深度搜索 API
api.openai.com
與 deepsearch.jina.ai
交換即可開始。reasoning_effort
參數。reasoning_effort
參數。{
"role": "user",
"content": "hi"
}
curl https://deepsearch.jina.ai/v1/chat/completions \
-H "Content-Type: application/json"\
-H "Authorization: Bearer " \
-d @- <<EOFEOF
{
"model": "jina-deepsearch-v1",
"messages": [
{
"role": "user",
"content": "Hi!"
},
{
"role": "assistant",
"content": "Hi, how can I help you?"
},
{
"role": "user",
"content": "what's the latest blog post from jina ai?"
}
],
"stream": true,
"reasoning_effort": "medium"
}
EOFEOF
{
"id": "1742181758589",
"object": "chat.completion.chunk",
"created": 1742181758,
"model": "jina-deepsearch-v1",
"system_fingerprint": "fp_1742181758589",
"choices": [
{
"index": 0,
"delta": {
"content": "The latest blog post from Jina AI is titled \"Snippet Selection and URL Ranking in DeepSearch/DeepResearch,\" published on March 12, 2025 [^1]. This post discusses how to improve the quality of DeepSearch by using late-chunking embeddings for snippet selection and rerankers to prioritize URLs before crawling. You can read the full post here: https://jina.ai/news/snippet-selection-and-url-ranking-in-deepsearch-deepresearch\n\n[^1]: Since our DeepSearch release on February 2nd 2025 we ve discovered two implementation details that greatly improved quality In both cases multilingual embeddings and rerankers are used in an in context manner operating at a much smaller scale than the traditional pre computed indices these models typically require [jina.ai](https://jina.ai/news/snippet-selection-and-url-ranking-in-deepsearch-deepresearch)",
"type": "text",
"annotations": [
{
"type": "url_citation",
"url_citation": {
"title": "Snippet Selection and URL Ranking in DeepSearch/DeepResearch",
"exactQuote": "Since our DeepSearch release on February 2nd 2025, we've discovered two implementation details that greatly improved quality. In both cases, multilingual embeddings and rerankers are used in an _\"in-context\"_ manner - operating at a much smaller scale than the traditional pre-computed indices these models typically require.",
"url": "https://jina.ai/news/snippet-selection-and-url-ranking-in-deepsearch-deepresearch",
"dateTime": "2025-03-13 06:48:01"
}
}
]
},
"logprobs": null,
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 169670,
"completion_tokens": 27285,
"total_tokens": 196526
},
"visitedURLs": [
"https://github.com/jina-ai/node-DeepResearch/blob/main/src/utils/url-tools.ts",
"https://huggingface.co/jinaai/jina-embeddings-v3",
"https://github.com/jina-ai/reader",
"https://zilliz.com/blog/training-text-embeddings-with-jina-ai",
"https://threads.net/@unwind_ai/post/DGmhWCVswbe/media",
"https://twitter.com/JinaAI_/status/1899840196507820173",
"https://jina.ai/news?tag=tech-blog",
"https://docs.llamaindex.ai/en/stable/examples/embeddings/jinaai_embeddings",
"https://x.com/jinaai_",
"https://x.com/JinaAI_/status/1899840202358784170",
"https://tracxn.com/d/companies/jina-ai/__IQ81fOnU0FsDpagFjG-LrG0DMWHELqI6znTumZBQF-A/funding-and-investors",
"https://jina.ai/models",
"https://linkedin.com/posts/imohitmayank_jinaai-has-unveiled-the-ultimate-developer-activity-7300401711242711040-VD64",
"https://medium.com/@tossy21/trying-out-jina-ais-node-deepresearch-c5b55d630ea6",
"https://huggingface.co/jinaai/jina-clip-v2",
"https://arxiv.org/abs/2409.10173",
"https://milvus.io/docs/embed-with-jina.md",
"https://seedtable.com/best-startups-in-china",
"https://threads.net/@sung.kim.mw/post/DGhG-J_vREu/jina-ais-a-practical-guide-to-implementing-deepsearchdeepresearchthey-cover-desi",
"https://elastic.co/search-labs/blog/jina-ai-embeddings-rerank-model-open-inference-api",
"http://status.jina.ai/",
"https://apidog.com/blog/recreate-openai-deep-research",
"https://youtube.com/watch?v=QxHE4af5BQE",
"https://sdxcentral.com/articles/news/cisco-engages-businesses-on-ai-strategies-at-greater-bay-area-2025/2025/02",
"https://aws.amazon.com/blogs/machine-learning/build-rag-applications-using-jina-embeddings-v2-on-amazon-sagemaker-jumpstart",
"https://reddit.com/r/perplexity_ai/comments/1ejbdqa/fastest_open_source_ai_search_engine",
"https://search.jina.ai/",
"https://sebastian-petrus.medium.com/build-openais-deep-research-open-source-alternative-4f21aed6d9f0",
"https://medium.com/@elmo92/jina-reader-transforming-web-content-to-feed-llms-d238e827cc27",
"https://openai.com/index/introducing-deep-research",
"https://python.langchain.com/docs/integrations/tools/jina_search",
"https://varindia.com/news/meta-is-in-talks-for-usd200-billion-ai-data-center-project",
"https://varindia.com/news/Mira-Murati%E2%80%99s-new-AI-venture-eyes-$9-billion-valuation",
"https://53ai.com/news/RAG/2025031401342.html",
"https://arxiv.org/abs/2409.04701",
"https://bigdatawire.com/this-just-in/together-ai-raises-305m-series-b-to-power-ai-model-training-and-inference",
"https://github.blog/",
"https://cdn-uploads.huggingface.co/production/uploads/660c3c5c8eec126bfc7aa326/MvwT9enRT7gOESHA_tpRj.jpeg",
"https://cdn-uploads.huggingface.co/production/uploads/660c3c5c8eec126bfc7aa326/JNs_DrpFbr6ok_pSRUK4j.jpeg",
"https://app.dealroom.co/lists/33530",
"https://api-docs.deepseek.com/news/news250120",
"https://sdxcentral.com/articles/news/ninjaone-raises-500-million-valued-at-5-billion/2025/02",
"https://linkedin.com/sharing/share-offsite?url=https%3A%2F%2Fjina.ai%2Fnews%2Fa-practical-guide-to-implementing-deepsearch-deepresearch%2F",
"https://twitter.com/intent/tweet?url=https%3A%2F%2Fjina.ai%2Fnews%2Fa-practical-guide-to-implementing-deepsearch-deepresearch%2F",
"https://platform.openai.com/docs/api-reference/chat/create",
"https://mp.weixin.qq.com/s/-pPhHDi2nz8hp5R3Lm_mww",
"https://huggingface.us17.list-manage.com/subscribe?id=9ed45a3ef6&u=7f57e683fa28b51bfc493d048",
"https://automatio.ai/",
"https://sdk.vercel.ai/docs/introduction",
"https://app.eu.vanta.com/jinaai/trust/vz7f4mohp0847aho84lmva",
"https://apply.workable.com/huggingface/j/AF1D4E3FEB",
"https://facebook.com/sharer/sharer.php?u=https%3A%2F%2Fjina.ai%2Fnews%2Fa-practical-guide-to-implementing-deepsearch-deepresearch%2F",
"https://facebook.com/sharer/sharer.php?u=http%3A%2F%2F127.0.0.1%3A3000%2Fen-US%2Fnews%2Fsnippet-selection-and-url-ranking-in-deepsearch-deepresearch%2F",
"https://reddit.com/submit?url=https%3A%2F%2Fjina.ai%2Fnews%2Fa-practical-guide-to-implementing-deepsearch-deepresearch%2F",
"https://apply.workable.com/huggingface",
"https://news.ycombinator.com/submitlink?u=https%3A%2F%2Fjina.ai%2Fnews%2Fa-practical-guide-to-implementing-deepsearch-deepresearch%2F",
"https://news.ycombinator.com/submitlink?u=http%3A%2F%2F127.0.0.1%3A3000%2Fen-US%2Fnews%2Fsnippet-selection-and-url-ranking-in-deepsearch-deepresearch%2F",
"https://docs.github.com/site-policy/privacy-policies/github-privacy-statement",
"https://discord.jina.ai/",
"https://docs.github.com/site-policy/github-terms/github-terms-of-service",
"https://bigdatawire.com/this-just-in/qumulo-announces-30-million-funding",
"https://x.ai/blog/grok-3",
"https://m-ric-open-deep-research.hf.space/",
"https://youtu.be/sal78ACtGTc?feature=shared&t=52",
"https://mp.weixin.qq.com/s/apnorBj4TZs3-Mo23xUReQ",
"https://perplexity.ai/hub/blog/introducing-perplexity-deep-research",
"https://githubstatus.com/",
"https://github.blog/changelog/2021-09-30-footnotes-now-supported-in-markdown-fields",
"https://openai.com/index/introducing-operator",
"mailto:[email protected]",
"https://resources.github.com/learn/pathways",
"https://status.jina.ai/",
"https://reuters.com/technology/artificial-intelligence/tencents-messaging-app-weixin-launches-beta-testing-with-deepseek-2025-02-16",
"https://scmp.com/tech/big-tech/article/3298981/baidu-adopts-deepseek-ai-models-chasing-tencent-race-embrace-hot-start",
"https://microsoft.com/en-us/research/articles/magentic-one-a-generalist-multi-agent-system-for-solving-complex-tasks",
"javascript:UC_UI.showSecondLayer();",
"https://resources.github.com/",
"https://storm-project.stanford.edu/research/storm",
"https://blog.google/products/gemini/google-gemini-deep-research",
"https://youtu.be/vrpraFiPUyA",
"https://chat.baidu.com/search?extParamsJson=%7B%22enter_type%22%3A%22ai_explore_home%22%7D&isShowHello=1&pd=csaitab&setype=csaitab&usedModel=%7B%22modelName%22%3A%22DeepSeek-R1%22%7D",
"https://app.dover.com/jobs/jinaai",
"http://localhost:3000/",
"https://docs.cherry-ai.com/",
"https://en.wikipedia.org/wiki/Delayed_gratification",
"https://support.github.com/?tags=dotcom-footer",
"https://docs.jina.ai/",
"https://skills.github.com/",
"https://partner.github.com/",
"https://help.x.com/resources/accessibility",
"https://business.twitter.com/en/help/troubleshooting/how-twitter-ads-work.html",
"https://business.x.com/en/help/troubleshooting/how-twitter-ads-work.html",
"https://support.twitter.com/articles/20170514",
"https://support.x.com/articles/20170514",
"https://t.co/jnxcxPzndy",
"https://t.co/6EtEMa9P05",
"https://help.x.com/using-x/x-supported-browsers",
"https://legal.twitter.com/imprint.html"
],
"readURLs": [
"https://jina.ai/news/a-practical-guide-to-implementing-deepsearch-deepresearch",
"https://github.com/jina-ai/node-DeepResearch",
"https://huggingface.co/blog/open-deep-research",
"https://jina.ai/news/snippet-selection-and-url-ranking-in-deepsearch-deepresearch",
"https://x.com/jinaai_?lang=en",
"https://jina.ai/news",
"https://x.com/joedevon/status/1896984525210837081",
"https://github.com/jina-ai/node-DeepResearch/blob/main/src/tools/jina-latechunk.ts"
],
"numURLs": 98
}
深度搜索參數指南
質量控制
在深度搜索中,通常存在一個權衡:系統執行的步驟越多,獲得的結果質量就越高,但同時也會消耗更多詞元。這種質量的提升源於更廣泛、更詳盡的搜索和更深入的反思。四個主要參數控制深度搜索的質量:budget_tokens
、max_attempts
、team_size
和 reasoning_effort
。reasoning_effort
參數本質上是 budget_tokens
和 max_attempts
的預設組合,並經過了精心調整。對於大多數用户來説,調整 reasoning_effort
是最簡單的方法。
預算詞元
budget_tokens
設置整個深度搜索過程中允許的最大詞元數量。這涵蓋了所有操作,包括網頁搜索、讀取網頁、反思、摘要和編碼。預算越大,響應質量自然也就越好。當預算耗盡或找到滿意答案時(以先到者為準),深度搜索過程將停止。如果預算先耗盡,您仍然會得到答案,但這可能不是最終的、完全優化的響應,因為它尚未通過max_attempts
定義的所有質量檢查。
最大嘗試次數
max_attempts
決定了系統在深度搜索過程中重試解決問題的次數。每次深度搜索生成答案時,它都必須通過內部評估器定義的某些質量測試。如果答案未通過這些測試,評估器會提供反饋,系統會使用此反饋繼續搜索和優化答案。將max_attempts
設置得太低意味着您可以快速獲得結果,但質量可能會受到影響,因為答案可能無法通過所有質量檢查。將其設置得太高可能會使流程陷入無休止的重試循環,不斷嘗試並失敗。
當budget_tokens
或max_attempts
超出(以先發生者為準),或者答案通過所有測試且仍有剩餘預算和嘗試次數時,系統會返回最終答案。
團隊規模
team_size
對質量的影響方式與max_attempts
和budget_tokens
截然不同。當team_size
設置為大於1時,系統會將原始問題分解為子問題,並分別進行研究。可以將其想象成Map-Reduce模式,將大型作業分解為並行運行的較小任務。最終答案是每個工作者結果的綜合。我們稱之為team_size
,因為它模擬了一個研究團隊,其中多個智能體調查同一問題的不同方面並協作完成最終報告。
請記住,所有智能體的詞元消耗都會計入您的budget_tokens
總數,但每個智能體都有獨立的max_attempts
。這意味着,如果team_size
較大,但budget_tokens
相同,由於預算限制,智能體可能會比預期更快地返回答案。我們建議同時增加team_size
和budget_tokens
,以便為每個智能體提供足夠的資源來完成全面的工作。
最後,您可以將team_size
視為控制搜索廣度的因素——它決定了要研究的不同方面的數量。同時,budget_tokens
和max_attempts
控制搜索的深度——即對每個方面的探索程度。
信息來源控制
深度搜索高度依賴基礎信息——它所使用的信息來源。質量不僅僅關乎算法的深度和廣度;深度搜索獲取信息的來源也同樣重要,而且往往是決定性因素。讓我們來探索一下控制這一點的關鍵參數。
不直接回答
no_direct_answer
是一個簡單的開關,它可以阻止系統在步驟 1 返回答案。啓用後,它會禁用系統使用內部知識的能力,並強制其始終首先搜索網頁。啓用此功能會使系統“過度思考”,即使是像“今天星期幾”、“你好嗎”這樣的簡單問題,或者像“誰是美國第 40 任總統”這樣的基本事實知識,這些知識肯定存在於模型的訓練數據中。
域名控制
三個參數——boost_hostnames
、bad_hostnames
和 only_hostnames
——告訴 深度搜索哪些網頁需要優先處理、避免或專門使用。要理解這些機制的工作原理,請回顧一下深度搜索中的搜索和讀取流程:
- 搜索階段:系統搜索網絡並檢索包含其摘要的網站 URL 列表
- 選擇階段:系統決定實際訪問哪些 URL(由於時間和成本限制,系統不會訪問所有 URL)
boost_hostnames
:此處列出的域名優先級更高,更有可能被訪問bad_hostnames
:這些域名永遠不會被訪問only_hostnames
:定義後,只有與這些域名匹配的 URL 才會被訪問
以下是一些關於域名參數的重要説明。首先,系統始終使用搜索引擎返回的摘要作為構建推理鏈的初始線索。這些域名參數僅影響系統訪問的網頁,而不會影響其如何制定搜索查詢。
其次,如果收集到的 URL 不包含 only_hostnames
中指定的域名,系統可能會完全停止讀取網頁。我們建議僅在您熟悉研究問題並瞭解潛在答案可能出現的位置(或絕對不應該出現的位置)時才使用這些參數。
特殊情況:學術研究
對於學術研究,您可能希望將搜索和讀取限制在 arxiv.org 上。在這種情況下,只需設置 "search_provider": "arxiv"
,所有內容都將以 arxiv 作為唯一來源。但是,對於一般性或瑣碎的問題,此限制可能無法獲得有效的答案,因此僅將 "search_provider": "arxiv"
用於嚴肅的學術研究。
搜索語言代碼
search_language_code
是另一個影響網絡資源的參數,它會強制系統以特定語言生成查詢,而不管原始輸入或中間推理步驟如何。通常,系統會自動確定查詢語言以獲得最佳搜索覆蓋率,但有時手動控制也很有用。
語言控制用例
國際市場調研:在研究本地品牌或公司在國際市場的影響力時,您可以強制查詢始終使用英語(使用 "search_language_code": "en"
來實現全球覆蓋),或者使用本地語言來獲取更具針對性的地區信息。
使用非英語提示進行全球調研:如果您的輸入始終使用中文或日語(因為您的最終用户主要使用這些語言),但您的調研範圍是全球性的,而不僅僅是本地的中文或日語網站,則系統可能會自動傾向於使用您提示的語言。使用此參數可以強制使用英語查詢,以實現更廣泛的國際覆蓋率。
與深度搜索聊天
什麼是深度搜索?
大模型
RAG範式和帶搜索的大模型
深度搜索
API價格表
產品 | API端口 | 描述arrow_upward | 無 API 密鑰key_off | 使用 API 密鑰key | 帶有高級 API 密鑰key | 平均延遲 | 詞元使用計數 | 請求類型 | |
---|---|---|---|---|---|---|---|---|---|
讀取器 API | https://r.jina.ai | 將 URL 轉換為大模型友好文本 | 20 RPM | 500 RPM | trending_up5000 RPM | 7.9s | 以輸出響應中的詞元數量為準。 | GET/POST | |
讀取器 API | https://s.jina.ai | 搜索網絡並將結果轉換為大模型友好文本 | block | 100 RPM | trending_up1000 RPM | 2.5s | 每個請求都需要固定數量的詞元,從 10000 個詞元開始 | GET/POST | |
深度搜索 | https://deepsearch.jina.ai/v1/chat/completions | 推理、搜索和迭代以找到最佳答案 | block | 50 RPM | 500 RPM | 56.7s | 統計整個過程中詞元的總數。 | POST | |
向量模型API | https://api.jina.ai/v1/embeddings | 將文本/圖片轉為定長向量 | block | 500 RPM & 1,000,000 TPM | trending_up2,000 RPM & 5,000,000 TPM | ssid_chart 取決於輸入大小 help | 以輸入請求中的詞元數量為準。 | POST | |
重排器 API | https://api.jina.ai/v1/rerank | 按查詢對文檔進行精排 | block | 500 RPM & 1,000,000 TPM | trending_up2,000 RPM & 5,000,000 TPM | ssid_chart 取決於輸入大小 help | 以輸入請求中的詞元數量為準。 | POST | |
分類器 API | https://api.jina.ai/v1/train | 使用訓練樣本訓練分類器 | block | 20 RPM & 200,000 TPM | 60 RPM & 1,000,000 TPM | ssid_chart 取決於輸入大小 | 詞元計數為:輸入詞元 × 迭代次數 | POST | |
分類器 API (少量樣本) | https://api.jina.ai/v1/classify | 使用經過訓練的少樣本分類器對輸入進行分類 | block | 20 RPM & 200,000 TPM | 60 RPM & 1,000,000 TPM | ssid_chart 取決於輸入大小 | 詞元計數為:輸入詞元 | POST | |
分類器 API (零樣本) | https://api.jina.ai/v1/classify | 使用零樣本分類對輸入進行分類 | block | 200 RPM & 500,000 TPM | 1,000 RPM & 3,000,000 TPM | ssid_chart 取決於輸入大小 | 詞元計數為:輸入詞元 加 標籤詞元 | POST | |
切分器 API | https://api.jina.ai/v1/segment | 對長文本進行分詞分句 | 20 RPM | 200 RPM | 1,000 RPM | 0.3s | 詞元不計算使用量。 | GET/POST |