极纳科技

Nailing these two details transforms your DeepSearch/DeepResearch from mid to GOAT: (1) selecting the best snippets from lengthy webpages and (2) ranking URLs before crawling. Check out our post and find out how to implement them right: https://lnkd.in/eyjKQvrq

We investigate embedding models on new "needle-in-haystack" tasks and find that beyond 4K tokens, they're just rolling dice - even with exact lexical matches or query expansion, they can't tell signal from noise in long context. https://lnkd.in/eZAXnxHa

This is either an extremely smart idea or an extremely stupid one—there's no in-between. https://lnkd.in/eFNRrv8A

QPS out, depth in. DeepSearch is the new norm. Find answers through read-search-reason loops. Learn what it is and how to build it. https://lnkd.in/e3kzHtZJ

Are you an Elasticsearch user? We are! We've learned a lot about search from Elastic, and now we're proud to be an official Elastic partner. Starting with Elasticsearch version 8.18, you can seamlessly integrate our state-of-the-art models such as jina-embeddings-v3 and jina-reranker-v2-base-multilingual in your RAG and search system. And that's not all – any new models we release will be readily available in Elasticsearch, ensuring a smooth user experience. Read on for more details:

Query expansion has come full circle - once essential, became obsolete, now making a comeback. If you've built an agentic search (like DeepSearch/DeepResearch), you immediately notice that query expansion is an extremely important component. The direct query from the user is often suboptimal - not concrete enough, not general enough, not specific enough for agentic tooling such as keyword-based search engines or structured databases to handle effectively. We desperately need query expansion/rewriting to match the right context, to think outside the box, to cover breadth while still digging deep.

Let's see if we can make search right. No ads, no login, no nonsense report. Pure. Deep. Search. ?? https://search.jina.ai/

Introducing DeepSearch, it search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, search, read, reason, ... ?? until the best answer is found! DeepSearch is our first agentic search system with reasoning and planning. Perfect for complex queries. Give it a vibe check here https://jina.ai/deepsearch or try it in your fav local chat client such as Chatwise, Cherry Studio, Chatbox. Our API is fully compatible with OpenAI Chat API schema - so simply swap the domain works!

For every model we release, there are three deployment options: Jina API, cloud service providers (CSP) like SageMaker, and self-hosted solutions. In this article, we compare five key performance metrics across these deployment scenarios to help you determine the optimal choice for your needs. Let's look at the cost-benefit analysis for each approach: https://lnkd.in/eYfrYciN

ModernBERT did three things right: high parameter-efficiency, code-friendly tokenizer and better long-context handling. In this post, we'll break down how ModernBERT stacks up against two models we know inside and out: jina-XLM-RoBERTa (the multilingual backbone behind jina-embeddings-v3) and RobERTa-large; and identify key development insights for future BERT-like models. Article: https://lnkd.in/eFAXa9NH