Perplexity has released pplx-embed-v2-context, a new family of open-weight vision-language embedding models available in 9-billion-parameter and 0.6-billion-parameter versions. The models are designed to improve information retrieval across text, images and visual documents by preserving the wider context of a document when generating embeddings for individual sections.
The release targets developers building search engines and retrieval-augmented generation (RAG) systems. Its central idea is to help AI applications retrieve not just the passage containing an answer, but also the surrounding information needed to understand and verify that answer. The models are accompanied by a research paper developed with turbopuffer and are available through Perplexity’s Hugging Face organisation.
Key Takeaways
- Perplexity has introduced pplx-embed-v2-context in 9B and 0.6B parameter sizes.
- Both models are released with open weights for developer use.
- The models support contextual embeddings for document-level and chunk-level retrieval.
- They are designed for searching text, images and visual documents.
- The 9B model leads Perplexity’s reported evaluations on Context-bench, ConTEB and other retrieval benchmarks.
- The models use compatible embedding spaces, allowing developers to explore different model sizes for indexing and querying.
- The research introduces Context-bench and Q2D-Chunk to evaluate retrieval of answers and supporting evidence.
- The models are embedding systems, not general-purpose conversational AI assistants.
What Is pplx-embed-v2-context?
Embedding models convert information such as text, images or documents into numerical representations called vectors.
These vectors allow search systems to compare the meaning of a user’s query with information stored in a database.
For example, when someone searches for a company’s revenue in a long financial report, an embedding model helps identify the document sections most relevant to that question.
Perplexity’s new models are designed to improve this process by considering the entire document when creating representations for individual sections.
This is particularly useful for long documents where the meaning of one paragraph may depend on information introduced several pages earlier.
The two released model sizes are:
| Model | Parameter size | Intended use |
|---|---|---|
| pplx-embed-v2-context-9B | 9 billion | Higher-capability document and query retrieval |
| pplx-embed-v2-context-0.6B | 600 million | More lightweight retrieval workloads |
The smaller model offers a lower-compute option, while the larger version is intended for applications that prioritise retrieval quality.
Both belong to Perplexity’s second-generation embedding model family.
How Contextual Embeddings Work
Traditional embedding systems commonly divide long documents into smaller sections, known as chunks.
Each chunk is then converted into a vector independently.
This approach makes searching large document collections more manageable, but it can remove important information.
Consider a financial report containing the following statements:
- A company announced a new manufacturing facility.
- The investment amount was disclosed several paragraphs later.
- The expected production capacity appeared in another section.
If these passages are separated, a conventional embedding model may struggle to understand how they relate to one another.
Perplexity’s contextual embedding approach attempts to solve this problem.
Instead of encoding every chunk in isolation, the model considers the full document when generating its representation.
This allows a section to retain information about its wider context even when the section itself contains incomplete references.
The result is a retrieval system that can identify relevant passages while preserving more of the information needed to interpret them.
Why Retrieving Supporting Evidence Matters
Finding the paragraph containing an answer is not always enough.
An AI system may retrieve a passage containing a financial figure but miss the paragraph explaining the reporting period, accounting basis or conditions attached to that figure.
This can lead to incomplete answers or incorrect interpretations.
Perplexity’s research focuses on retrieving three related types of information:
| Retrieval objective | What it means |
|---|---|
| Document retrieval | Find the correct document |
| Answer retrieval | Find the section containing the answer |
| Evidence retrieval | Find additional sections needed to verify the answer |
The company argues that retrieval systems should identify both the answer and the supporting evidence.
This is particularly relevant for legal documents, financial reports, research papers and technical manuals, where an answer may depend on information spread across multiple sections.
New Research Benchmarks: Context-bench and Q2D-Chunk
Perplexity developed the models using a contextual retrieval approach and introduced evaluation benchmarks to measure their performance.
One of these is Context-bench, developed with turbopuffer.
The benchmark evaluates whether retrieval systems can identify relevant documents, find answer-containing sections and recover the supporting evidence needed to verify an answer.
Perplexity reports that its 9B contextual model achieves leading results on Context-bench and the established ConTEB benchmark.
The research also introduces Q2D-Chunk, a benchmark based on real-world search queries that evaluates retrieval at the document-chunk level.
The company says the 9B model achieves leading results on these evaluations as well.
These benchmarks address a limitation in conventional retrieval testing, where systems are often judged primarily on whether they find one designated answer passage.
The new approach places greater emphasis on recovering the information surrounding that answer.
Performance Across Text and Visual Documents
The models are built around a vision-language architecture, allowing them to work with more than conventional text documents.
Their intended applications include:
- Searching PDF documents and scanned pages.
- Finding information inside charts and tables.
- Retrieving relevant sections from research papers.
- Searching images using natural-language queries.
- Building knowledge bases containing mixed text and visual information.
Perplexity reports strong performance on the ViDoRe-V3 visual-document retrieval benchmark.
The company also says the 9B model achieves state-of-the-art results among dense retrieval models on visual-document tasks.
An important distinction is that the contextual models are designed to preserve document-level information while producing embeddings for sections. They should not be confused with Perplexity’s separate late-interaction model family, which retains multiple token-level vectors and uses a different query-document scoring approach.
The two approaches address related retrieval challenges but have different architectures and storage requirements.
Open Weights Give Developers More Flexibility
Perplexity has made the new model weights available through Hugging Face.
This gives developers the option to experiment with the models outside a conventional hosted chatbot interface.
The 9B model card provides an example of loading the model through the Transformers library using AutoModel and trust_remote_code=True.
This is useful for developers building custom retrieval systems, internal search tools and RAG applications.
However, open weights do not automatically mean that a model can run efficiently on every computer.
The 9B version has substantial hardware requirements. Its published model files occupy tens of gigabytes, so developers should evaluate available memory, acceleration hardware and deployment costs before attempting local inference.
The 0.6B version provides a smaller alternative for more constrained environments.
Developers should also review the model’s loading instructions and custom code before deploying it in production.
How Businesses Could Use the New Models
The biggest potential opportunity is enterprise search.
Companies maintain large collections of documents containing information about customers, products, finances, operations and internal policies.
Conventional search systems can struggle when relevant information is distributed across multiple documents or when a passage depends on wider context.
Perplexity’s contextual embedding approach could help businesses build more capable retrieval systems.
Financial research: An AI assistant could search annual reports and retrieve financial figures together with the accounting notes needed to interpret them.
Legal research: A retrieval system could identify contract clauses alongside definitions and supporting provisions elsewhere in the document.
Technical documentation: Developers could search manuals and retrieve instructions together with related technical explanations.
Enterprise knowledge management: Companies could build searchable internal databases containing reports, presentations, PDFs and other visual documents.
These are potential applications rather than confirmed customer deployments.
The Bigger Picture
The release highlights a growing shift in AI infrastructure from generating answers towards retrieving the right information before generating them.
As companies build more sophisticated AI agents, the quality of their underlying retrieval systems becomes increasingly important.
A model may be capable of producing a convincing response, but if its retrieval system misses important evidence, the final answer can still be incomplete or misleading.
Perplexity is attempting to address this challenge by improving how embedding models represent document context.
The availability of two model sizes also gives developers more flexibility to balance retrieval quality, inference costs and deployment requirements.
Looking Ahead
The next important test will be how pplx-embed-v2-context performs on independent enterprise datasets and production search workloads. Developers will need to compare retrieval accuracy, latency, memory consumption and indexing costs against existing embedding systems before adopting it at scale.
If contextual embeddings consistently improve the retrieval of both answers and supporting evidence, they could become an important component of future RAG applications and AI-powered enterprise search. Perplexity’s open-weight release gives developers an opportunity to evaluate that approach directly.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
