Collect Video Testimonials That Build TrustBest Embedding Models
By Venkatraman C
Last updated on Oct 7, 2026Embedding models convert text, images, audio, video and other data into numerical representations called vectors that capture meaning and relationships.
- Contextual document chunk embeddings that incorporate surrounding document information.
- 2048 dimensional embeddings with Matryoshka trained support for 1024 dimensional representations.
- Native INT8 quantized embeddings with dedicated query and document encoding methods.
Context Aware Retrieval
pplx-embed-v2-context-9b-preview differs from conventional independent chunk embedding workflows by encoding the chunks of a document together. Each resulting vector therefore represents its individual chunk while incorporating surrounding document information. Perplexity trains the model to retrieve both answer bearing chunks and evidence needed to support or verify those answers, making contextual retrieval its central technical objective.
Embedding Representation and Storage
The model outputs 2048 dimensional embeddings and was trained with Matryoshka losses at 1024 and 2048 dimensions. It also natively produces INT8 quantized embeddings. Perplexity states that contextualisation does not increase the number of vectors compared with an equivalent conventional chunk level index because the model still creates one vector per chunk. Dimension truncation and INT8 quantization provide documented options for reducing storage requirements.
Deployment and Preview Status
The model is distributed through Hugging Face and supports loading with Transformers using custom model code. It requires transformers>=5.4.0 together with PyTorch, NumPy, Safetensors and tqdm. Perplexity identifies this version as a preview and warns that weights, embeddings and the interface may change without backward compatibility. At release, Perplexity was still working on making the model available through its API.
- Multimodal retrieval across text, code, images, video and audio in a shared 768 dimensional embedding space
- Modular architecture ranging from a 270M text and code configuration to a 740M full multimodal configuration
- Matryoshka Representation Learning with 768, 512, 256 and 128 dimensional embeddings
Multimodal Retrieval Architecture
EmbeddingGemma 2 uses specialized encoders for text, code, vision and audio while projecting their representations into the same 768 dimensional space. This allows semantic similarity comparisons across modalities rather than requiring separate incompatible indexes. The architecture is modular, with documented configurations of 270M, 440M, 570M and 740M parameters depending on which modality encoders are loaded.
Context and Embedding Flexibility
The model provides an 8,192 token context window shared across its supported modalities. Matryoshka Representation Learning permits embeddings to be reduced from 768 dimensions to 512, 256 or 128 dimensions. Google states that 256 dimensional embeddings retain most text and code quality and about 95% of image, video and speech retrieval quality, while reducing vector storage requirements.
Retrieval Evaluation and Deployment
Google reports that EmbeddingGemma 2 scores 14% higher than EmbeddingGemma 1 on the Massive Text Embedding Benchmark for code. The model can run through sentence-transformers and is documented for use with tools including Hugging Face Transformers, vLLM, SGLang, MLX, Ollama, LM Studio and LiteRT. Google also documents efficient fine tuning through Unsloth.