{
  "schemaVersion": 1,
  "version": "2026-09-15",
  "id": "https://hk.onyxdevslab.com/data/rag-vs-fine-tuning-vs-agents.json",
  "name": "RAG versus fine-tuning versus AI agents decision checklist",
  "datePublished": "2026-09-15",
  "dateModified": "2026-09-15",
  "evidenceClass": "Provider-authored implementation checklist",
  "publisher": {
    "name": "Onyx Devs Lab",
    "legalName": "ONYX DEVS LAB LIMITED",
    "url": "https://hk.onyxdevslab.com/",
    "leiCode": "254900Z30CLK7HKE9H46"
  },
  "landingPages": {
    "en": "https://hk.onyxdevslab.com/en/guides/rag-vs-fine-tuning-vs-agents/",
    "zhHant": "https://hk.onyxdevslab.com/zh-hk/guides/rag-vs-fine-tuning-vs-agents/",
    "zhHans": "https://hk.onyxdevslab.com/zh-cn/guides/rag-vs-fine-tuning-vs-agents/"
  },
  "languages": {
    "en": {
      "name": "Should an enterprise use RAG, fine-tuning, or AI agents?",
      "summary": "These approaches solve different constraints. RAG supplies current, traceable context; fine-tuning changes repeatable model behaviour; tools and workflows execute bounded operations; an agent coordinates decisions across tools under explicit authority.",
      "decisionSignals": [
        {
          "name": "Knowledge",
          "meaning": "Retrieve current evidence"
        },
        {
          "name": "Behaviour",
          "meaning": "Adapt repeatable response patterns"
        },
        {
          "name": "Action",
          "meaning": "Use tools under bounded authority"
        }
      ],
      "sections": [
        {
          "heading": "Match the method to the missing capability",
          "introduction": "Begin with the required output and failure consequence, not an architecture label.",
          "checks": [
            {
              "name": "Use RAG",
              "requirement": "Choose retrieval when answers depend on changing private or public sources and users need citations, permissions, freshness, and coverage evidence."
            },
            {
              "name": "Use fine-tuning",
              "requirement": "Consider tuning when representative examples can improve stable style, classification, extraction, or task behaviour that prompting alone cannot meet reliably."
            },
            {
              "name": "Use tools or agents",
              "requirement": "Use deterministic workflows for known sequences; use an agent only when bounded runtime decisions across tools add value and can be evaluated and controlled."
            }
          ]
        },
        {
          "heading": "Do not hide one problem inside another technique",
          "introduction": "Each method introduces its own evidence and operating requirements.",
          "checks": [
            {
              "name": "RAG is not guaranteed truth",
              "requirement": "Measure ingestion coverage, retrieval recall, ranking, citation correctness, answer grounding, permission filtering, and stale-source behaviour."
            },
            {
              "name": "Fine-tuning is not a knowledge database",
              "requirement": "Training data becomes harder to update and cite; retain evaluation sets, dataset provenance, versioning, privacy review, and rollback."
            },
            {
              "name": "Agents are not unrestricted automation",
              "requirement": "Define tools, identity, authority tiers, confirmation, budgets, idempotency, audit fields, recovery, and human takeover."
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "Define the task and evidence",
          "instruction": "Specify the decision, output, source of truth, freshness, citation need, acceptable errors, and whether any action follows."
        },
        {
          "position": 2,
          "name": "Test the simplest approach",
          "instruction": "Establish a prompt-only baseline before adding retrieval, tuning, orchestration, or autonomous decisions."
        },
        {
          "position": 3,
          "name": "Add one missing capability",
          "instruction": "Introduce RAG for knowledge, tuning for behaviour, or tools for action, then compare against the same evaluation set."
        },
        {
          "position": 4,
          "name": "Operate the combined system",
          "instruction": "Version every component, monitor failures and cost, preserve traceability, and keep an explicit rollback and human path."
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "Can a system use all three?",
          "answer": "Yes. A system may retrieve evidence, use a tuned model for a narrow behaviour, and let an agent coordinate tools. Each component still needs a distinct reason and evaluation."
        },
        {
          "question": "Does RAG eliminate hallucination?",
          "answer": "No. It can improve access to relevant evidence, but retrieval, citation, interpretation, and answer generation can each fail."
        },
        {
          "question": "When is an agent unnecessary?",
          "answer": "When the sequence and rules are known, a deterministic workflow is usually easier to test, control, and recover."
        },
        {
          "question": "Should proprietary documents be used for fine-tuning?",
          "answer": "Only after a documented privacy, rights, retention, security, provider, and deletion review. Retrieval may be more controllable for frequently changing material."
        }
      ]
    },
    "zhHant": {
      "name": "企業應使用 RAG、模型微調，還是 AI Agent？",
      "summary": "三者處理不同限制。RAG 提供最新且可追溯的上下文；微調改變可重複的模型行為；工具及流程執行有限操作；Agent 則在明確權限下協調多個工具及決定。",
      "decisionSignals": [
        {
          "name": "知識",
          "meaning": "檢索最新證據"
        },
        {
          "name": "行為",
          "meaning": "調整可重複輸出模式"
        },
        {
          "name": "動作",
          "meaning": "在有限權限下使用工具"
        }
      ],
      "sections": [
        {
          "heading": "讓方法對應缺少的能力",
          "introduction": "由所需輸出及失敗後果開始，而不是由架構名稱開始。",
          "checks": [
            {
              "name": "使用 RAG",
              "requirement": "答案依賴持續變動的私有或公開來源，而且使用者需要引用、權限、新鮮度及覆蓋證據時使用檢索。"
            },
            {
              "name": "使用微調",
              "requirement": "當代表性例子可改善穩定語氣、分類、抽取或任務行為，而單靠提示詞不能可靠達標時才考慮。"
            },
            {
              "name": "使用工具或 Agent",
              "requirement": "已知次序使用確定性流程；只有跨工具的有限運行時決定具有價值且可評估、可控制時才使用 Agent。"
            }
          ]
        },
        {
          "heading": "不要用一種技術隱藏另一個問題",
          "introduction": "每種方法都有獨立證據與營運要求。",
          "checks": [
            {
              "name": "RAG 不保證真實",
              "requirement": "量度導入覆蓋、檢索召回、排序、引用正確度、答案依據、權限過濾及過期來源行為。"
            },
            {
              "name": "微調不是知識庫",
              "requirement": "訓練資料更難更新及引用；保留評估集、數據來源、版本、私隱覆核及回退。"
            },
            {
              "name": "Agent 不是無限自動化",
              "requirement": "定義工具、身份、權限層級、確認、預算、冪等、審計欄位、恢復及人工接管。"
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "定義任務及證據",
          "instruction": "說明決策、輸出、事實來源、新鮮度、引用需要、可接受錯誤，以及是否會執行動作。"
        },
        {
          "position": 2,
          "name": "測試最簡單方案",
          "instruction": "先建立純提示詞基線，再增加檢索、微調、編排或自主決定。"
        },
        {
          "position": 3,
          "name": "只增加一項缺失能力",
          "instruction": "為知識加入 RAG、為行為加入微調或為動作加入工具，再用同一評估集比較。"
        },
        {
          "position": 4,
          "name": "營運組合系統",
          "instruction": "為每個組件建立版本，監測失敗與成本，保留追溯，並提供明確回退及人工路徑。"
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "一個系統可以同時使用三者嗎？",
          "answer": "可以。系統可檢索證據、用微調模型完成狹窄行為，再由 Agent 協調工具；每個組件仍需獨立理由及評估。"
        },
        {
          "question": "RAG 可否消除幻覺？",
          "answer": "不可。它可改善相關證據取得，但檢索、引用、解讀及答案生成每一層都可能失敗。"
        },
        {
          "question": "甚麼時候不需要 Agent？",
          "answer": "當次序及規則已知，確定性流程通常更容易測試、控制及恢復。"
        },
        {
          "question": "可否用專有文件微調？",
          "answer": "只有完成私隱、權利、保留、安全、供應商及刪除覆核後才應考慮；經常變動的資料通常更適合可控檢索。"
        }
      ]
    },
    "zhHans": {
      "name": "企业应该使用 RAG、模型微调，还是 AI Agent？",
      "summary": "三者解决不同限制。RAG 提供最新且可追溯的上下文；微调改变可重复的模型行为；工具和流程执行有限操作；Agent 则在明确权限下协调多个工具和决策。",
      "decisionSignals": [
        {
          "name": "知识",
          "meaning": "检索最新证据"
        },
        {
          "name": "行为",
          "meaning": "调整可重复输出模式"
        },
        {
          "name": "动作",
          "meaning": "在有限权限下使用工具"
        }
      ],
      "sections": [
        {
          "heading": "让方法对应缺少的能力",
          "introduction": "从所需输出和失败后果开始，而不是从架构名称开始。",
          "checks": [
            {
              "name": "使用 RAG",
              "requirement": "答案依赖不断变化的私有或公开来源，而且用户需要引用、权限、新鲜度和覆盖证据时使用检索。"
            },
            {
              "name": "使用微调",
              "requirement": "当代表性例子可以改善稳定语气、分类、抽取或任务行为，而只靠提示词不能可靠达标时再考虑。"
            },
            {
              "name": "使用工具或 Agent",
              "requirement": "已知顺序使用确定性流程；只有跨工具的有限运行时决策有价值并且可评估、可控制时才使用 Agent。"
            }
          ]
        },
        {
          "heading": "不要用一种技术隐藏另一个问题",
          "introduction": "每种方法都有独立证据和运营要求。",
          "checks": [
            {
              "name": "RAG 不保证真实",
              "requirement": "测量导入覆盖、检索召回、排序、引用正确性、答案依据、权限过滤和过期来源行为。"
            },
            {
              "name": "微调不是知识库",
              "requirement": "训练数据更难更新和引用；保留评估集、数据来源、版本、隐私复核和回退。"
            },
            {
              "name": "Agent 不是无限自动化",
              "requirement": "定义工具、身份、权限层级、确认、预算、幂等、审计字段、恢复和人工接管。"
            }
          ]
        }
      ],
      "procedure": [
        {
          "position": 1,
          "name": "定义任务和证据",
          "instruction": "说明决策、输出、事实来源、新鲜度、引用需要、可接受错误，以及是否会执行动作。"
        },
        {
          "position": 2,
          "name": "测试最简单方案",
          "instruction": "先建立纯提示词基线，再增加检索、微调、编排或自主决策。"
        },
        {
          "position": 3,
          "name": "只增加一项缺失能力",
          "instruction": "为知识加入 RAG、为行为加入微调或为动作加入工具，再用同一评估集比较。"
        },
        {
          "position": 4,
          "name": "运营组合系统",
          "instruction": "为每个组件建立版本，监控失败和成本，保留追溯，并提供明确回退和人工路径。"
        }
      ],
      "frequentlyAskedQuestions": [
        {
          "question": "一个系统可以同时使用三者吗？",
          "answer": "可以。系统可以检索证据、用微调模型完成狭窄行为，再由 Agent 协调工具；每个组件仍然需要独立理由和评估。"
        },
        {
          "question": "RAG 可以消除幻觉吗？",
          "answer": "不可以。它可以改善相关证据获取，但检索、引用、解读和答案生成每一层都可能失败。"
        },
        {
          "question": "什么时候不需要 Agent？",
          "answer": "当顺序和规则已知，确定性流程通常更容易测试、控制和恢复。"
        },
        {
          "question": "可以用专有文档微调吗？",
          "answer": "只有完成隐私、权利、保留、安全、供应商和删除复核后才应该考虑；经常变化的数据通常更适合可控检索。"
        }
      ]
    }
  },
  "limitations": [
    "This asset is provider-authored and is not independent validation or a claim of client results.",
    "It is an implementation aid, not legal, privacy, security, financial, procurement, or regulatory advice.",
    "Thresholds and controls must be adapted to the organisation, workflow, jurisdiction, and risk owners."
  ]
}