Gemini 多模態大型語言模型大小事 Day11 - 使用 Gemini API 產生結構化輸出

16th鐵人賽

kevin_chiu

2024-09-21 01:11:07

43 瀏覽

分享至

前言

    程式環境都會用colab 來執行程式，如果要在其他環境執行，請自行修改哦

colab 事前準備：設定專案和 API 金鑰
載入gemini

#pip install -q -U google-generativeai
import google.generativeai as genai

API 金鑰

from google.colab import userdata
API_KEY=userdata.get('GOOGLE_API_KEY')

#genai.configure(api_key="YOUR_API_KEY")

# Configure the client library by providing your API key.
genai.configure(api_key=API_KEY)

使用 Gemini API 產生結構化輸出

Gemini 預設產生非結構化文本，但某些應用程式需要結構化文字。對於這些用例，您可以限制 Gemini 使用 JSON 回應，JSON 是一種適合自動處理的結構化資料格式。您也可以限制模型使用列舉中指定的選項之一進行回應。

以下是一些可能需要模型結構化輸出的用例：

透過從報紙文章中提取公司資訊來建立公司資料庫。
從簡歷中提取標準化資訊。
從食譜中提取成分，並顯示每種成分的雜貨網站連結。
在提示中，您可以要求 Gemini 產生 JSON 格式的輸出，但請注意，該模型不保證會產生 JSON，並且只產生 JSON。為了獲得更具確定性的回應，您可以在 responseSchema 欄位中傳遞特定的 JSON 模式，以便 Gemini 始終以預期的結構進行回應。

如何使用 generateContent您選擇的 SDK 或直接使用 REST API 的方法來產生 JSON。顯示了純文字輸入，儘管 Gemini 還可以對包含圖像、視訊和音訊的多模式請求產生 JSON 回應。

在提示中提供模式作為文本

model = genai.GenerativeModel("gemini-1.5-pro-latest")
prompt = """以 JSON 格式列出一些好吃的中餐食譜。

使用此 JSON 架構：

Recipe = {'recipe_name': str, 'ingredients': list[str]}
Return: list[Recipe]"""
result = model.generate_content(prompt)
print(result)

回答

response:
GenerateContentResponse(
    done=True,
    iterator=None,
    result=protos.GenerateContentResponse({
      "candidates": [
        {
          "content": {
            "parts": [
              {
                "text": "```json\n[\n  {\n    \"recipe_name\": \"\u5bae\u4fdd\u96de\u4e01\",\n    \"ingredients\": [\n      \"\u96de\u8089\u4e01\",\n      \"\u82b1\u751f\",\n      \"\u4e7e\u8fa3\u6912\",\n      \"\u82b1\u6912\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9ebb\u5a46\u8c46\u8150\",\n    \"ingredients\": [\n      \"\u5ae9\u8c46\u8150\",\n      \"\u8c6c\u7d5e\u8089\",\n      \"\u8c46\u74e3\u91ac\",\n      \"\u8fa3\u6912\u91ac\",\n      \"\u82b1\u6912\u7c89\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u7cd6\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u7cd6\u918b\u6392\u9aa8\",\n    \"ingredients\": [\n      \"\u8c6c\u5c0f\u6392\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9b5a\u9999\u8304\u5b50\",\n    \"ingredients\": [\n      \"\u8304\u5b50\",\n      \"\u8c6c\u7d5e\u8089\",\n      \"\u8c46\u74e3\u91ac\",\n      \"\u8fa3\u6912\u91ac\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9752\u6912\u8089\u7d72\",\n    \"ingredients\": [\n      \"\u8c6c\u8089\u7d72\",\n      \"\u9752\u6912\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u91ac\u6cb9\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  }\n]\n```"
              }
            ],
            "role": "model"
          },
          .............
)

透過模型配置提供模式

import typing_extensions as typing

class Recipe(typing.TypedDict):
    recipe_name: str
    ingredients: list[str]

model = genai.GenerativeModel("gemini-1.5-pro-latest")
result = model.generate_content(
    "列出一些好吃的中餐食譜。",
    generation_config=genai.GenerationConfig(
        response_mime_type="application/json", response_schema=list[Recipe]
    ),
)
print(result)

回答

......
response:
GenerateContentResponse(
.........

"parts": [
              {
                "text": "[{\"ingredients\": [\"\u732a\u8089\", \"\u767d\u83dc\", \"\u8c46\u8150\", \"\u7c89\u6761\", \"\u6728\u8033\", \"\u8471\", \"\u59dc\", \"\u849c\", \"\u9171\u6cb9\", \"\u6599\u9152\", \"\u76d0\", \"\u7cd6\"], \"recipe_name\": \"\u767d\u83dc\u7096\u7c89\u6761\"}, {\"ingredients\": [\"\u9e21\u8089\", \"\u9752\u6912\", \"\u7ea2\u6912\", \"\u8471\", \"\u59dc\", \"\u849c\", \"\u8fa3\u6912\", \"\u82b1\u6912\", \"\u9171\u6cb9\", \"\u6599\u9152\", \"\u76d0\", \"\u7cd6\"], \"recipe_name\": \"\u5bab\u4fdd\u9e21\u4e01\"}, {\"ingredients\": [\"\u571f\u8c46\", \"\u9752\u6912\", \"\u7ea2\u6912\", \"\u8471\", \"\u59dc\", \"\u849c\", \"\u9171\u6cb9\", \"\u918b\", \"\u76d0\", \"\u7cd6\"], \"recipe_name\": \"\u571f\u8c46\u4e1d\"}, {\"ingredients\": [\"\u756a\u8304\", \"\u9e21\u86cb\", \"\u8471\", \"\u59dc\", \"\u849c\", \"\u76d0\", \"\u7cd6\"], \"recipe_name\": \"\u756a\u8304\u7092\u86cb\"}, {\"ingredients\": [\"\u8304\u5b50\", \"\u8089\u672b\", \"\u8471\", \"\u59dc\", \"\u849c\", \"\u8c46\u74e3\u9171\", \"\u9171\u6cb9\", \"\u6599\u9152\", \"\u76d0\", \"\u7cd6\"], \"recipe_name\": \"\u9c7c\u9999\u8304\u5b50\"}] "
              }
            ],
            "role": "model"
.......

定義模式最簡單的方法是使用類型提示註解。

generation_config={"response_mime_type": "application/json",
                   "response_schema": list[Recipe]}
print(result)

回答

response:
GenerateContentResponse(
    done=True,
    iterator=None,
    result=protos.GenerateContentResponse({
      "candidates": [
        {
          "content": {
            "parts": [
              {
                "text": "```json\n[\n  {\n    \"recipe_name\": \"\u5bae\u4fdd\u96de\u4e01\",\n    \"ingredients\": [\n      \"\u96de\u8089\u4e01\",\n      \"\u82b1\u751f\",\n      \"\u4e7e\u8fa3\u6912\",\n      \"\u82b1\u6912\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9ebb\u5a46\u8c46\u8150\",\n    \"ingredients\": [\n      \"\u5ae9\u8c46\u8150\",\n      \"\u8c6c\u7d5e\u8089\",\n      \"\u8c46\u74e3\u91ac\",\n      \"\u8fa3\u6912\u91ac\",\n      \"\u82b1\u6912\u7c89\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u7cd6\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u7cd6\u918b\u6392\u9aa8\",\n    \"ingredients\": [\n      \"\u8c6c\u5c0f\u6392\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9b5a\u9999\u8304\u5b50\",\n    \"ingredients\": [\n      \"\u8304\u5b50\",\n      \"\u8c6c\u7d5e\u8089\",\n      \"\u8c46\u74e3\u91ac\",\n      \"\u8fa3\u6912\u91ac\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u8525\",\n      \"\u91ac\u6cb9\",\n      \"\u918b\",\n      \"\u7cd6\"\n    ]\n  },\n  {\n    \"recipe_name\": \"\u9752\u6912\u8089\u7d72\",\n    \"ingredients\": [\n      \"\u8c6c\u8089\u7d72\",\n      \"\u9752\u6912\",\n      \"\u8591\",\n      \"\u849c\",\n      \"\u91ac\u6cb9\",\n      \"\u592a\u767d\u7c89\"\n    ]\n  }\n]\n```"
              }
            ],
            "role": "model"
          },
          
          ..............

使用列舉來限制輸出

import google.generativeai as genai
import enum
from typing import TypedDict

from typing_extensions import TypedDict # Use typing_extensions.TypedDict


# Define a TypedDict for the schema
class ChoiceSchema(TypedDict):
    value: str 

# Define your enum class
class Choice(enum.Enum):
    PERCUSSION = "打擊樂"
    STRING = "字串"
    WOODWIND = "木管樂器"
    BRASS = "黃銅"
    KEYBOARD = "鍵盤"

model = genai.GenerativeModel("gemini-1.5-pro-latest")

organ = genai.upload_file(path="organ.jpg",
                            display_name="organ")
result = model.generate_content(
    ["這是什麼樂器：",organ],
    generation_config=genai.GenerationConfig(
        response_mime_type="text/x.enum", response_schema=ChoiceSchema
    ),
)


print(result)  # Keyboard

回答

response:
GenerateContentResponse(
    done=True,
    iterator=None,
    result=protos.GenerateContentResponse({
      "candidates": [
        {
          "content": {
            "parts": [
              {
                "text": "\u9019\u4ef6\u6a02\u5668\u662f\u4e00\u53f0**\u7ba1\u98a8\u7434**\u3002"
              }
            ],
            "role": "model"
          },
................

model = genai.GenerativeModel("gemini-1.5-pro-latest")

organ = genai.upload_file("organ.jpg")
result = model.generate_content(
    ["請用繁體中文回答這是什麼樂器：", organ],
    generation_config=genai.GenerationConfig(
        response_mime_type="text/x.enum",
        response_schema={
            "type": "STRING",
            "enum": ["打擊樂", "字串", "木管樂器", "黃銅", "鍵盤"],
        },
    ),
)
print(result)  # Keyboard

回答

response:
GenerateContentResponse(
    done=True,
    iterator=None,
    result=protos.GenerateContentResponse({
      "candidates": [
        {
          "content": {
            "parts": [
              {
                "text": " \u9375\u76e4"
              }
            ],
            "role": "model"
          },
          
          ............

向模型詢問食譜標題列表，並使用列舉Grade為每個標題提供評份等級

import enum
from typing import List, TypedDict
from typing_extensions import TypedDict
from pydantic import BaseModel, Field

class Grade(enum.Enum):
    A_PLUS = "a+"
    A = "a"
    B = "b"
    C = "c"
    D = "d"
    F = "f"

class Recipe(TypedDict):
    grade: Grade
    recipe_name: str
    
model = genai.GenerativeModel("gemini-1.5-pro-latest")

result = model.generate_content(
    "列出大約 10 種中餐食譜，根據受歡迎程度對它們進行評分",
    generation_config=genai.GenerationConfig(
        response_mime_type="application/json",
        response_schema={
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "grade": {"type": "string", "enum": [e.value for e in Grade]},
                    "recipe_name": {"type": "string"},
                },
            },
        }
    ),
    request_options={"timeout": 600},
)
print(result)  # [{"grade": "a+", "recipe_name": "Mapo Tofu"}, ...]

回答

response:
GenerateContentResponse(
    done=True,
    iterator=None,
    result=protos.GenerateContentResponse({
      "candidates": [
        {
          "content": {
            "parts": [
              {
                "text": "[{\"grade\": \"a+\", \"recipe_name\": \"Kung Pao Chicken\"}, {\"grade\": \"a+\", \"recipe_name\": \"Mapo Tofu\"}, {\"grade\": \"a\", \"recipe_name\": \"Chow Mein\"}, {\"grade\": \"a\", \"recipe_name\": \"Sweet and Sour Pork\"}, {\"grade\": \"a\", \"recipe_name\": \"Dumplings\"}, {\"grade\": \"a\", \"recipe_name\": \"Spring Rolls\"}, {\"grade\": \"b\", \"recipe_name\": \"Peking Duck\"}, {\"grade\": \"b\", \"recipe_name\": \"Wontons\"}, {\"grade\": \"b\", \"recipe_name\": \"Fried Rice\"}, {\"grade\": \"b\", \"recipe_name\": \"Lo Mein\"}] "
              }
            ],
            "role": "model"
          },
          ...........