Every research agent hits the same wall: parallel search returns a mass of near-identical passages, and reading them all is unaffordable. Having reranking and semantic deduplication as callable tools — rather than something the model attempts by reading — turns that mass into a shortlist before it enters the context. guess_datetime_url is the other quiet favourite: knowing when a page was actually published is the check most agents skip.
Jina's official endpoint, which listed 22 tools to our probe as version 1.7.0. It is the most complete research pipeline in this directory — retrieval, parallel retrieval, and then the embedding operations that make a large result set usable.
- read_url and search_web — page content as clean markdown, and web search with locale and time filters
- search_web_deep — search, read each result, and return the passage that best answers the question
- parallel_search_web, parallel_search_arxiv, parallel_search_ssrn, parallel_read_url — several queries or pages at once
- search_arxiv, search_ssrn, search_bibtex, search_images, search_jina_blog — academic and image search, with BibTeX citations from DBLP
- sort_by_relevance, deduplicate_strings, deduplicate_images, classify_text — rerank, take the top-k semantically unique items, and label texts
- extract_pdf — figures, tables and equations from a PDF via layout detection
- capture_screenshot_url, guess_datetime_url, expand_query, primer, show_api_key
A Jina API key is optional. The parallel and embedding tools are what most benefit from a paid quota.
Paste a URL with a key — add the endpoint to your client