Julius Lipp

This two-stage method uses GRPO and fine-grained scoring from token logits to enable small language models (SLMs) to outperform much larger competitors in document reranking.

Blog post demonstrating how OCR errors create a performance ceiling for RAG systems, and how multimodal models can improve retrieval by a good margin.

Second-generation reranking models using reinforcement learning, supporting 100+ languages with up to 32k token context.

Compact 22.7M parameter embedding model with support for binary quantization and matryoshka embeddings.

A library that adds dynamic batching to inference systems, grouping requests/tensors to maximize model throughput.

Open-source information retrieval testing framework supporting sparse, dense, and hybrid search with unified benchmarking.

Enhanced lexical search algorithm improving BM25 with entropy-weighted similarity and weighted query augmentation.

Collaboration with deepset producing high-performance German/English embedding models with support for binary quantization/MRL.

ColBERT model for reranking and retrieval, outperforming other ColBERT and cross-encoder models on BEIR.

Embedding model supporting Matryoshka for both hidden layers and embeddings, enabling flexible 2D dimensionality reduction.

State-of-the-art embedding model (mxbai-embed-large-v1) trained on 700M+ data pairs, outperforming OpenAI's text-embedding-v3.

Family of open-source reranking models (xsmall, base, large).

AI-powered college finder for admiring students

Blog post demonstrating how OCR errors create a performance ceiling for RAG systems, and how multimodal models can improve retrieval by a good margin.

Second-generation reranking models using reinforcement learning, supporting 100+ languages with up to 32k token context.

Compact 22.7M parameter embedding model with support for binary quantization and matryoshka embeddings.

A library that adds dynamic batching to inference systems, grouping requests/tensors to maximize model throughput.

Open-source information retrieval testing framework supporting sparse, dense, and hybrid search with unified benchmarking.

Enhanced lexical search algorithm improving BM25 with entropy-weighted similarity and weighted query augmentation.

Collaboration with deepset producing high-performance German/English embedding models with support for binary quantization/MRL.

ColBERT model for reranking and retrieval, outperforming other ColBERT and cross-encoder models on BEIR.

Embedding model supporting Matryoshka for both hidden layers and embeddings, enabling flexible 2D dimensionality reduction.

State-of-the-art embedding model (mxbai-embed-large-v1) trained on 700M+ data pairs, outperforming OpenAI's text-embedding-v3.

Family of open-source reranking models (xsmall, base, large).

This two-stage method uses GRPO and fine-grained scoring from token logits to enable small language models (SLMs) to outperform much larger competitors in document reranking.

Enhanced lexical search algorithm improving BM25 with entropy-weighted similarity and weighted query augmentation.

Second-generation reranking models using reinforcement learning, supporting 100+ languages with up to 32k token context.

Compact 22.7M parameter embedding model with support for binary quantization and matryoshka embeddings.

Collaboration with deepset producing high-performance German/English embedding models with support for binary quantization/MRL.

ColBERT model for reranking and retrieval, outperforming other ColBERT and cross-encoder models on BEIR.

Embedding model supporting Matryoshka for both hidden layers and embeddings, enabling flexible 2D dimensionality reduction.

State-of-the-art embedding model (mxbai-embed-large-v1) trained on 700M+ data pairs, outperforming OpenAI's text-embedding-v3.

Family of open-source reranking models (xsmall, base, large).

A library that adds dynamic batching to inference systems, grouping requests/tensors to maximize model throughput.

Open-source information retrieval testing framework supporting sparse, dense, and hybrid search with unified benchmarking.