1.背景

当前OpenAI提供了开放接口，支持通过api的方式调用LLM进行文本推理、图片生成等能力，但目前官方只提供了Python SDK。为了后续更方便集成和应用，可以采用Golang对核心推理调用接口进行封装，提供模型调用能力。

2.相关准备

官方OpenAPI文档：https://platform.openai.com/docs/overview

1. 首先需要注册OpenAI账号，并且创建OpenAPI Key：https://platform.openai.com/api-keys，账号内需要充值5美元用于API调用计费。充值需要有美联储值卡，可以选择找代充，也可以直接买已有的账号，链接：https://eylink.cn/
2. OpenAPI官方调用域名为：https://api.openai.com，国内需要开启全局科学上网才可调用，可以用代理域名：https://api.openai-proxy.com
3. OpenAPI调用模型计费规则以消耗的Tokens计费：https://openai.com/api/pricing/
   1. gpt-3.5-turbo：2美元/百万Tokens
   2. gpt-4-turbo：40美元/百万Tokens
   3. gpt-4o：20美元/百万Tokens

3.实现代码

代码地址已上传：https://github.com/pbrong/llm_hub/blob/master/pkg/llm_caller/gpt_caller.go

• helper.go

package llm_callerimport "context"var (_ LLMCaller = &gptLLMCaller{}
)type LLMCaller interface {Call(ctx context.Context, userPrompt string) (completions string, err error)
}type Message struct {Role    string `json:"role"`Content string `json:"content"`
}type GptCompletion struct {Created int `json:"created"`Usage   struct {CompletionTokens int `json:"completion_tokens"`PromptTokens     int `json:"prompt_tokens"`TotalTokens      int `json:"total_tokens"`} `json:"usage"`Model   string `json:"model"`ID      string `json:"id"`Choices []struct {FinishReason string `json:"finish_reason"`Index        int    `json:"index"`Message      struct {Role    string `json:"role"`Content string `json:"content"`} `json:"message"`} `json:"choices"`SystemFingerprint interface{} `json:"system_fingerprint"`Object            string      `json:"object"`
}

• gpt_caller.go

package llm_callerimport ("context""encoding/json""fmt""llm_hub/conf""llm_hub/pkg/http"
)const (// 请求路径CompletionsURL = "/v1/chat/completions"// 可用模型Gpt35TurboModel = "gpt-3.5-turbo"
)type gptLLMCaller struct {openAiKey   stringsystemText  stringtemperature float64maxTokens   int64
}func NewGptLLMCaller(ctx context.Context, systemText string, temperature float64, maxTokens int64) (*gptLLMCaller, error) {return &gptLLMCaller{openAiKey:   conf.LLMHubConfig.Openai.Key,systemText:  systemText,temperature: temperature,maxTokens:   maxTokens,}, nil
}func (caller *gptLLMCaller) Call(ctx context.Context, userPrompt string) (completion string, err error) {reqURL := conf.LLMHubConfig.Openai.Host + CompletionsURLbody := map[string]interface{}{"model":       Gpt35TurboModel,"temperature": caller.temperature,"stream":      false,"max_tokens":  caller.maxTokens,"messages":    nil,}body["messages"] = buildPromptMessages(caller.systemText, userPrompt)headers := buildAuthHeaders(caller.openAiKey)resp, err := http.PostWithHeader(reqURL, body, headers)if err != nil {return "", fmt.Errorf("GPT调用失败, err = %v", err)}var gptCompletion GptCompletion_ = json.Unmarshal(resp, &gptCompletion)if len(gptCompletion.Choices) > 0 {completion = gptCompletion.Choices[0].Message.Content}return completion, nil
}func buildAuthHeaders(key string) map[string]string {headers := map[string]string{"Authorization": "Bearer " + key,}return headers
}func buildPromptMessages(system string, user string) []*Message {var messages []*Messagemessages = append(messages, &Message{Role:    "system",Content: system,})messages = append(messages, &Message{Role:    "user",Content: user,})return messages
}

• 测试：gpt_caller_test.go

package llm_callerimport ("context""github.com/stretchr/testify/assert""llm_hub/conf""testing"
)func Test_gptLLMCaller_Call(t *testing.T) {conf.Init()ctx := context.Background()caller, err := NewGptLLMCaller(ctx, "hello gpt-3.5-turbo", 0.5, 128)assert.Nil(t, err)completion, err := caller.Call(ctx, "hello world")assert.Nil(t, err)t.Logf("gpt call success, completion = %v", completion)
}

成功实现gpt-3.5.turbo模型调用：

2024/05/26 22:17:11 post with header, url = https://api.openai-proxy.com/v1/chat/completions, request = {"max_tokens": 128,"messages": [{"role": "system","content": "hello gpt-3.5-turbo"},{"role": "user","content": "hello world"}],"model": "gpt-3.5-turbo","stream": false,"temperature": 0.5
}, response = {"id": "chatcmpl-9T8yMDeJDEHKyN70Skk3prHvHZBkz","object": "chat.completion","created": 1716733030,"model": "gpt-3.5-turbo-0125","choices": [{"index": 0,"message": {"role": "assistant","content": "Hello! How can I assist you today?"},"logprobs": null,"finish_reason": "stop"}],"usage": {"prompt_tokens": 23,"completion_tokens": 9,"total_tokens": 32},"system_fingerprint": null
}gpt_caller_test.go:17: gpt call success, completion = Hello! How can I assist you today?
--- PASS: Test_gptLLMCaller_Call (1.31s)
PASS