Learn more on how submodular formulation transforms an ad-hoc "select diverse queries" heuristic into a rigorous optimization problem with provable guarantees, efficient algorithms, and measurable objectives.

By formulating the problem into submodular optimization, we are able to use lazy greedy algorithm which runs O(nk) time comparing the original (nk) combinations, and have strong theoretical guarantees to find the optimal set.

Submodular function and submodularity may sound unfamiliar to many, but you may well have heard of the idea of "diminishing returns" - well, submodularity is the mathematical representation of that. In plain English: adding an element to a smaller set gives at least as much

But since it's cheap to generate a large number of queries, which eventually yields some good ones, why don't we treat this as a subset selection problem? i.e. first generate a large set, and then find the optimal diverse set?. Unfortunately, finding the optimal subset of k

First, we want to check if prompting is an effective way for generating diverse queries. We use jina-embeddings-v3 to measure the cosine similarity between the original query and generated queries and the within generated queries. The LLM is gemini-2.5-flash. One can see that

Many know the importance of diverse query generation in DeepResearch, but few take its implementation seriously. Most DeepResearch simply hardcode "diversity" into the prompt. We show a more rigorous approach to diverse query generation using sentence embeddings and submodular

Quantization can act as a regularizer, preventing the model from overfitting to the training data. By forcing the model to operate with lower precision, QAT can improve generalization performance and robustness. In we show QAT models sometimes.

Cliché is that quantization hurts performance—that you must trade quality for space. Reality? Skill issue. Learn how we trained quantized versions of jina-embeddings-v4 using Quantization-Aware Training (QAT), where models learn to work with rounding rather than fight against it.

Learn more about v4's training, design and benchmark Try it out today via our Search Foundation API or Hugging Face🤗 Let us know what you think.

jina-embeddings-v4 is built on the Qwen2.5-VL-3B-Instruct. Text and image inputs are processed through a shared pathway: images are first converted to token sequences via a vision encoder, then both modalities are jointly processed by the decoder. Three task-specific LoRA

jina-embeddings-v4 is our most ambitious embedding model yet. As an open-source model, v4 outperforms leading closed-source embedding models from major providers, delivering 12% better performance than OpenAI's text-embedding-3-large on multilingual retrieval (66.49 vs 59.27),

Today we're releasing jina-embeddings-v4, our new 3.8B universal embedding model for retrieving text, images, visual documents and code. V4 achieves state-of-the-art retrieval performance on multimodal and multilingual tasks across MTEB, MMTEB, CoIR, LongEmbed, STS, Jina-VDR,

But who rules the ruler? This is a fundamental problem with all synthetic benchmarks: How do you validate the validator? We provide some suggestive evidence that models rank similarly on AIR-BENCH vs established benchmarks and that AIR-Bench correlates.

The motivation comes from the limitations of MTEB and BEIR, where models suffer from "teaching to the test," making scores less meaningful for real-world applications. We naturally thought synthetic data could provide a better benchmark. However, naive generation produces poor

AIR-BENCH: Automated Heterogeneous Information Retrieval Benchmark - Our colab with BAAI (the team behind bge-m3/bge-1.5) was accepted by #ACL2025 main conf. Today, using LLMs for evals has become the norm: RAGAS, ARES, and AlpacaEval have applied this to different problems.

There are tons of interesting use cases for this visualization tool. We use it for validating citations in deep search, debugging late chunking, etc. Go try it out on and let us know what you think!.

One of the really useful features is being able to drag and select interesting areas to find out what's behind them. You can simply drag an area on the heatmap and it'll pop up a small panel showing you a mini heatmap of that area and the corresponding chunks. This is super

One interesting question people ask us is: "How do you guys vibe-check your embeddings?" Sure, there's MTEB for more serious quantitative evaluation on public benchmarks, but what do you do for open-domain or new problem? Today we want to share a small internal tool we use for

Learn more about fair scoring of multimodal docs: and try jina-reranker-m0 via our API with 10M free tokens on every new API key, or on AWS, GCP, Azure and 🤗.

This very simple two-stage approach delivers a 62% improvement in recall because the system finally leverages what humans do naturally: considering both what we read and what we see to determine relevance. Here's a visual example that shows jina-reranker-m0 consistently ranks the