Perplexity released preview weights for a model that searches document passages in their wider context. Training gives useful supporting passages credit alongside the passage containing an answer. The model can run on users’ own infrastructure under the MIT license; Perplexity API access is pending, and later versions may change its embeddings and interface.
Perplexity released pplx-embed-v2-context-9b-preview on September 30, a contextual embedding model intended to find answers and supporting passages together. Embedding models turn text into numerical representations that a search system can compare. This preview encodes document chunks with their surrounding document in view, so a passage can retain context supplied elsewhere in the text.[1], [2]
Conventional training can mark one answer-bearing passage as useful while treating the others as negatives, including passages needed to interpret that answer. Perplexity uses a context-compression model as a teacher. It reads a query and document together, scores individual tokens and turns those scores into guidance for the document chunks. The student model learns how relevance is distributed across them. The teacher operates during training, rather than requiring a second compression-model call during live retrieval.[1]
Preview weights are published on Hugging Face under the MIT license. The model supports 2048-dimensional vectors and a 1024-dimensional representation, with native eight-bit integer embeddings. Running it requires the Transformers model library, version 5.4.0 or later and permission to load its custom model code.[1]
Perplexity API access has not yet opened for this preview. Queries and document chunks use different encoding methods. The preview’s weights, embeddings and interface may change without backward compatibility; its model card warns against mixing current vectors with those generated by a future release.[1]
