pplx-embed-v2-context-9b-preview

About pplx-embed-v2-context-9b-preview

pplx-embed-v2-context-9b-preview is a contextual embedding model from Perplexity designed for document chunk retrieval in Retrieval Augmented Generation systems. Instead of encoding each chunk independently, it accepts a document as a list of chunks and encodes them together so that each returned chunk embedding reflects surrounding document context.

The model produces one embedding for each document chunk and is trained to retrieve both answer bearing chunks and supporting context. Queries and document chunks use separate encoding methods and fixed query and document prefixes. Perplexity provides dedicated encode_queries functionality for queries and encode for document chunks.

The model produces 2048 dimensional embeddings and was trained with Matryoshka losses at both 1024 and 2048 dimensions. It natively produces unnormalized INT8 quantized embeddings and uses mean pooling. The model is available as a preview release through Hugging Face under the MIT license and can be loaded with Hugging Face Transformers using its custom model code.

Key Features

• Contextual Chunk Embeddings: Encodes document chunks together so that each chunk embedding incorporates information from surrounding document context.
• RAG Retrieval: Designed specifically for retrieving document chunks in Retrieval Augmented Generation systems.
• Answer and Evidence Retrieval: Training targets both answer chunks and supporting contextual evidence instead of focusing only on a single relevant chunk.
• 2048 Dimensional Embeddings: Produces 2048 dimensional embeddings with Matryoshka training at 1024 and 2048 dimensions.
• INT8 Embeddings: Natively produces unnormalized INT8 quantized embeddings that can be compared using cosine similarity.
• Separate Query Encoding: Provides separate encode_queries and encode methods for queries and document chunks, using fixed query and document prefixes.