Perplexity Releases pplx-embed-v2-context-9b-preview: A Contextual Embedding Model That Retrieves Answers and Their Supporting Evidence
https://www.marktechpost.com/2026/09/30/perplexity-releases-pplx-embed-v2-context-9b-preview-a-contextual-embedding-model-that-retrieves-answers-and-their-supporting-evidence/📌 Perplexity 新 embedding 模型：檢索答案的同時，連「證據」一起找出來
TL;DR：pplx-embed-v2-context-9b-preview 訓練時連答案的佐證段落一起學，解決 RAG 單一金標段落的盲點。
RAG 系統最尷尬的場景之一：三份租約文件裡都寫著「每月租金為…」，但只有一份屬於你要找的那間房子。傳統的 chunk 檢索常常只盯著「哪一段最像答案」，卻漏掉了「哪一段能證明這個答案是對的」。Perplexity Research 與 turbopuffer 聯手釋出的 pplx-embed-v2-context-9b-preview，想解決的正是這個落差。
🤔 問題：金標段落之外的證據，全被當成負樣本
RAG 系統會把長文件切成多個 chunk，但一個 chunk 的意義經常依賴文件別處提到的實體、標題或定義。近期的 contextual embedding 模型多半採用 late chunking 來處理：先把整份文件編碼一次，再針對每個 chunk 做 pooling。問題在於訓練階段通常只標記「一個」金標 chunk，其餘全部當作負樣本，連能幫忙驗證答案的佐證句子也不例外。Perplexity 另外點出三個延伸問題：二元標籤給出的訓練訊號太粗糙；用 LLM 標註資料的成本會隨資料量線性增加；而且標籤往往綁死在單一種切分策略上。
🧩 方法：用「教師模型」打分，取代單一金標籤
pplx-embed-v2-context-9b-preview 的核心改動在訓練訊號：模型學習的目標是同時檢索出答案「以及」驗證這個答案所需要的上下文，而不是單一金標段落。負責打分的教師模型是 Perplexity 自家的 query-aware context compression 模型，它會同時讀入 query 與文件，為每個 token 評分。每個訓練批次會隨機抽樣一種切分策略，chunk 之間以一個學習到的 <|chunk_sep|> token 分隔並做 mean pooling。由於教師模型只在訓練階段運作，推論時不會增加延遲或儲存成本。
模型本身是從 Perplexity 內部的 9B ColBERT 檢索模型出發，接上線性投影輸出 2048 維向量，並透過 Matryoshka 訓練同時支援 1024 維；再搭配 quantization-aware training，讓模型原生支援 int8 embedding。這次釋出的版本是多個 checkpoint 的 model soup，訓練資料涵蓋約 430 個資料集、超過 50 種語言，且不含 ConTEB 資料。
📊 資料與評測：context-bench 上的表現
Perplexity 以自建的 context-bench 進行評測（2,099 筆查詢、38,894 份文件、2,458,072 個句子層級 chunk，採取窮舉式排序）。在 K=10 的設定下，Perplexity 報告其模型與 voyage-context-4 相比分別領先 14.4 分與 5.0 分（文中 Voyage 數值是以 Perplexity 公布的差距反推所得）。在成本面，Perplexity 指出以 1024 維 int8（每筆 1 KB）儲存，在其 chunk 檢索測試集上的表現，還略微超越 2048 維 float32（每筆 8 KB）的 voyage-context-4；而跨 74 項 MTEB 任務的平均 nDCG@10（敏感度指標）也在報告中一併公布。
🧩 怎麼用：已開源，但還沒進 API
這個模型現在就可以自行部署：權重已上架 Hugging Face，採用 MIT 授權，載入時需要 transformers>=5.4.0 並開啟 trust_remote_code=True。目前尚未整合進 Perplexity 自家 API，官方在模型卡上也特別註明，這是 preview 版本，權重與介面之後可能會有不向下相容的變動。
⚠️ 限制
模型卡明確標註這是 preview 版本，介面與權重可能隨時改動；目前也只能自行架設推論服務，無法透過 Perplexity API 直接呼叫。
🎯 實務啟示
如果你的 RAG 系統常常遇到「答案對了，但驗證不了」的狀況，這種把驗證證據也納入訓練目標的做法值得關注。由於權重已經開源且支援低維度 int8 embedding，對儲存成本敏感的團隊可以優先評估用 1024 維版本取代既有的高維浮點數 embedding，在縮小儲存體積的同時不犧牲太多檢索品質。
🔗 來源
- 標題：Perplexity Releases pplx-embed-v2-context-9b-preview: A Contextual Embedding Model That Retrieves Answers and Their Supporting Evidence
- 作者／機構：Asif Razzaq, MarkTechPost
- 連結：https://www.marktechpost.com/2026/09/30/perplexity-releases-pplx-embed-v2-context-9b-preview-a-contextual-embedding-model-that-retrieves-answers-and-their-supporting-evidence/
#Perplexity #RAG #EmbeddingModel #ContextualRetrieval #InformationRetrieval #OpenSource #HuggingFace #VectorSearch #MachineLearning #NLP
原始資料 MarkTechPost · 收集於 2026-10-01
摘要原文
Perplexity Research and turbopuffer have released pplx-embed-v2-context-9b-preview , a contextual embedding model for RAG pipelines. Each chunk is embedded with the full document in view. The real change is the training signal. The model learns to retrieve the answer along with the context needed to verify it, not one ‘gold passage.’ P Is it deployable? Yes, as a self-hosted preview. Weights are on Hugging Face under the MIT license. Loading requires transformers>=5.4.0 with trust_remote_code=True . It is not yet on the Perplexity API. The model card notes that weights and interface may change without backward compatibility. RAG systems split long documents into chunks. A chunk often depends on an entity, heading, or definition stated elsewhere. Contextual models address this with late chunking . The document is encoded in one pass, then pooled per chunk. Training, however, usually marks one gold chunk per query. Every other chunk becomes a negative, including the sentences that make the answer checkable. Perplexity lists 3 more problems. Binary labels give a coarse signal. LLM annotation cost grows linearly with dataset size. Labels are also tied to one chunking strategy. The teacher is Perplexity’s query-aware context compression model . It reads the query and document together and scores every token. Each batch samples a random chunking strategy. Chunks are separated by a learned <|chunk_sep|> token and mean-pooled. The teacher runs only during training, so inference adds no latency or storage. The model starts from an in-house 9B ColBERT retrieval model. A linear projection outputs 2048 dimensions. Matryoshka training also supports 1024 dimensions. Quantization-aware training enables native int8 embeddings. The release is a model soup of several checkpoints. Training used roughly 430 datasets covering over 50 languages, with no ConTEB data. An interactive walkthrough of Perplexity’s contextual embedding preview: why isolated chunks fail, how teacher distillation replaces the single gold passage, and what the reported numbers mean. Three lease files share the sentence “Monthly rent is …” . Only one belongs to 5 Park Avenue. Switch modes and run retrieval. Illustrative example modeled on the lease scenario in Perplexity’s post. Scores are for explanation only, not model outputs. A context compression model acts as teacher. It scores every token for the query. Those scores are pooled per chunk (mean of the top n tokens) and turned into a soft target, instead of a one-hot gold label. Teacher target (soft, from token scores) Change the boundaries: the same token scores re-aggregate without re-annotation. That is the “flexible chunk boundaries” property Perplexity describes. Token scores here are illustrative. context-bench (2,099 queries, 38,894 documents, 2,458,072 sentence chunks, exhaustive ranking). Numbers below are as reported by Perplexity at K = 10. Voyage values are computed as Perplexity’s figure minus the stated gap (14.4 and 5.0 points). Other Voyage metrics appear only in Perplexity’s chart and are not shown here. Contextual embeddings store one vector per chunk, same as a normal chunk index. Cost depends on vector size. Perplexity reports that 1024-dim int8 (1 KB) slightly exceeds voyage-context-4 at 2048-dim float32 (8 KB) on its chunk-retrieval suite. Bytes = dimensions x bytes per value. Sensitivity is mean nDCG@10 across 74 MTEB tasks, as reported by Perplexity.
由 claude-code/sonnet 自動生成