Refactor entire system to be more contextual so that conversation flow can be more easily managed

llm.go

@@ -2,6 +2,7 @@ package go_llm
 
 import (
 	"context"
+	"github.com/sashabaranov/go-openai"
 )
 
 type Role string
@@ -18,6 +19,26 @@ type Image struct {
 	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
@@ -25,10 +46,66 @@ type Message struct {
 	Images []Image
 }
 
-type Request struct {
-	Messages    []Message
-	Toolbox     *ToolBox
-	Temperature *float32
+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,
+				},
+			})
+		}
+	}
+
+	return []openai.ChatCompletionMessage{res}
 }
 
 type ToolCall struct {
@@ -36,16 +113,54 @@ type ToolCall struct {
 	FunctionCall FunctionCall
 }
 
-type ResponseChoice struct {
-	Index   int
-	Role    Role
-	Content string
-	Refusal string
-	Name    string
-	Calls   []ToolCall
+func (t ToolCall) toRaw() map[string]any {
+	res := map[string]any{
+		"id": t.ID,
+	}
+
+	res["function"] = t.FunctionCall.toRaw()
+
+	return res
 }
-type Response struct {
-	Choices []ResponseChoice
+
+func (t ToolCall) toChatCompletionMessages() []openai.ChatCompletionMessage {
+	return []openai.ChatCompletionMessage{{
+		Role:       openai.ChatMessageRoleTool,
+		ToolCallID: t.ID,
+	}}
+}
+
+type ToolCallResponse struct {
+	ID     string
+	Result string
+	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()
+	}
+
+	return []openai.ChatCompletionMessage{{
+		Role:       openai.ChatMessageRoleTool,
+		Content:    t.Result,
+		Refusal:    refusal,
+		ToolCallID: t.ID,
+	}}
 }
 
 type ChatCompletion interface {