add image edit utility

This commit is contained in:
Jicheng Lu 2024-07-19 14:18:22 -05:00
parent de5750dc28
commit b1fb9d6367
12 changed files with 231 additions and 22 deletions

View file

@ -1,16 +0,0 @@
namespace BotSharp.Abstraction.Files.Models;
public class LlmFileContext
{
[JsonPropertyName("user_request")]
[JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
public string? UserRequest { get; set; }
[JsonPropertyName("file_types")]
[JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
public string? FileTypes { get; set; }
[JsonPropertyName("image_description")]
[JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
public string? ImageDescription { get; set; }
}

View file

@ -73,7 +73,7 @@ public class HandleEmailRequestFn : IFunctionCallback
private async Task<IEnumerable<MessageFileModel>> GetConversationFiles()
{
var convService = _services.GetService<IConversationService>();
var convService = _services.GetRequiredService<IConversationService>();
var fileService = _services.GetRequiredService<IBotSharpFileService>();
var conversationId = convService.ConversationId;
var dialogs = convService.GetDialogHistory(fromBreakpoint: false);

View file

@ -11,12 +11,15 @@
</PropertyGroup>
<ItemGroup>
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\functions\edit_image.json" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\functions\generate_image.json" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\functions\read_image.json" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\functions\read_pdf.json" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\edit_image.fn.liquid" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\generate_image.fn.liquid" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\read_image.fn.liquid" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\read_pdf.fn.liquid" />
<None Remove="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\select_edit_image_prompt.liquid" />
</ItemGroup>
<ItemGroup>
@ -38,6 +41,15 @@
<Content Include="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\read_pdf.fn.liquid">
<CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
</Content>
<Content Include="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\functions\edit_image.json">
<CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
</Content>
<Content Include="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\edit_image.fn.liquid">
<CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
</Content>
<Content Include="data\agents\6745151e-6d46-4a02-8de4-1c4f21c7da95\templates\select_edit_image_prompt.liquid">
<CopyToOutputDirectory>PreserveNewest</CopyToOutputDirectory>
</Content>
</ItemGroup>
<ItemGroup>

View file

@ -4,5 +4,6 @@ public class UtilityName
{
public const string ImageGenerator = "image-generator";
public const string ImageReader = "image-reader";
public const string ImageEditor = "image-editor";
public const string PdfReader = "pdf-reader";
}

View file

@ -0,0 +1,155 @@
using BotSharp.Abstraction.Templating;
using System.IO;
namespace BotSharp.Plugin.FileHandler.Functions;
public class EditImageFn : IFunctionCallback
{
public string Name => "edit_image";
public string Indication => "Editing image";
private readonly IServiceProvider _services;
private readonly ILogger<EditImageFn> _logger;
private string _conversationId;
private string _messageId;
public EditImageFn(
IServiceProvider services,
ILogger<EditImageFn> logger)
{
_services = services;
_logger = logger;
}
public async Task<bool> Execute(RoleDialogModel message)
{
var args = JsonSerializer.Deserialize<LlmContextIn>(message.FunctionArgs);
var descrpition = args?.UserRequest ?? string.Empty;
Init(message);
SetImageOptions();
var image = await SelectConversationImage();
var response = await GetImageEditGeneration(message, descrpition, image);
message.Content = response;
return true;
}
private void Init(RoleDialogModel message)
{
var convService = _services.GetRequiredService<IConversationService>();
_conversationId = convService.ConversationId;
_messageId = message.MessageId;
}
private void SetImageOptions()
{
var state = _services.GetRequiredService<IConversationStateService>();
state.SetState("image_format", "bytes");
state.SetState("image_count", "1");
}
private async Task<MessageFileModel?> SelectConversationImage()
{
var convService = _services.GetRequiredService<IConversationService>();
var fileService = _services.GetRequiredService<IBotSharpFileService>();
var dialogs = convService.GetDialogHistory(fromBreakpoint: false);
var messageIds = dialogs.Select(x => x.MessageId).Distinct().ToList();
var images = fileService.GetMessageFiles(_conversationId, messageIds, FileSourceType.User, imageOnly: true);
return await SelectImage(images, dialogs);
}
private async Task<MessageFileModel?> SelectImage(IEnumerable<MessageFileModel> images, List<RoleDialogModel> dialogs)
{
if (images.IsNullOrEmpty()) return null;
var llmProviderService = _services.GetRequiredService<ILlmProviderService>();
var render = _services.GetRequiredService<ITemplateRender>();
var db = _services.GetRequiredService<IBotSharpRepository>();
try
{
var promptImages = images.Where(x => x.ContentType == MediaTypeNames.Image.Png).Select((x, idx) =>
{
return $"id: {idx + 1}, image_name: {x.FileName}.{x.FileType}";
}).ToList();
if (promptImages.IsNullOrEmpty()) return null;
var prompt = db.GetAgentTemplate(BuiltInAgentId.UtilityAssistant, "select_edit_image_prompt");
prompt = render.Render(prompt, new Dictionary<string, object>
{
{ "image_list", promptImages }
});
var agent = new Agent
{
Id = BuiltInAgentId.UtilityAssistant,
Name = "Utility Assistant",
Instruction = prompt
};
var provider = llmProviderService.GetProviders().FirstOrDefault(x => x == "openai");
var model = llmProviderService.GetProviderModel(provider: provider, id: "gpt-4");
var completion = CompletionProvider.GetChatCompletion(_services, provider: provider, model: model.Name);
var response = await completion.GetChatCompletions(agent, dialogs);
var content = response?.Content ?? string.Empty;
var fid = JsonSerializer.Deserialize<int?>(content);
return images.Where((x, idx) => idx == fid - 1).FirstOrDefault();
}
catch (Exception ex)
{
_logger.LogWarning($"Error when getting the image edit response. {ex.Message}\r\n{ex.InnerException}");
return null;
}
}
private async Task<string> GetImageEditGeneration(RoleDialogModel message, string description, MessageFileModel? image)
{
if (image == null)
{
return "Failed to find an image. Please provide an image.";
}
try
{
var completion = CompletionProvider.GetImageCompletion(_services, provider: "openai", model: "dall-e-2");
var text = !string.IsNullOrWhiteSpace(description) ? description : message.Content;
var dialog = RoleDialogModel.From(message, AgentRole.User, text);
var agent = new Agent
{
Id = BuiltInAgentId.UtilityAssistant,
Name = "Utility Assistant"
};
using var stream = File.OpenRead(image.FileStorageUrl);
var result = await completion.GetImageEdits(agent, dialog, stream, image.FileName ?? string.Empty);
stream.Close();
SaveGeneratedImage(result?.GeneratedImages?.FirstOrDefault());
return !string.IsNullOrWhiteSpace(result?.Content) ? result.Content : "Image edit is completed.";
}
catch (Exception ex)
{
var error = $"Error when getting image edit response. {ex.Message}";
_logger.LogWarning($"{error}\r\n{ex.InnerException}");
return error;
}
}
private void SaveGeneratedImage(ImageGeneration? image)
{
if (image == null) return;
var files = new List<BotSharpFile>()
{
new BotSharpFile
{
FileName = $"{Guid.NewGuid()}.png",
FileData = $"data:{MediaTypeNames.Image.Png};base64,{image.ImageData}"
}
};
var fileService = _services.GetRequiredService<IBotSharpFileService>();
fileService.SaveMessageFiles(_conversationId, _messageId, FileSourceType.Bot, files);
}
}

View file

@ -20,7 +20,7 @@ public class GenerateImageFn : IFunctionCallback
public async Task<bool> Execute(RoleDialogModel message)
{
var args = JsonSerializer.Deserialize<LlmFileContext>(message.FunctionArgs);
var args = JsonSerializer.Deserialize<LlmContextIn>(message.FunctionArgs);
Init(message);
SetImageOptions();

View file

@ -5,6 +5,7 @@ public class FileHandlerHook : AgentHookBase, IAgentHook
private const string READ_IMAGE_FN = "read_image";
private const string READ_PDF_FN = "read_pdf";
private const string GENERATE_IMAGE_FN = "generate_image";
private const string EDIT_IMAGE_FN = "edit_image";
public override string SelfId => string.Empty;
@ -19,9 +20,11 @@ public class FileHandlerHook : AgentHookBase, IAgentHook
if (isConvMode)
{
AddUtility(agent, UtilityName.ImageReader, READ_IMAGE_FN);
AddUtility(agent, UtilityName.PdfReader, READ_PDF_FN);
AddUtility(agent, UtilityName.ImageGenerator, GENERATE_IMAGE_FN);
AddUtility(agent, UtilityName.ImageReader, READ_IMAGE_FN);
AddUtility(agent, UtilityName.ImageEditor, EDIT_IMAGE_FN);
AddUtility(agent, UtilityName.PdfReader, READ_PDF_FN);
}
base.OnAgentLoaded(agent);

View file

@ -4,8 +4,9 @@ public class FileHandlerUtilityHook : IAgentUtilityHook
{
public void AddUtilities(List<string> utilities)
{
utilities.Add(UtilityName.ImageReader);
utilities.Add(UtilityName.PdfReader);
utilities.Add(UtilityName.ImageGenerator);
utilities.Add(UtilityName.ImageReader);
utilities.Add(UtilityName.ImageEditor);
utilities.Add(UtilityName.PdfReader);
}
}

View file

@ -7,4 +7,8 @@ public class LlmContextIn
[JsonPropertyName("user_request")]
[JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
public string? UserRequest { get; set; }
[JsonPropertyName("image_description")]
[JsonIgnore(Condition = JsonIgnoreCondition.WhenWritingNull)]
public string? ImageDescription { get; set; }
}

View file

@ -0,0 +1,14 @@
{
"name": "edit_image",
"description": "If the user requests you editting or changing an image or a picture, you can call this function to edit an image.",
"parameters": {
"type": "object",
"properties": {
"user_request": {
"type": "string",
"description": "The request posted by user, which is related to editing the requested image."
}
},
"required": [ "user_request" ]
}
}

View file

@ -0,0 +1 @@
Please call edit_image if user wants to edit or change an image in the conversation.

View file

@ -0,0 +1,34 @@
Please take a look at the images in the [IMAGES] section from the conversation and select ONLY one image based on the conversation with user.
Your response must be an interger number.
** Please ONLY output the interger number. Do not prepend or append anything else.
** If you think user requests multiple images. Please ONLY select the first image and output its id.
Suppose there are three images:
id: 1, image_name: example_image_a.png
id: 2, image_name: example_image_b.png
id: 3, image_name: example_image_c.png
=====
Example 1:
USER: I want to add a dog in the first file.
OUTPUT: 1
Example 2:
USER: Add a coffee cup in the second image I uploaded.
OUTPUT: 2
Example 3:
USER: Please remove the left tree in the third and the first images.
OUTPUT: 3
Example 4:
USER: Add a boat in the images.
OUTPUT: 1
=====
[IMAGES]
{% for image in image_list -%}
{{ image }}{{ "\r\n" }}
{%- endfor %}