models pplx-embed-v2-late, namely 0.6B and 9B. The two models’ vectors are mutually compatible. Developers can use the 9B large model to build a knowledge base, then use the 0.6B small model to handle routine searches, without running the large model for every query.
The new models also support retrieval from text, images, and PDF pages. Conventional embedding models typically compress a piece of content into a single vector, which can easily lose details. The new models retain a 128-dimensional vector for each token, allowing search terms to match relevant content in documents individually. When processing PDF、PPT and scanned documents, page images can be converted directly into vectors without first using OCR to extract text, while preserving charts, tables, and layout information.
The two models were trained from the same 18B teacher model and share a single vector space. In the ViDoRe v3 image retrieval benchmark, the official evaluation found that using the 0.6B model for indexing and querying produced a score of 62.3%; using 9B for indexing and 0.6B for querying increased the score to 63.5%. The score when using the 9B model exclusively was 65.2%.
The weights of both models have been released on Hugging Face under the MIT license.