Add support for multi part chat messages (and gpt-4-vision-preview model) (#580)

* Add support for multi part chat messages

OpenAI has recently introduced a new model called gpt-4-visual-preview,
which now supports images as input. The chat completion endpoint accepts
multi-part chat messages, where the content can be an array of structs
in addition to the usual string format.

This commit introduces new structures and constants to represent
different types of content parts. It also implements the json.Marshaler
and json.Unmarshaler interfaces on ChatCompletionMessage.

* Add ImageURLDetail and ChatMessagePartType types

* Optimize ChatCompletionMessage deserialization

* Add ErrContentFieldsMisused error

Author: rkintzi

chat.go

@@ -2,6 +2,7 @@ package openai
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"net/http"
 )
@@ -20,6 +21,7 @@ const chatCompletionsSuffix = "/chat/completions"
 var (
 	ErrChatCompletionInvalidModel       = errors.New("this model is not supported with this method, please use CreateCompletion client method instead") //nolint:lll
 	ErrChatCompletionStreamNotSupported = errors.New("streaming is not supported with this method, please use CreateChatCompletionStream")              //nolint:lll
+	ErrContentFieldsMisused             = errors.New("can't use both Content and MultiContent properties simultaneously")
 )
 
 type Hate struct {
@@ -51,9 +53,36 @@ type PromptAnnotation struct {
 	ContentFilterResults ContentFilterResults `json:"content_filter_results,omitempty"`
 }
 
+type ImageURLDetail string
+
+const (
+	ImageURLDetailHigh ImageURLDetail = "high"
+	ImageURLDetailLow  ImageURLDetail = "low"
+	ImageURLDetailAuto ImageURLDetail = "auto"
+)
+
+type ChatMessageImageURL struct {
+	URL    string         `json:"url,omitempty"`
+	Detail ImageURLDetail `json:"detail,omitempty"`
+}
+
+type ChatMessagePartType string
+
+const (
+	ChatMessagePartTypeText     ChatMessagePartType = "text"
+	ChatMessagePartTypeImageURL ChatMessagePartType = "image_url"
+)
+
+type ChatMessagePart struct {
+	Type     ChatMessagePartType  `json:"type,omitempty"`
+	Text     string               `json:"text,omitempty"`
+	ImageURL *ChatMessageImageURL `json:"image_url,omitempty"`
+}
+
 type ChatCompletionMessage struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
+	Role         string `json:"role"`
+	Content      string `json:"content"`
+	MultiContent []ChatMessagePart
 
 	// This property isn't in the official documentation, but it's in
 	// the documentation for the official library for python:
@@ -70,6 +99,64 @@ type ChatCompletionMessage struct {
 	ToolCallID string `json:"tool_call_id,omitempty"`
 }
 
+func (m ChatCompletionMessage) MarshalJSON() ([]byte, error) {
+	if m.Content != "" && m.MultiContent != nil {
+		return nil, ErrContentFieldsMisused
+	}
+	if len(m.MultiContent) > 0 {
+		msg := struct {
+			Role         string            `json:"role"`
+			Content      string            `json:"-"`
+			MultiContent []ChatMessagePart `json:"content,omitempty"`
+			Name         string            `json:"name,omitempty"`
+			FunctionCall *FunctionCall     `json:"function_call,omitempty"`
+			ToolCalls    []ToolCall        `json:"tool_calls,omitempty"`
+			ToolCallID   string            `json:"tool_call_id,omitempty"`
+		}(m)
+		return json.Marshal(msg)
+	}
+	msg := struct {
+		Role         string            `json:"role"`
+		Content      string            `json:"content"`
+		MultiContent []ChatMessagePart `json:"-"`
+		Name         string            `json:"name,omitempty"`
+		FunctionCall *FunctionCall     `json:"function_call,omitempty"`
+		ToolCalls    []ToolCall        `json:"tool_calls,omitempty"`
+		ToolCallID   string            `json:"tool_call_id,omitempty"`
+	}(m)
+	return json.Marshal(msg)
+}
+
+func (m *ChatCompletionMessage) UnmarshalJSON(bs []byte) error {
+	msg := struct {
+		Role         string `json:"role"`
+		Content      string `json:"content"`
+		MultiContent []ChatMessagePart
+		Name         string        `json:"name,omitempty"`
+		FunctionCall *FunctionCall `json:"function_call,omitempty"`
+		ToolCalls    []ToolCall    `json:"tool_calls,omitempty"`
+		ToolCallID   string        `json:"tool_call_id,omitempty"`
+	}{}
+	if err := json.Unmarshal(bs, &msg); err == nil {
+		*m = ChatCompletionMessage(msg)
+		return nil
+	}
+	multiMsg := struct {
+		Role         string `json:"role"`
+		Content      string
+		MultiContent []ChatMessagePart `json:"content"`
+		Name         string            `json:"name,omitempty"`
+		FunctionCall *FunctionCall     `json:"function_call,omitempty"`
+		ToolCalls    []ToolCall        `json:"tool_calls,omitempty"`
+		ToolCallID   string            `json:"tool_call_id,omitempty"`
+	}{}
+	if err := json.Unmarshal(bs, &multiMsg); err != nil {
+		return err
+	}
+	*m = ChatCompletionMessage(multiMsg)
+	return nil
+}
+
 type ToolCall struct {
 	// Index is not nil only in chat completion chunk object
 	Index    *int         `json:"index,omitempty"`

chat_test.go

@@ -3,6 +3,7 @@ package openai_test
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"net/http"
@@ -296,6 +297,108 @@ func TestAzureChatCompletions(t *testing.T) {
 	checks.NoError(t, err, "CreateAzureChatCompletion error")
 }
 
+func TestMultipartChatCompletions(t *testing.T) {
+	client, server, teardown := setupAzureTestServer()
+	defer teardown()
+	server.RegisterHandler("/openai/deployments/*", handleChatCompletionEndpoint)
+
+	_, err := client.CreateChatCompletion(context.Background(), openai.ChatCompletionRequest{
+		MaxTokens: 5,
+		Model:     openai.GPT3Dot5Turbo,
+		Messages: []openai.ChatCompletionMessage{
+			{
+				Role: openai.ChatMessageRoleUser,
+				MultiContent: []openai.ChatMessagePart{
+					{
+						Type: openai.ChatMessagePartTypeText,
+						Text: "Hello!",
+					},
+					{
+						Type: openai.ChatMessagePartTypeImageURL,
+						ImageURL: &openai.ChatMessageImageURL{
+							URL:    "URL",
+							Detail: openai.ImageURLDetailLow,
+						},
+					},
+				},
+			},
+		},
+	})
+	checks.NoError(t, err, "CreateAzureChatCompletion error")
+}
+
+func TestMultipartChatMessageSerialization(t *testing.T) {
+	jsonText := `[{"role":"system","content":"system-message"},` +
+		`{"role":"user","content":[{"type":"text","text":"nice-text"},` +
+		`{"type":"image_url","image_url":{"url":"URL","detail":"high"}}]}]`
+
+	var msgs []openai.ChatCompletionMessage
+	err := json.Unmarshal([]byte(jsonText), &msgs)
+	if err != nil {
+		t.Fatalf("Expected no error: %s", err)
+	}
+	if len(msgs) != 2 {
+		t.Errorf("unexpected number of messages")
+	}
+	if msgs[0].Role != "system" || msgs[0].Content != "system-message" || msgs[0].MultiContent != nil {
+		t.Errorf("invalid user message: %v", msgs[0])
+	}
+	if msgs[1].Role != "user" || msgs[1].Content != "" || len(msgs[1].MultiContent) != 2 {
+		t.Errorf("invalid user message")
+	}
+	parts := msgs[1].MultiContent
+	if parts[0].Type != "text" || parts[0].Text != "nice-text" {
+		t.Errorf("invalid text part: %v", parts[0])
+	}
+	if parts[1].Type != "image_url" || parts[1].ImageURL.URL != "URL" || parts[1].ImageURL.Detail != "high" {
+		t.Errorf("invalid image_url part")
+	}
+
+	s, err := json.Marshal(msgs)
+	if err != nil {
+		t.Fatalf("Expected no error: %s", err)
+	}
+	res := strings.ReplaceAll(string(s), " ", "")
+	if res != jsonText {
+		t.Fatalf("invalid message: %s", string(s))
+	}
+
+	invalidMsg := []openai.ChatCompletionMessage{
+		{
+			Role:    "user",
+			Content: "some-text",
+			MultiContent: []openai.ChatMessagePart{
+				{
+					Type: "text",
+					Text: "nice-text",
+				},
+			},
+		},
+	}
+	_, err = json.Marshal(invalidMsg)
+	if !errors.Is(err, openai.ErrContentFieldsMisused) {
+		t.Fatalf("Expected error: %s", err)
+	}
+
+	err = json.Unmarshal([]byte(`["not-a-message"]`), &msgs)
+	if err == nil {
+		t.Fatalf("Expected error")
+	}
+
+	emptyMultiContentMsg := openai.ChatCompletionMessage{
+		Role:         "user",
+		MultiContent: []openai.ChatMessagePart{},
+	}
+	s, err = json.Marshal(emptyMultiContentMsg)
+	if err != nil {
+		t.Fatalf("Unexpected error")
+	}
+	res = strings.ReplaceAll(string(s), " ", "")
+	if res != `{"role":"user","content":""}` {
+		t.Fatalf("invalid message: %s", string(s))
+	}
+}
+
 // handleChatCompletionEndpoint Handles the ChatGPT completion endpoint by the test server.
 func handleChatCompletionEndpoint(w http.ResponseWriter, r *http.Request) {
 	var err error