pplx-embed-context-v1-4b
by perplexity-ai
Contextual embedding that encodes document chunks with awareness of surrounding content
perplexity-ai/pplx-embed-context-v1-4bmixpeek://text_extractor@v1/perplexity_pplx_embed_context_4b_v1Overview
PPLX Embed Context is the first open-weight contextual embedding model. Unlike standard embedders that process each chunk in isolation, it sees the full document while encoding each chunk — so the resulting vector captures both local content and the chunk's relationship to surrounding text. This eliminates the retrieval failure mode where a chunk about 'it increased 15%' is meaningless without knowing what 'it' refers to.
On Mixpeek, contextual embeddings improve retrieval precision for document-heavy pipelines where chunks reference earlier sections, use pronouns, or contain relative statements. One model call replaces the traditional approach of prepending context summaries to each chunk.
Architecture
Diffusion-pretrained Qwen3 backbone with contextual attention mechanism. 4B parameters. Encodes each chunk while attending to the full document context. Produces dense embeddings where each chunk vector encodes both local semantics and positional context within the document.
Mixpeek SDK Integration
import { Mixpeek } from "mixpeek";const mx = new Mixpeek({ apiKey: "API_KEY" });await mx.collections.ingest({collection_id: "document-collection",source: { url: "https://example.com/report.pdf" },feature_extractors: [{feature: "text_embeddings",model: "perplexity-ai/pplx-embed-context-v1-4b"}]});
Capabilities
- Document-aware chunk embedding (sees full document while encoding each chunk)
- SOTA on ConTEB contextual retrieval benchmark (81.96 nDCG@10)
- Competitive on standard MTEB retrieval (69.66 nDCG@10)
- Eliminates pronoun/reference ambiguity in chunk embeddings
- MIT license for unrestricted commercial use
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| ConTEB (contextual retrieval) | nDCG@10 | 81.96% | Perplexity Research, 2026 |
| MTEB Multilingual v2 (retrieval) | nDCG@10 | 69.66% | Perplexity Research, 2026 |
Performance
Specification
Research Paper
pplx-embed: State-of-the-Art Embedding Models for Web-Scale Retrieval
arxiv.orgBuild a pipeline with pplx-embed-context-v1-4b
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Open Studio