# Perplexity pplx-embed-v2-context: Models, API Availability and Retrieval Tradeoffs

Perplexity's contextual embedding models can search text and visual documents, with a shared vector space that changes how teams divide indexing and query costs.

By Paras Madan | 2026-11-03


Perplexity has announced pplx-embed-v2-context, open-weight embedding models in 9B and 0.6B sizes for searching text, images and visual documents. They represent document chunks with surrounding context, helping retrieval systems find supporting evidence rather than only a passage that appears to contain an answer.

The [official announcement](https://community.perplexity.ai/t/introducing-pplx-embed-v2-context-open-weight-vision-language-embedding-models-in-9b-and-0-6b/6466) appeared on November 3. The research paper was submitted on October 28, and the model repositories predate the announcement. This is the public introduction of that release, not a claim that the research or weights first appeared today.

## What changed from the preview

Perplexity previously published a 9B contextual embedding preview. The current family includes a smaller 0.6B model and documents retrieval across text and visual inputs. The final model names omit the preview suffix.

That matters for deployment. A team can consider a smaller query encoder alongside a larger document encoder, rather than running the larger model every time someone searches. The models share a compatible 1,024-dimensional embedding space, although the 9B model also supports a larger representation.

**Compatibility between these two final models does not establish compatibility with an old preview index.** The preview card warned that later embeddings could change. Existing users should test a separate index rather than mix vectors based on similar model names.

## Models and vector dimensions

The [9B model card](https://huggingface.co/perplexity-ai/pplx-embed-v2-context-9b) and [0.6B model card](https://huggingface.co/perplexity-ai/pplx-embed-v2-context-0.6b) describe these configurations:

| Model | Vector dimensions | Retrieval modes |
| --- | --- | --- |
| pplx-embed-v2-context-9b | 2,048, or a trained 1,024-dimensional prefix | Contextual and independent text inputs; images |
| pplx-embed-v2-context-0.6b | 1,024 | Contextual and independent text inputs; images |

For cross-model retrieval, use the 9B model's 1,024-dimensional prefix and apply the documented normalization. Arbitrarily cutting vectors to another size is not equivalent to using a trained dimension.

The models produce int8-quantized embeddings. That describes the output representation, not a promise that the model weights or inference memory requirements are eight-bit. Vector storage and the hardware needed to create those vectors are different costs.

## How contextual retrieval works

In ordinary chunk retrieval, a paragraph may be encoded without the document around it. A phrase such as “the renewal date” can then lose the information that tells a search system which agreement it refers to.

A contextual embedding is computed with surrounding document content available. Perplexity's [technical report](https://arxiv.org/abs/2610.38854) describes training with relevance signals distilled from a query-aware context-compression model. The aim is to retrieve both answer-bearing chunks and evidence that helps verify the answer.

The paper reports leading results on its evaluated contextual-retrieval benchmarks. These are research results from the authors, not a guarantee that the same ranking holds on a company's documents. Evaluation should include questions where the answer and its qualifying details sit in different parts of a file.

Visual retrieval follows a related idea. A document page can be split into horizontal bands, with a vector for each band, while the image is encoded contextually. These vectors work with standard dense approximate-nearest-neighbor indexes. More bands also mean more stored vectors, so the indexing cost still needs measuring.

## API availability and pricing

**Open weights are available; a hosted v2-context API price was not verified.** At publication, Perplexity's [API pricing page](https://docs.perplexity.ai/docs/getting-started/pricing) listed v1 standard and contextual embedding models, without a v2-context rate. Neither final model card listed a Hugging Face inference provider.

Do not apply the older models' token prices to this release. For now, the verified implementation route is the published model code and weights, with deployment and compute costs handled by the team running them.

The cards require Transformers 5.4.0 or newer and custom model code. Review that code before enabling remote-code loading in a production environment.

## Integration details worth checking first

Queries must use `encode_queries`; document chunks use `encode`. The cards warn that using the document method for queries silently reduces retrieval quality. Contextual document chunks should be non-overlapping, and oversized text inputs need splitting or truncation before encoding.

These details can matter more than a headline benchmark. A retrieval pipeline can run successfully while producing weaker search results because the wrong encoder path or normalization was used.

For teams with large, slowly changing collections, the useful experiment is a larger model at indexing time and a smaller model at query time. Keep that split only if answer quality survives on real queries. The opportunity is to spend more once when preparing the evidence, rather than pay the same cost every time someone needs it.

I'm Paras. I work at the intersection of AI, engineering and distribution.

If you're building an AI product, I help turn technical launches and product thinking into stories developers and founders read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).