開發人員可透過 AI 應用程式,運用 Google 的基礎模型、搜尋專業知識和對話式 AI 技術,打造企業級生成式 AI 應用程式。這項技術仍處於早期階段,功能和用途仍在演進中,因此可能遭到誤用、濫用,並產生意料之外或無法預測的後果。舉例來說,AI 應用程式可能會生成預期外的輸出內容,包括令人反感、未顧及感受或與事實不符的文字。
考量到這些風險和複雜性,AI 應用程式的設計依據是 Google 的 AI 原則。不過,開發人員必須瞭解及測試模型,才能安全且負責任地部署。為協助開發人員,AI Applications 內建安全篩選器,可協助客戶在用途中封鎖潛在有害的輸出內容。詳情請參閱「Vertex AI Search 的安全設定」。
將 AI 應用程式整合至客戶的專屬用途和情境時,可能需要考量其他負責任的 AI 考量事項和模型限制。我們建議客戶採用公平性、可解釋性、隱私權和安全性建議做法。
Vertex AI Search 的安全設定
Vertex AI Search 提供安全搜尋篩選器,可過濾情色露骨的網站搜尋結果。包括色情、暴力與血腥內容。安全搜尋過濾功能不適用於私人內容。
安全搜尋功能可套用至 Vertex AI Search 建立索引的公開內容,例如網站。開啟安全搜尋功能後,無論顧客是查詢圖片、影片還是網站,Vertex AI Search 都會將煽情露骨內容從顧客的 Vertex AI Search 結果中過濾掉。雖然這些篩選器無法百分之百準確,但我們會持續評估及更新篩選功能。
在特定情境和用途中,客戶可能需要存取各種資訊 (包括露骨內容),才能從 Vertex AI Search 收集全面深入的洞察資料和結果。客戶可以透過 API 將 safe_search 設為「true」或「false」,啟用及停用安全篩選器。
safe_search 設定會套用至搜尋結果,以及 Vertex AI Search 輸出內容,例如摘要 (摘要內容會根據搜尋結果產生)。
系統會根據一份安全屬性清單,評估 Vertex AI Search 的輸出內容 (例如摘要),即使 safe_search 處於關閉狀態,也會篩除可能有害的內容。
私人內容摘要和網站搜尋
系統會根據一系列安全屬性評估 Vertex AI Search 的輸出內容 (例如摘要),篩除可能有害的內容,例如貶抑、情色露骨、惡意或暴力資訊,即使關閉網站搜尋的 safe_search 也是如此。此外,您也可以在 Vertex AI Search API 回應中使用 safetyAttributes,篩除包含「有害類別」的回覆,以及可能被視為敏感的主題。
備用回覆
如果摘要模型對要求的回覆是空白字串或一般回覆,表示輸入內容或輸出內容觸發了安全防護篩選器。在 Vertex AI Search API 回應中,如果偵測到可能違反政策的情況,SummarySkippedReason 會設為 BLOCKED_DUE_TO_POTENTIAL_POLICY_VIOLATION。
[[["容易理解","easyToUnderstand","thumb-up"],["確實解決了我的問題","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["難以理解","hardToUnderstand","thumb-down"],["資訊或程式碼範例有誤","incorrectInformationOrSampleCode","thumb-down"],["缺少我需要的資訊/範例","missingTheInformationSamplesINeed","thumb-down"],["翻譯問題","translationIssue","thumb-down"],["其他","otherDown","thumb-down"]],["上次更新時間:2025-09-05 (世界標準時間)。"],[[["\u003cp\u003eVertex AI Agent Builder utilizes Google's AI technology to create generative AI applications, but developers should be aware of potential risks like generating offensive, incorrect, or unintended content.\u003c/p\u003e\n"],["\u003cp\u003eGoogle's AI Principles guide Vertex AI Agent Builder's design, and built-in safety filters are included to help block harmful outputs, although these are not 100% accurate.\u003c/p\u003e\n"],["\u003cp\u003eVertex AI Search offers SafeSearch filters for public content to help filter out explicit materials, with the option to enable or disable these filters via API settings.\u003c/p\u003e\n"],["\u003cp\u003eVertex AI Search outputs, such as summarization, are automatically filtered for harmful content, including derogatory, sexually explicit, toxic, or violent information, even if SafeSearch is turned off.\u003c/p\u003e\n"],["\u003cp\u003eDevelopers are encouraged to follow Responsible AI best practices and be mindful of the model's limitations, especially when using private data, while also recognizing that human validation and oversight are crucial.\u003c/p\u003e\n"]]],[],null,["# Responsible AI\n\nAI Applications lets developers tap into the power of Google's foundation\nmodels, search expertise, and conversational AI technologies to create\nenterprise-grade generative AI applications.\nAs an early-stage technology, its\nevolving capabilities and uses create potential for misapplication, misuse, and\nunintended or unforeseen consequences. For example, AI Applications can\ngenerate output that you don't expect, including text that's offensive,\ninsensitive, or factually incorrect.\n\nGiven these risks and complexities, AI Applications is designed with\nGoogle's [AI Principles](https://ai.google/principles/) in mind. However, it is\nimportant for developers to understand and test their models to deploy safely\nand responsibly. To aid developers, AI Applications has built-in safety\nfilters to help customers block potentially harmful outputs within their use\ncase. For more information, see [Safety settings for Vertex AI Search](#safety-settings-search).\n\nWhen AI Applications is integrated into a customer's unique use case and\ncontext, additional responsible AI considerations and model limitations might need\nto be considered. We encourage customers to leverage fairness, interpretability,\nprivacy, and security [recommended\npractices](https://ai.google/responsibilities/responsible-ai-practices/).\n\nSafety settings for Vertex AI Search\n------------------------------------\n\nVertex AI Search offers SafeSearch filters to filter explicit website\nsearch results. Explicit results might include content like pornography, violence,\nand gore. SafeSearch filtering is not available for private content.\n\nSummarization (both for website search and private content) filters out\nresponses such as derogatory, sexually explicit, toxic, or violent information.\nAdditionally, the summarization response contains safety attributes, which\ninclude \"harmful categories\" and topics that might be considered sensitive.\n\n### Website search\n\nSafeSearch can be applied to public content indexed by\nVertex AI Search, such as\n[websites](/generative-ai-app-builder/docs/enterprise-search-introduction#websites). When SafeSearch is on,\nVertex AI Search helps filter out explicit content in a customer's\nVertex AI Search results for all queries across images, videos, and\nwebsites. While those filters are not 100% accurate, we continuously evaluate\nand update filtering capabilities.\n\nIn certain contexts and use cases, customers might require access to a wide range\nof information, even if it includes explicit content, to gather comprehensive\ninsights and results from Vertex AI Search. Customers can enable and\ndisable safety filters using the API by setting `safe_search`\nto \"true\" or \"false\".\n\nTo enable SafeSearch for Vertex AI Search, use the following API\ncall: \n\n curl -X POST\\\n -H \"Authorization: Bearer $(gcloud auth application-default print-access-token)\"\n -H \"Content-Type: application/json\"\\\n \"https://discoveryengine.googleapis.com/v1/projects/\u003cvar translate=\"no\"\u003ePROJECT_ID\u003c/var\u003e/locations/global/collections/default_collection/dataStores/\u003cvar translate=\"no\"\u003eDATA_STORE_ID\u003c/var\u003e/servingConfigs/default_config:search\"\n -d '{\n \"query\": { \"input\": \"\u003cvar translate=\"no\"\u003eFREE_TEXT\u003c/var\u003e\"},\n \"serving_config\": \"\u003cvar translate=\"no\"\u003eSERVING_CONFIG_ID\u003c/var\u003e\",\n \"safe_search\": \"true\",\n }'\n\nTo enable SafeSearch for Search with follow-ups in Vertex AI Search,\nuse the following API call: \n\n curl -X POST \\\n -H \"Authorization: Bearer $(gcloud auth print-access-token)\" \\\n -H \"Content-Type: application/json\" \\\n \"https://discoveryengine.googleapis.com/v1beta/projects/\u003cvar translate=\"no\"\u003ePROJECT_ID\u003c/var\u003e/locations/global/collections/default_collection/dataStores/\u003cvar translate=\"no\"\u003eDATA_STORE_ID\u003c/var\u003e/conversations/\u003cvar translate=\"no\"\u003eCONVERSATION_ID\u003c/var\u003e:converse\" \\\n -d '{\n \"query\": { \"input\": \"\u003cvar translate=\"no\"\u003eFREE_TEXT\u003c/var\u003e\"},\n \"serving_config\": \"\u003cvar translate=\"no\"\u003eSERVING_CONFIG_ID\u003c/var\u003e\",\n \"safe_search\": \"true\",\n }'\n\nThe `safe_search` setting applies to search results and also to\nVertex AI Search outputs like summarization (which is based on the search\nresults).\n\nVertex AI Search outputs, such as summarization, are assessed against a list of\nsafety attributes which filter potentially harmful content even when\n`safe_search` is turned off.\n\n### Summarization for private content and website search\n\nVertex AI Search outputs, such as summarization, are assessed against a list of\nsafety attributes, which filters potentially harmful content such as derogatory,\nsexually explicit, toxic, or violent information, even when `safe_search` for\nwebsite search is turned off. Additionally, you can filter out responses which\ninclude \"harmful categories\" and topics that might be considered sensitive by\nusing `safetyAttributes` in the Vertex AI Search API response.\n\n### Fallback responses\n\nIf the summarization model responds to a request with an empty string or a\ngeneric response, it means that either the input or the output is triggering a\nsafety filter. In the Vertex AI Search API response,\n`SummarySkippedReason` is set to `BLOCKED_DUE_TO_POTENTIAL_POLICY_VIOLATION` if\na potential policy violation is detected.\n\nIf you think that a safety filter is being inappropriately\ntriggered, [report this issue](/support/docs/issue-trackers) via the\nGoogle Cloud console.\n\nAdditional resources\n--------------------\n\n- Learn more about Google's recommendations for [Responsible AI practices](https://ai.google/responsibility/responsible-ai-practices/)\n- Read our blog, [A shared agenda for responsible AI progress](https://blog.google/technology/ai/a-shared-agenda-for-responsible-ai-progress/)"]]