Retrieval stops chasing one gold chunk: Perplexity's new embedders rank answers and their supporting evidence together
Synopsis
Perplexity released the pplx-embed-v2-context family, which distills graded chunk-level relevance from a query-aware context compression model instead of gold-chunk labels, so training targets no longer depend on how a document is chunked. The 9B model beats the strongest baseline, voyage-context-4, by 14.4 points at Answer@10 on context-bench and leads on Q2D-Chunk and ConTEB; the 0.6B model shares the 9B embedding space, so a small model can query an index built with the large one.
Interpretation
A retriever is now asked not only to find the chunk that answers a query but also to surface the chunks needed to interpret and verify that answer. Earlier contextual embedding training and evaluation assumed one relevant chunk per query, treated the document's other chunks as negatives, and rewarded only ranking that single chunk first. On context-bench, the 9B model reaches 45.5 Answer@10 against 31.1 for the strongest baseline, voyage-context-4, a 14.4-point gap, with turbopuffer running all evaluations blindly during development.
The training signal becomes graded chunk relevance distilled from a query-aware context compression model rather than a single gold chunk picked by an annotator. The compressor scores every document token with a keep probability, and aggregating those scores yields chunk relevance for any chunking, which decouples the target from chunk boundaries. Training mixes about 440 public and in-house query-document datasets in over 50 languages, roughly 57M pairs for the 9B model, with no chunk-level annotations, no ConTEB training data and no context-bench or Q2D-Chunk data.
Two new benchmarks label both the chunks that answer a query and the evidence chunks needed to verify the answer. Existing benchmarks label one gold passage per query and cannot test whether a model retrieves the entity, date or condition that makes an answer checkable. Q2D-Chunk holds 6,069 base queries and 11,719 with rewordings across 10 languages and 26 domains, averaging 14.0 answer spans per query; context-bench holds 2,099 queries over 38,894 documents, with 97% of queries carrying evidence.
Perspective
The demonstrated scope is text chunk retrieval and page-level visual document retrieval, evaluated on context-bench, Q2D-Chunk, ConTEB, ViDoRe-V3 and MIRACL-Vision, with documents of up to 32,768 tokens encoded in a single pass.
Evidence retrieval remains difficult on both benchmarks, with the 9B model at 46.6 corpus-level Evidence@10 on Q2D-Chunk, and the non-contextual Nemotron-3-Embed-8B leads on that evidence metric; Q2D-Chunk labels come from LLM labelers and an LLM adjudicator and should be read as model-assigned judgments.