Description
pplx-embed-v2-context-9b-preview is a contextual embedding model specifically designed for use in retrieval-augmented generation (RAG) systems. This model processes documents by dividing them into chunks and encoding these chunks together. The resulting embeddings reflect the context of each chunk, providing a more nuanced representation. This is a preview release, and users should be aware that weights, embeddings, and the interface may change in future versions, potentially affecting backward compatibility.
Read More
The model employs different encoding methods for queries and documents, using encode_queries for queries and encode for document chunks. It is important to note that encoding queries with the document method can degrade retrieval quality. The model produces unnormalized int8-quantized embeddings, which can be compared using cosine similarity or normalized for dot product comparison.
The pplx-embed-v2-context-9b-preview model requires specific dependencies, including transformers version 5.4.0 or higher, torch, numpy, safetensors, and tqdm. It utilizes custom code, necessitating the use of trust_remote_code=True when loading the model. Users can adjust various options such as batch size, normalization of embeddings, conversion to NumPy arrays, and device allocation.
The model was trained with Matryoshka losses at dimensions of 1024 and 2048, allowing for flexibility in embedding size. However, only the 1024 and 2048 dimensions were trained, and other truncation sizes are not supported. This model is ideal for developers and researchers working with RAG systems who require contextual embeddings for document processing.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
pplx-embed-v2-context-9b-preview Highlights
Contextual embedding model
Supports RAG systems
Encodes document chunks
Produces int8-quantized embeddings
Separate encoding for queries and documents
Custom code support
Matryoshka loss training
Dimensions: 1024 and 2048
Getting Started with pplx-embed-v2-context-9b-preview
Access model: Visit Hugging Face
Authenticate: Use your credentials
Set up environment: Install dependencies
Integrate via API: Use encode and encode_queries
Optimize: Adjust batch size and normalization
pplx-embed-v2-context-9b-preview's Use Cases
- Document Chunking
- Query Encoding
- Contextual Analysis
- RAG System Integration
- Embedding Comparison
