Philipp Krenn

"Float Bloat: vector serialization gone wrong" by storing float64 instead of float32. it's a surprising problem but with elasticsearch you're in luck: * since 9.2 (index creation) `index.mapping.exclude_source_vectors` is on by default — you don't store the floats in source at all any more 1/3

* `semantic_text generates` the embeddings server-side and has always avoided the problem and for even more optimized ingestion, use base64 encoded strings: www.elastic.co/search-labs/... 2/3

are you in the weights? it's the new klout also works for products (though reluctantly) #elasticsearch intheweights.com

BM25: "why won't you die?!" > the lexical retriever BM25 with appropriate setup outperforms neural rankers in most cases; notably, gpt-oss-20b with BM25 on the passage corpus achieves the highest answer accuracy across all retrieval settings in our study;

obligatory GTC keynote tweet when you make the top slides: building HNSW graphs with up to 12x the throughput and 7x faster merges on elasticsearch and NVIDIA cuVS (a GPU-accelerated library for vector search)

*claude code is the new gatekeeper* fascinating report on the choices claude makes, build vs buy, the personality of different models, and who is winning / losing 1/5 PS: when creating code is cheap, the new moat is distribution

models have personalities in their recommendations and how conservative or cutting edge they are — massive differences even within a close "family" 4/5

*claude code is the new gatekeeper* fascinating report on the choices claude makes, build vs buy, the personality of different models, and who is winning / losing 1/5 PS: when creating code is cheap, the new moat is distribution

depending on the area, there is a clear bias towards building vs buying feature toggles as a full blown company were always a weird choice... 2/5

2 new #jina models have entered the embedding arena — v5 multilingual: * small: 1024 dim, 32K context * nano: 768 dim, 8K context both support matryoshka dimension truncation (32+) and are at the top of current benchmarks — especially for their parameter size publicly accessible (non-commercial)

* OpenClaw agent opening a PR on matplotlib * reviewer rejects it per the repo's policies * agent replies with a personal attack github.com/matplotlib/m...

new public #elastic roadmap: * covering key initiatives like ES|QL, better dashboards,... * recently shipped features (those are our fiscal quarters) * upcoming features as in-progress, near-term, and mid-term 1/2

long US weekend — great time to catch up on some @JinaAI_ papers about rerankers and code / multilingual / multimodal embeddings: * jina-reranker-v3 * jina-code-embeddings * jina-embeddings-v4 * ReaderLM-v2 * jina-clip-v2

jina-clip-v2 uses a multi-task, multi-stage contrastive learning strategy to align multilingual text and image representations to work well at both cross-modal and text-only retrieval. it excels at visually rich documents while flexibly truncating embedding dimensions arxiv.org/abs/2412.08802

long US weekend — great time to catch up on some @JinaAI_ papers about rerankers and code / multilingual / multimodal embeddings: * jina-reranker-v3 * jina-code-embeddings * jina-embeddings-v4 * ReaderLM-v2 * jina-clip-v2

jina-embeddings-v4 works by projecting text & images into a shared semantic space using a unified Qwen2.5-VL backbone and task-specific LoRA adapters, which minimizes the modality gap and enables SOTA retrieval of visually rich documents via single- and multi-vector outputs arxiv.org/abs/2506.18902

long US weekend — great time to catch up on some @JinaAI_ papers about rerankers and code / multilingual / multimodal embeddings: * jina-reranker-v3 * jina-code-embeddings * jina-embeddings-v4 * ReaderLM-v2 * jina-clip-v2

jina-reranker-v3 uses a new "last but not late" interaction strategy that processes the query and multiple documents simultaneously in a single shared context window, allowing it to capture cross-document and query-document relationships with its compact 0.6B parameter model arxiv.org/abs/2509.25085