go-llm/llm.go

package go_llm

import (
	"context"
	"fmt"

	"github.com/sashabaranov/go-openai"
)

type Role string

const (
	RoleSystem    Role = "system"
	RoleUser      Role = "user"
	RoleAssistant Role = "assistant"
)

type Image struct {
	Base64      string
	ContentType string
	Url         string
}

func (i Image) toRaw() map[string]any {
	res := map[string]any{
		"base64":      i.Base64,
		"contenttype": i.ContentType,
		"url":         i.Url,
	}

	return res
}

func (i *Image) fromRaw(raw map[string]any) Image {
	var res Image

	res.Base64 = raw["base64"].(string)
	res.ContentType = raw["contenttype"].(string)
	res.Url = raw["url"].(string)

	return res
}

type Message struct {
	Role   Role
	Name   string
	Text   string
	Images []Image
}

func (m Message) toRaw() map[string]any {
	res := map[string]any{
		"role": m.Role,
		"name": m.Name,
		"text": m.Text,
	}

	images := make([]map[string]any, 0, len(m.Images))
	for _, img := range m.Images {
		images = append(images, img.toRaw())
	}

	res["images"] = images

	return res
}

func (m *Message) fromRaw(raw map[string]any) Message {
	var res Message

	res.Role = Role(raw["role"].(string))
	res.Name = raw["name"].(string)
	res.Text = raw["text"].(string)

	images := raw["images"].([]map[string]any)
	for _, img := range images {
		var i Image

		res.Images = append(res.Images, i.fromRaw(img))
	}

	return res
}

func (m Message) toChatCompletionMessages() []openai.ChatCompletionMessage {
	var res openai.ChatCompletionMessage

	res.Role = string(m.Role)
	res.Name = m.Name
	res.Content = m.Text

	for _, img := range m.Images {
		if img.Base64 != "" {
			res.MultiContent = append(res.MultiContent, openai.ChatMessagePart{
				Type: "image_url",
				ImageURL: &openai.ChatMessageImageURL{
					URL: "data:" + img.ContentType + ";base64," + img.Base64,
				},
			})
		} else if img.Url != "" {
			res.MultiContent = append(res.MultiContent, openai.ChatMessagePart{
				Type: "image_url",
				ImageURL: &openai.ChatMessageImageURL{
					URL: img.Url,
				},
			})
		}
	}

	// openai does not support messages with both content and multi-content
	if len(res.MultiContent) > 0 && res.Content != "" {
		res.MultiContent = append([]openai.ChatMessagePart{{
			Type: "text",
			Text: res.Content,
		}}, res.MultiContent...)
		res.Content = ""
	}

	return []openai.ChatCompletionMessage{res}
}

type ToolCall struct {
	ID           string
	FunctionCall FunctionCall
}

func (t ToolCall) toRaw() map[string]any {
	res := map[string]any{
		"id": t.ID,
	}

	res["function"] = t.FunctionCall.toRaw()

	return res
}

func (t ToolCall) toChatCompletionMessages() []openai.ChatCompletionMessage {
	return []openai.ChatCompletionMessage{{
		Role:       openai.ChatMessageRoleTool,
		ToolCallID: t.ID,
	}}
}

type ToolCallResponse struct {
	ID     string
	Result any
	Error  error
}

func (t ToolCallResponse) toRaw() map[string]any {
	res := map[string]any{
		"id":     t.ID,
		"result": t.Result,
	}

	if t.Error != nil {
		res["error"] = t.Error.Error()
	}

	return res
}

func (t ToolCallResponse) toChatCompletionMessages() []openai.ChatCompletionMessage {
	var refusal string
	if t.Error != nil {
		refusal = t.Error.Error()
	}

	if refusal != "" {
		if t.Result != "" {
			t.Result = fmt.Sprint(t.Result) + " (error in execution: " + refusal + ")"
		} else {
			t.Result = "error in execution:" + refusal
		}
	}

	return []openai.ChatCompletionMessage{{
		Role:       openai.ChatMessageRoleTool,
		Content:    fmt.Sprint(t.Result),
		ToolCallID: t.ID,
	}}
}

type ChatCompletion interface {
	ChatComplete(ctx context.Context, req Request) (Response, error)
}

type LLM interface {
	ModelVersion(modelVersion string) (ChatCompletion, error)
}

func OpenAI(key string) LLM {
	return openaiImpl{key: key}
}

func Anthropic(key string) LLM {
	return anthropic{key: key}
}

func Google(key string) LLM {
	return google{key: key}
}