Perplexity
Perplexity opens embedding models that search text and scanned pages in context
pplx-embed-v2-context comes in 9B and 0.6B sizes, embeds each chunk with its whole document in view, and handles PDFs and images.
On 2 November 2026 Perplexity released pplx-embed-v2-context, a pair of open-weights embedding models in 9-billion and 0.6-billion-parameter sizes. Embedding models turn text and images into vectors so a system can find the passages most relevant to a question. They sit underneath most AI search and ‘chat with your documents’ tools.
Both return one int8 vector per document, or one per chunk that sees the whole document.
Chunks that know their document
Most retrieval systems split a document into chunks and embed each one on its own, so a paragraph that says ‘the rate rises to 4% from July’ loses track of which contract it came from. Contextual embeddings encode each chunk with the rest of the document in view. Perplexity’s paper, written with the vector database company turbopuffer, goes a step further: it argues that a retriever should find not only the chunk that answers a question but also the evidence needed to verify it, which is often spread across several chunks.
The models handle images as well. Pages from PDFs and scans are split into horizontal bands and each band gets its own embedding, which the paper says improves retrieval on visual documents.
Reported results and practical details
- The paper reports that the 9B model sets a new state of the art on two benchmarks the team introduced, Context-bench and Q2D-Chunk, and on the existing ConTEB benchmark. Q2D-Chunk is published as a public leaderboard.
- It also reports state-of-the-art results among dense retrievers on the visual ViDoRe-V3 benchmark.
- Embeddings are int8, which keeps storage small. The 9B model outputs 2,048 dimensions, and its first 1,024 share a space with the 0.6B model, so you can index documents with one model and query with the other.
- The 9B model card lists the MIT licence.
What it means for your business
When an AI assistant gives a wrong answer about your own policies or contracts, the cause is often retrieval rather than the language model: it was handed the wrong passage. Better embeddings, especially for scanned and image-heavy documents, are one of the cheapest ways to improve answers, and an open 0.6B model is small enough to run on ordinary hardware.
The paper’s focus on evidence is the useful part for business. An answer that arrives with the passages that support it can be checked by a person in seconds. That’s the standard to hold any document assistant to: show the source, so your team can verify before acting.
Sources
- Introducing pplx-embed-v2-context (opens in a new tab) · Perplexity Developers on X
- Contextual Embeddings Beyond the Gold Chunk (opens in a new tab) · arXiv
- pplx-embed-v2-context-9b model card (opens in a new tab) · Perplexity on Hugging Face
- pplx-embed-v2-context-0.6b model card (opens in a new tab) · Perplexity on Hugging Face
General information only, not legal, financial or professional advice. Check the sources for your circumstances.