Jina Embeddings V2: From Raw Data to Bilingual Hybrid Search
summary
German‑English embedding model trained on ~400–500M web documents with ~50/50 English and German data, using a BERT/ALiBi backbone for long context, mined positive pairs, consistency filtering and hard negatives retrieved from an index then filtered by a cross‑encoder. Evaluation beats comparable bilingual/multilingual baselines on MTB tasks and in an e‑commerce test of 70k products shows MRR@20: bm25 30% vector 51% hybrid 53%; hybrid combines normalized BM25 and cosine with a 0.2/0.8 weighting. Model and code are open source on Hugging Face and used in a demo search pipeline, plus a multimodal Gina Clip release.
AI-written summary · about
chapters (12)
- 0:00Introduction to Jina AI
- 1:34Understanding embeddings
- 5:02Training BERT models
- 7:58Bi-encoder training
- 11:05Triplet and hard negatives
- 13:24Bilingual data distribution
- 15:34Hybrid search overview
- 18:14Hybrid search implementation
- 23:41Deployment and ranking
- 24:44Multimodal Jina CLIP
- 29:40Future models and roadmap
- 34:10Q&A session
related talks
45mHybrid search: Greater than the sum of its parts?
25mSemantic Search
24mRediscover Your Data with a New Multimodal Search Capability
39mAI-powered Semantic Search; A story of broken promises?
01h 30mBuilding a Search Engine
41mSemantic Multimodal Multilingual Similarity Engine
22mMultilingual embeddings to scale NLP models to multiple...
41mGenAI, vector/semantic/hybrid search, RRF, NLP, (...) and other buzzwords