pplx-embed-v1-late-0.6b
by perplexity-ai
Late-interaction (ColBERT-style) embedding model from Perplexity AI
perplexity-ai/pplx-embed-v1-late-0.6bmixpeek://text_extractor@v1/perplexity_pplx_embed_late_06b_v1Overview
pplx-embed-v1-late is a 0.6B parameter late-interaction embedding model from Perplexity AI that uses ColBERT-style token-level representations with MaxSim scoring. Unlike dense single-vector embeddings, it produces 128-dimensional vectors for each token, enabling fine-grained matching that captures partial document relevance. It outperforms ColBERT-zero on BEIR (56.61 nDCG@10) and jina-colbert-v2 on MIRACL multilingual retrieval (66.62).
Architecture
Late-interaction architecture based on the pplx-embed-v1-0.6b backbone. Produces per-token 128-dimensional vectors instead of a single document vector. Scoring uses MaxSim — for each query token, find the maximum similarity to any document token, then sum across query tokens. This enables fine-grained partial matching that dense embeddings miss. Optimized CUDA and Metal kernels available for efficient scoring.
Mixpeek SDK Integration
from mixpeek import Mixpeekmx = Mixpeek(api_key="YOUR_KEY")mx.ingest.documents(source="s3://corpus/legal-docs/",collection="legal_search",feature_extractors=[{"name": "text_embeddings","model": "perplexity-ai/pplx-embed-v1-late-0.6b","params": {"interaction": "late", "dim": 128}}])
Capabilities
- Fine-grained token-level document matching via MaxSim
- Better partial relevance detection than dense embeddings
- Multilingual retrieval (strong MIRACL performance)
- Compatible with existing ColBERT indexing infrastructure
- Optimized GPU/Metal kernels for production scoring
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| BEIR | nDCG@10 | 56.61 | Beats ColBERT-zero |
| MIRACL | nDCG@10 | 66.62 | Beats jina-colbert-v2 on multilingual |
Performance
Specification
Build a pipeline with pplx-embed-v1-late-0.6b
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Open Studio