jina-web-search-dsh-plugin

DeepSeek Harness ships a reliable built-in web_search for everyday queries, but it falls short the moment a search turns academic. arXiv results often come back as mirror-site links instead of official ones, SSRN social-science publications are simply out of scope, and there is no native path to text embedding, relevance reranking, or PDF formula extraction. jina-web-search-dsh-plugin bridges that gap by wiring all twelve jina_* model tools directly into dsh, so an agent can call them mid-conversation without the user juggling a separate CLI or hand-rolling API requests.

For research workflows the plugin splits arXiv and SSRN into two dedicated tools whose names map one-to-one to the target domain, so the model routes straight to arxiv.org or papers.ssrn.com official links. General web search adds time-range filtering, region and language hints, and priority ranking for official Chinese-language sources such as government and corporate sites. Beyond search the toolkit covers web-page-to-Markdown conversion, full-page screenshots, query expansion, text vectorization via jina-embeddings-v5, relevance reranking with jina-reranker-v3.5, text classification, PDF chart and formula extraction, and a jina_primer tool that hands the model the host clock, public IP, and Jina account balance on demand.

The plugin is well suited to graduate students and researchers doing literature reviews inside DSH, Chinese-language content creators who need authoritative and time-filtered results, and developers building headless NLP pipelines that chain search into embedding and reranking in a single session. Users on mainland China networks also benefit from automatic system-proxy discovery on every call, self-healing when a VPN changes port, and clear error messages when no tunnel is available.