AI Gateway multiple providers

Mix.install([
  {:kino, "~> 0.14"},
  {:req, "~> 0.4"}
])

認証情報の設定

以下の値を Secrets に登録する

• OPENAI_API_KEY: OpenAI の API キー
• GEMINI_API_KEY: Google Gemini の API キー
• CF_ACCOUNT_ID: Cloudfaler のアカウントID
• WORKERS_AI_API_KEY: Workers AI の API キー

OpenAI API の直接呼出

openai_api_key = System.get_env("LB_OPENAI_API_KEY")

openai_base_url = "https://api.openai.com/v1/chat/completions"

openai_model_id = "gpt-4o-mini"

openai_headers = %{
  "Content-Type" => "application/json",
  "Authorization" => "Bearer #{openai_api_key}"
}

request_body = %{
  model: openai_model_id,
  messages: [
    %{
      role: "system",
      content: "You are a helpful assistant."
    },
    %{
      role: "user",
      content: "あんパンとカレーパンはどちらが人気ですか"
    }
  ]
}

response =
  "#{openai_base_url}"
  |> Req.post!(json: request_body, headers: openai_headers)
  |> Map.get(:body)

response["choices"]
|> Enum.at(0)
|> Map.get("message")
|> Map.get("content")
|> Kino.Markdown.new()

Google Gemini API の直接呼出

gemini_api_key = System.get_env("LB_GEMINI_API_KEY")

gemini_base_url = "https://generativelanguage.googleapis.com/v1beta/models"

gemini_model_id = "gemini-1.5-flash-latest"

gemini_headers = %{
  "Content-Type" => "application/json"
}

request_body = %{
  contents: [
    %{
      parts: [
        %{
          text: "日本で最も人気のあるパンは何ですか"
        }
      ]
    }
  ]
}

response =
  "#{gemini_base_url}/#{gemini_model_id}:generateContent?key=#{gemini_api_key}"
  |> Req.post!(json: request_body, headers: gemini_headers)
  |> Map.get(:body)

response["candidates"]
|> Enum.at(0)
|> Map.get("content")
|> Map.get("parts")
|> Enum.at(0)
|> Map.get("text")
|> Kino.Markdown.new()

Workers AI の直接呼出

cf_account_id = System.get_env("LB_CF_ACCOUNT_ID")

workers_ai_api_key = System.get_env("LB_WORKERS_AI_API_KEY")

workers_ai_base_url = "https://api.cloudflare.com/client/v4/"

workers_ai_model_id = "llama-3.1-8b-instruct"

workers_ai_headers = %{
  "Content-Type" => "application/json",
  "Authorization" => "Bearer #{workers_ai_api_key}"
}

request_body = %{
  messages: [
    %{
      role: "system",
      content: "あなたは優秀なAIアシスタントです。日本語で簡潔に質問に回答してください"
    },
    %{
      role: "user",
      content: "食パンの一番美味しいところはどこですか"
    }
  ]
}

response =
  "#{workers_ai_base_url}/accounts/#{cf_account_id}/ai/run/@cf/meta/#{workers_ai_model_id}"
  |> Req.post!(json: request_body, headers: workers_ai_headers)
  |> Map.get(:body)

response["result"]["response"]
|> Kino.Markdown.new()

AI Gateway からの呼出

cf_gateway_name_input = Kino.Input.text("CLOUDFLARE GATEWAY_NAME")

cf_gateway_name = Kino.Input.read(cf_gateway_name_input)
cf_host = "gateway.ai.cloudflare.com"

gw_url =
  "https://#{cf_host}/v1/#{cf_account_id}/#{cf_gateway_name}"

gw_headers = %{
  "Content-Type" => "application/json"
}

question = "コロネはどこから食べるべきですか"

openai_request_body = %{
  provider: "openai",
  endpoint: "chat/completions",
  headers: openai_headers,
  query: %{
    model: openai_model_id,
    messages: [
      %{role: "system", content: "You are a helpful assistant."},
      %{role: "user", content: question}
    ]
  }
}

gemini_request_body = %{
  provider: "google-ai-studio",
  endpoint: "v1beta/models/#{gemini_model_id}:generateContent?key=#{gemini_api_key}",
  headers: gemini_headers,
  query: %{
    contents: [%{parts: [%{text: question}]}]
  }
}

workers_request_body = %{
  provider: "workers-ai",
  endpoint: "@cf/meta/#{workers_ai_model_id}",
  headers: workers_ai_headers,
  query: %{
    messages: [
      %{role: "system", content: "あなたは優秀なAIアシスタントです。日本語で簡潔に質問に回答してください"},
      %{role: "user", content: question}
    ]
  }
}

request_body = [
  openai_request_body,
  gemini_request_body,
  workers_request_body
]

response =
  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)

request_body = [
  gemini_request_body,
  workers_request_body,
  openai_request_body
]

response =
  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)

request_body = [
  workers_request_body,
  openai_request_body,
  gemini_request_body
]

response =
  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)

認証エラーを発生させる

# 認証キーをリクエストから外す
request_body = [
  %{
    provider: "google-ai-studio",
    endpoint: "v1beta/models/#{gemini_model_id}:generateContent",
    headers: gemini_headers,
    query: %{
      contents: [%{parts: [%{text: question}]}]
    }
  },
  workers_request_body,
  openai_request_body
]

response =
  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)

キャッシュ

AI Gateway でキャッシュを有効にしてから2回実行する

request_body = [
  openai_request_body,
  gemini_request_body,
  workers_request_body
]

response =
  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)

レート制限

AI Gateway で1分間のレート制限を3回にしてから5回実行する

1..5
|> Enum.map(fn index ->
  request_body = [
    %{
      provider: "openai",
      endpoint: "chat/completions",
      headers: openai_headers,
      query: %{
        model: openai_model_id,
        messages: [
          %{role: "system", content: "You are a helpful assistant."},
          %{role: "user", content: "#{index}といえば、何を表す数字ですか"}
        ]
      }
    }
  ]

  gw_url
  |> Req.post!(json: request_body, headers: gw_headers)
  |> Map.get(:body)
end)