🤖

Gemini REST APIでのJSON Modeの利用(1.5 Flash編)

2024/06/24に公開

JSON

rest

Gemini

tech

generationConfig で response_mime_type に application/json を指定する
Gemini 1.5 Proでは generationConfig でさらに responseSchema を指定して厳密にスキーマを定義することが可能でしたが、Gemini 1.5 Flashではプロンプト内でスキーマを指定する必要があります

プロンプトの例

Translate to Japanese.
A string in `<paragraph>` tag to `</paragraph>` tag is one paragraph.
If a paragraph of input is translated and a paragraph consists of multiple sentences, output an array consisting of multiple String.
Using this JSON schema:
Paragraph = {\"line\": number, \"text\": list[string]}
Return a `list[Paragraph]`

request

curl "https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent?key=$API_KEY" \
    -H 'Content-Type: application/json' \
    -X POST \
    -d '{
      "systemInstruction": {
        "parts": [
          {
            "text": "Translate to Japanese.\n\nA string in `<paragraph>` tag to `</paragraph>` tag is one paragraph.\n\nIf a paragraph of input is translated and a paragraph consists of multiple sentences, output an array consisting of multiple String.\n\nUsing this JSON schema:\n\n Paragraph = {\"line\": number, \"text\": list[string]}\n\nReturn a `list[Paragraph]`\n"
          }
        ]
      },
      "contents": [
        {
          "parts": [
            {
              "text": "<paragraph>Cat</paragraph>"
            },
            {
              "text": "<paragraph>Dog</paragraph>"
            }
          ]
        }
      ],
      "generationConfig": {
            "response_mime_type": "application/json",
      }
    }'

レスポンスの candidates -> content -> parts -> text に JSONで返ってくるので更にパースして上げる必要があります

response

{
  "candidates": [
    {
      "content": {
        "parts": [
          {
            "text": "[{\"line\": 1, \"text\": [\"猫\"]}, {\"line\": 2, \"text\": [\"犬\"]}]\n"
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "index": 0,
      "safetyRatings": [
        {
          "category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_HATE_SPEECH",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_HARASSMENT",
          "probability": "NEGLIGIBLE"
        },
        {
          "category": "HARM_CATEGORY_DANGEROUS_CONTENT",
          "probability": "NEGLIGIBLE"
        }
      ]
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 128,
    "candidatesTokenCount": 60,
    "totalTokenCount": 188
  }
}

Discussion