An Answer, Its Evidence, and Everything Missing in Between

With two open-weight models of 9B and 0.6B, Perplexity improves contextual search and introduces a benchmark based on 71,100 web pages.

An answer can be perfectly accurate and yet impossible to verify if the information that justifies it remains scattered throughout a document. It is this limitation of the search engines used by AI that Perplexity seeks to address with pplx-embed-v2-context, a new family of embedding models designed to retrieve both answers and the passages needed to understand them.

Developed with the contribution of researchers from turbopuffer, the two models with 9 and 0.6 billion parameters are released as open weights. Their distinctive feature lies in the way they represent documents: each passage receives a numerical representation that takes the overall content into account, rather than just its text. This approach is detailed in the study Contextual Embeddings Beyond the Gold Chunk, devoted to the limits of contextual search systems. An answer can depend on multiple passages

In a retrieval-augmented generation (RAG) system, documents are generally split into fragments, or chunks, which are then indexed to retrieve the information relevant to a question. The problem arises when a passage contains the answer without specifying what it refers to. A rent amount, for example, may be listed in a lease while the address of the property appears several pages earlier.

Classic approaches often favor a single reference passage, called the gold chunk. This method overlooks the complementary fragments that establish context, identify a person, or provide an essential condition. Worse, these passages may be considered non-relevant during training, even when they contribute to verifying the answer.

To go beyond this restriction, the researchers repurposed a contextual compression system already used in Perplexity’s answer engine. It assigns relevance scores to elements of a document based on a query. These signals are then used to train the embedding models, which learn to distinguish passages that directly contain an answer from those that provide supporting evidence.

The compression model is used only during training. During retrieval, pplx-embed-v2-context directly produces its numerical representations, without requiring this additional step. The models can operate with a single vector per document or with one vector per fragment, with each fragment then retaining knowledge of the entire document. 71,100 pages to test retrieval and its evidence

Alongside the models, the team introduces two new benchmarks designed to measure capabilities that have so far received little attention. The first, context-bench, challenges systems with very similar documents to verify whether they can retrieve the correct answer while taking into account elements scattered throughout the text.

The second, Q2D-Chunk, is based on queries from Perplexity’s real traffic, reformulated to form standalone questions. It includes 6,069 initial queries, expanded into 11,719 formulations in total, and a corpus of 71,100 pages split into 9.17 million fragments of up to 128 tokens each. The evaluation covers ten languages and 26 domains, with annotations distinguishing passages containing answers from those needed to verify them.

This distinction changes the nature of the metrics. A search engine may quickly retrieve the right document while missing the conditions that make an answer valid. Q2D-Chunk therefore evaluates separately the retrieval of answers, that of evidence, and the amount of text that must be examined to gather all of the expected information.

On this benchmark, the 9-billion-parameter model reaches 86.4% in Answer@10, compared with 83.9% for voyage-context-4 and 83.1% for Perplexity’s 0.6-billion version. To cover all of the expected answers and justifications, its median result is 1,954 tokens, compared with 2,412 for voyage-context-4. The 9B model also ranks first in the published comparisons on context-bench and ConTEB.

These results come from the study conducted by the researchers and from their evaluation protocols. The Q2D-Chunk corpus remains private and the results are calculated by the team, even though a public leaderboard is provided. The reported performance therefore does not constitute an entirely reproducible independent validation. Visual search inherited from a model trained on text

The capabilities of the new models are not limited to text documents. Their architecture is based on a vision-language model capable of representing images and visual documents in a shared retrieval space.

One detail is particularly noteworthy: the 9B version was trained only on textual data for this new stage, while retaining visual search capabilities inherited from its initial architecture. The researchers also experimented with splitting pages into several horizontal bands, each with its own representation.

On ViDoRe-V3, a benchmark dedicated to searching within visual documents, the 9B model achieves 61.18 nDCG@10 points with 128-token fragments, compared with 58.23 for Nemotron-3-Embed-8B in the published comparison. With two 1,024-dimensional representations per page, the researchers also claim better performance than Nemotron-3.5-Embed-8B while using an index that is four times smaller.

The 0.6-billion-parameter version offers another advantage: it shares a representation space compatible with the first 1,024 dimensions of the 9B model. This makes it possible to index documents with the larger model and perform searches with the lighter one, without fully rebuilding the index. Two open-weights models for RAG systems

The 9B and 0.6B versions are listed on Hugging Face. They produce int8-quantized embeddings, with 2,048-dimensional representations for the larger model and 1,024-dimensional ones for the smaller. Retrieval can be performed at the level of an entire document or contextualized fragments, depending on the precision required.

This flexibility is particularly relevant for RAG systems that work on large corpora, where the relevance of an answer often depends on several distant passages. The published results nevertheless show