为 retrieve-and-generate RAG 评估作业创建提示数据集 - Amazon Bedrock
为 Amazon Bedrock 调用您的知识库的 retrieve-and-generate评估任务准备数据集使用您自己的推理响应数据为 retrieve-and-generate评估任务准备数据集

本文属于机器翻译版本。若本译文内容与英语原文存在差异，则一律以英文原文为准。

为 retrieve-and-generate RAG 评估作业创建提示数据集

retrieve-and-generate评估任务需要使用 JSON 行格式的提示数据集。数据集最多可包含 1000 条提示

为 Amazon Bedrock 调用您的知识库的 retrieve-and-generate评估任务准备数据集

要创建 Amazon Bedrock 在其中调用知识库的“仅检索”评测作业，您的提示数据集必须包含以下键值对：

  • referenceResponses – 此父键用于指定您期望 RetrieveAndGenerate 返回的基础事实响应。在 text 键中指定基础事实。如果您在评测作业中选择上下文覆盖程度指标，则 referenceResponses 是必需的。

  • prompt – 此父键用于指定在评测作业运行时，您希望模型进行响应的提示（用户查询）。

下面是一个包含 6 个输入并使用了 JSON 行格式的自定义数据集示例。

{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you want to use during inference"}]},"referenceResponses":[{"content":[{"text":"Specify a ground-truth response"}]}]}]}

为清楚起见，已展开以下提示。在实际提示数据集内，每一行（提示）都必须是一个有效的 JSON 对象。

{
    "conversationTurns": [
        {
            "prompt": {
                "content": [
                    {
                        "text": "What is the recommended service interval for your product?"
                    }
                ]
            },
            "referenceResponses": [
                {
                    "content": [
                        {
                            "text": "The recommended service interval for our product is two years."
                        }
                    ]
                }
            ]
        }
    ]
}

使用您自己的推理响应数据为 retrieve-and-generate评估任务准备数据集

要创建在其中提供自己的推理响应数据的 retrieve-and-generate评估作业，您的提示数据集是对话回合的列表，其中包含每回合的以下内容。每个作业只能评测一个 RAG 来源。

  • prompt – 您提供给模型以用于生成结果的提示。

  • referenceResponses – 此父键用于指定在 LLM 摄取检索结果和输入查询后，您期望其生成的最终输出所对应的基础事实响应。

  • referenceContexts（可选）– 此可选父键用于指定您希望从 RAG 来源中检索到的基础事实段落。如果您需要在自己的自定义评测指标中使用此键，只需包含此键即可。Amazon Bedrock 提供的内置指标不使用此属性。

  • output – 来自 RAG 来源的输出，包括以下内容：

    • text – 来自 RAG 系统中 LLM 的最终输出。

    • retrievedPassages – 此父键用于指定 RAG 来源检索到的内容。

您的 output 数据还必须包含字符串 knowledgeBaseIdentifier，该字符串定义用于生成推理响应的 RAG 来源。您还可以包含一个可选 modelIdentifier 字符串，用于标识您使用的 LLM。对于 retrievalResults 和 retrievedReferences，您可以提供可选的名称和元数据。

下面是一个包含 6 个输入并使用了 JSON 行格式的自定义数据集示例。

{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}
{"conversationTurns":[{"prompt":{"content":[{"text":"Provide the prompt you used to generate the response"}]},"referenceResponses":[{"content":[{"text":"A ground truth for the final response generated by the LLM"}]}],"referenceContexts":[{"content":[{"text":"A ground truth for a received passage"}]}],"output":{"text":"The output of the LLM","modelIdentifier":"(Optional) a string identifying your model","knowledgeBaseIdentifier":"A string identifying your RAG source","retrievedPassages":{"retrievalResults":[{"name":"(Optional) a name for your retrieval","content":{"text":"The retrieved content"},"metadata":{"(Optional) a key for your metadata":"(Optional) a value for your metadata"}}]}}}]}

为清楚起见，下面显示了展开的提示数据集格式。在实际提示数据集内，每一行（提示）都必须是一个有效的 JSON 对象。

{
    "conversationTurns": [
        {
            "prompt": {
                "content": [
                    {
                        "text": "Provide the prompt you used to generate the responses"
                    }
                ]
            },
            "referenceResponses": [
                {
                    "content": [
                        {
                            "text": "A ground truth for the final response generated by the LLM"
                        }
                    ]
                }
            ],
            "referenceContexts": [
                {
                    "content": [
                        {
                            "text": "A ground truth for a received passage"
                        }
                    ]
                }
            ],
            "output": {
                "text": "The output of the LLM",
                "modelIdentifier": "(Optional) a string identifying your model",
                "knowledgeBaseIdentifier": "A string identifying your RAG source",
                "retrievedPassages": {
                    "retrievalResults": [
                        {
                            "name": "(Optional) a name for your retrieval",
                            "content": {
                                "text": "The retrieved content"
                            },
                            "metadata": {
                                "(Optional) a key for your metadata": "(Optional) a value for your metadata"
                            }
                        }
                    ]
                }
            }
        }
    ]
}