feat(OCR): 添加基于大模型的OCR作为PaddleOCR的替代方案

TelegramSearchBot.LLM.Test/Service/AI/LLM/GeneralLLMServiceTests.cs

@@ -214,5 +214,37 @@ public async Task AnalyzeImageAsync_NoChannels_ReturnsErrorMessage() {
             var result = await _service.AnalyzeImageAsync("/tmp/test.jpg", 123, CancellationToken.None);
             Assert.StartsWith("Error:", result);
         }
+
+        [Fact]
+        public async Task AnalyzeImageAsync_WithCustomPrompt_ForwardsPromptToProvider() {
+            var providerMock = new Mock<ILLMService>();
+            var channel = new LLMChannel {
+                Name = "vision-channel",
+                Gateway = "https://example.com",
+                ApiKey = "key",
+                Provider = LLMProvider.OpenAI,
+                Parallel = 1,
+                Priority = 1
+            };
+
+            providerMock
+                .Setup(s => s.AnalyzeImageAsync("image.jpg", "vision-model", channel, GeneralLLMService.DefaultVisionOcrPrompt))
+                .ReturnsAsync("recognized text");
+
+            var results = new List<string>();
+            await foreach (var result in _service.AnalyzeImageAsync(
+                "image.jpg",
+                123,
+                "vision-model",
+                providerMock.Object,
+                channel,
+                GeneralLLMService.DefaultVisionOcrPrompt,
+                CancellationToken.None)) {
+                results.Add(result);
+            }
+
+            Assert.Single(results);
+            Assert.Equal("recognized text", results[0]);
+        }
     }
 }

TelegramSearchBot.LLM/Interface/AI/LLM/IGeneralLLMService.cs

@@ -21,7 +21,9 @@ public interface IGeneralLLMService {
         IAsyncEnumerable<string> ResumeFromSnapshotAsync(LlmContinuationSnapshot snapshot, LlmExecutionContext executionContext, CancellationToken cancellationToken = default);
 
         Task<string> AnalyzeImageAsync(string PhotoPath, long ChatId, CancellationToken cancellationToken = default);
+        Task<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string prompt, CancellationToken cancellationToken = default);
         IAsyncEnumerable<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string modelName, ILLMService service, LLMChannel channel, CancellationToken cancellationToken = default);
+        IAsyncEnumerable<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string modelName, ILLMService service, LLMChannel channel, string prompt, CancellationToken cancellationToken = default);
         Task<float[]> GenerateEmbeddingsAsync(Message message, long ChatId);
         Task<float[]> GenerateEmbeddingsAsync(string message, CancellationToken cancellationToken = default);
         IAsyncEnumerable<float[]> GenerateEmbeddingsAsync(string message, string modelName, ILLMService service, LLMChannel channel, CancellationToken cancellationToken = default);

TelegramSearchBot.LLM/Interface/AI/LLM/ILLMService.cs

@@ -10,7 +10,7 @@
 namespace TelegramSearchBot.Interface.AI.LLM {
     public interface ILLMService {
         public IAsyncEnumerable<string> ExecAsync(Message message, long ChatId, string modelName, LLMChannel channel,
                                                   [System.Runtime.CompilerServices.EnumeratorCancellation] CancellationToken cancellationToken = default);
 
         /// <summary>
         /// Execute LLM with an execution context that can signal iteration limit reached
@@ -18,7 +18,7 @@
         /// </summary>
         public IAsyncEnumerable<string> ExecAsync(Message message, long ChatId, string modelName, LLMChannel channel,
                                                   LlmExecutionContext executionContext,
                                                   [System.Runtime.CompilerServices.EnumeratorCancellation] CancellationToken cancellationToken = default) {
             // Default: delegate to the original ExecAsync (backward compatible)
             return ExecAsync(message, ChatId, modelName, channel, cancellationToken);
         }
@@ -29,7 +29,7 @@
         /// </summary>
         public IAsyncEnumerable<string> ResumeFromSnapshotAsync(LlmContinuationSnapshot snapshot, LLMChannel channel,
                                                                  LlmExecutionContext executionContext,
                                                                  [System.Runtime.CompilerServices.EnumeratorCancellation] CancellationToken cancellationToken = default) {
             // Default: empty stream (services that don't support it just return nothing)
             return AsyncEnumerable.Empty<string>();
         }
@@ -42,7 +42,7 @@
         /// </summary>
         public Task<IEnumerable<ModelWithCapabilities>> GetAllModelsWithCapabilities(LLMChannel channel);
 
-        public Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel);
+        public Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel, string prompt = null);
         public virtual async Task<bool> IsHealthyAsync(LLMChannel channel) => ( await GetAllModels(channel) ).Any();
     }
 }

TelegramSearchBot.LLM/Service/AI/LLM/AnthropicService.cs

@@ -971,11 +971,13 @@ public Task<float[]> GenerateEmbeddingsAsync(string text, string modelName, LLMC
 
         #region Image Analysis
 
-        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel) {
+        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel, string prompt = null) {
             if (string.IsNullOrWhiteSpace(modelName)) {
                 modelName = "claude-sonnet-4-20250514";
             }
 
+            prompt = string.IsNullOrWhiteSpace(prompt) ? GeneralLLMService.DefaultAltPhotoPrompt : prompt;
+
             if (channel == null || string.IsNullOrWhiteSpace(channel.ApiKey)) {
                 _logger.LogError("{ServiceName}: Channel or ApiKey is not configured.", ServiceName);
                 return $"Error: {ServiceName} channel/apikey is not configured.";
@@ -990,8 +992,6 @@ public async Task<string> AnalyzeImageAsync(string photoPath, string modelName,
                 var imgArray = imgData.ToArray();
                 var base64Image = Convert.ToBase64String(imgArray);
 
-                var prompt = "请根据这张图片生成一句准确、详尽的中文alt文本，说明画面中重要的元素、场景和含义，避免使用'图中显示'或'这是一张图片'这类通用表达。";
-
                 var imageSource = new Base64ImageSource {
                     Data = base64Image,
                     MediaType = MediaType.ImagePng,

TelegramSearchBot.LLM/Service/AI/LLM/GeminiService.cs

@@ -458,11 +458,13 @@ public async Task<float[]> GenerateEmbeddingsAsync(string text, string modelName
             }
         }
 
-        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel) {
+        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel, string prompt = null) {
             if (string.IsNullOrWhiteSpace(modelName)) {
                 modelName = "gpt-4-vision-preview";
             }
 
+            prompt = string.IsNullOrWhiteSpace(prompt) ? GeneralLLMService.DefaultAltPhotoPrompt : prompt;
+
             if (channel == null || string.IsNullOrWhiteSpace(channel.Gateway) || string.IsNullOrWhiteSpace(channel.ApiKey)) {
                 _logger.LogError("{ServiceName}: Channel, Gateway or ApiKey is not configured.", ServiceName);
                 return $"Error: {ServiceName} channel/gateway/apikey is not configured.";
@@ -471,8 +473,6 @@ public async Task<string> AnalyzeImageAsync(string photoPath, string modelName,
             var googleAI = new GoogleAi(channel.ApiKey, client: _httpClientFactory.CreateClient());
             var model = googleAI.CreateGenerativeModel("models/" + modelName);
             try {
-                var prompt = $"请根据这张图片生成一句准确、详尽的中文alt文本，说明画面中重要的元素、场景和含义，避免使用'图中显示'或'这是一张图片'这类通用表达。";
-
                 var chat = model.StartChat();
 
 

TelegramSearchBot.LLM/Service/AI/LLM/GeneralLLMService.cs

@@ -31,6 +31,8 @@
         public const string MaxImageRetryCountKey = "LLM:MaxImageRetryCount";
         public const string AltPhotoModelName = "LLM:AltPhotoModelName";
         public const string EmbeddingModelName = "LLM:EmbeddingModelName";
+        public const string DefaultAltPhotoPrompt = "请根据这张图片生成一句准确、详尽的中文alt文本，说明画面中重要的元素、场景和含义，避免使用'图中显示'或'这是一张图片'这类通用表达。";
+        public const string DefaultVisionOcrPrompt = "请使用视觉能力识别并提取这张图片里可见的文字内容，优先输出原文，并按从上到下、从左到右的自然阅读顺序整理。尽量保留换行和段落，不要改写、总结或解释；只有在完全没有可识别文字时，才用一句中文简要描述图片主要内容。不要添加“图中显示”或“这是一张图片”这类套话。";
         public const int DefaultMaxRetryCount = 100;
         public const int DefaultMaxImageRetryCount = 1000;
 
@@ -104,7 +106,7 @@
                 yield return e;
             }
         }
         public async IAsyncEnumerable<string> ExecAsync(
             Model.Data.Message message,
             long ChatId,
             string modelName,
@@ -225,7 +227,13 @@
             _logger.LogWarning($"所有{modelName}关联的渠道当前都已满载，重试{maxRetries}次后放弃");
 
         }
-        public async Task<string> AnalyzeImageAsync(string PhotoPath, long ChatId, CancellationToken cancellationToken = default) {
+        public Task<string> AnalyzeImageAsync(string PhotoPath, long ChatId, CancellationToken cancellationToken = default) {
+            return AnalyzeImageAsync(PhotoPath, ChatId, DefaultAltPhotoPrompt, cancellationToken);
+        }
+
+        public async Task<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string prompt, CancellationToken cancellationToken = default) {
+            prompt = string.IsNullOrWhiteSpace(prompt) ? DefaultAltPhotoPrompt : prompt;
+
             // 1. 获取模型名称
             var modelName = "gemma3:27b";
             var config = await _dbContext.AppConfigurationItems
@@ -235,7 +243,7 @@
             }
 
             await using var enumerator = ExecOperationAsync<string>((service, channel, cancel) => {
-                return AnalyzeImageAsync(PhotoPath, ChatId, modelName, service, channel, cancel);
+                return AnalyzeImageAsync(PhotoPath, ChatId, modelName, service, channel, prompt, cancel);
             }, modelName, cancellationToken).GetAsyncEnumerator();
 
             if (await enumerator.MoveNextAsync()) {
@@ -245,8 +253,15 @@
             _logger.LogWarning($"未能获取 {modelName} 模型的图片分析结果");
             return $"Error:未能获取 {modelName} 模型的图片分析结果";
         }
         public async IAsyncEnumerable<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string modelName, ILLMService service, LLMChannel channel, CancellationToken cancellationToken = default) {
-            yield return await service.AnalyzeImageAsync(PhotoPath, modelName, channel);
+            await foreach (var result in AnalyzeImageAsync(PhotoPath, ChatId, modelName, service, channel, DefaultAltPhotoPrompt, cancellationToken)) {
+                yield return result;
+            }
+        }
+
+        public async IAsyncEnumerable<string> AnalyzeImageAsync(string PhotoPath, long ChatId, string modelName, ILLMService service, LLMChannel channel, string prompt, CancellationToken cancellationToken = default) {
+            prompt = string.IsNullOrWhiteSpace(prompt) ? DefaultAltPhotoPrompt : prompt;
+            yield return await service.AnalyzeImageAsync(PhotoPath, modelName, channel, prompt);
             yield break;
         }
 
@@ -280,7 +295,7 @@
             _logger.LogWarning($"未能获取 {modelName} 模型的嵌入向量");
             return Array.Empty<float>();
         }
         public async IAsyncEnumerable<float[]> GenerateEmbeddingsAsync(string message, string modelName, ILLMService service, LLMChannel channel, CancellationToken cancellationToken = default) {
             yield return await service.GenerateEmbeddingsAsync(message, modelName, channel);
             yield break;
         }

TelegramSearchBot.LLM/Service/AI/LLM/OllamaService.cs

@@ -389,16 +389,17 @@ public async Task<float[]> GenerateEmbeddingsAsync(string text, string modelName
             }
         }
 
-        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel) {
+        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel, string prompt = null) {
             if (string.IsNullOrWhiteSpace(modelName)) {
                 modelName = "gemma3:27b";
             }
 
+            prompt = string.IsNullOrWhiteSpace(prompt) ? GeneralLLMService.DefaultAltPhotoPrompt : prompt;
+
             var httpClient = _httpClientFactory?.CreateClient() ?? new HttpClient();
             httpClient.BaseAddress = new Uri(channel.Gateway);
             var ollama = new OllamaApiClient(httpClient, modelName);
             ollama.SelectedModel = modelName;
-            var prompt = "请根据这张图片生成一句准确、详尽的中文alt文本，说明画面中重要的元素、场景和含义，避免使用'图中显示'或'这是一张图片'这类通用表达。";
             var chat = new Chat(ollama);
             chat.Options = new RequestOptions();
             chat.Options.Temperature = 0.1f;

TelegramSearchBot.LLM/Service/AI/LLM/OpenAIService.cs

@@ -1461,11 +1461,13 @@ public async Task<string> GetModel(long ChatId) {
             return ModelName;
         }
 
-        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel) {
+        public async Task<string> AnalyzeImageAsync(string photoPath, string modelName, LLMChannel channel, string prompt = null) {
             if (string.IsNullOrWhiteSpace(modelName)) {
                 modelName = "gpt-4-vision-preview";
             }
 
+            prompt = string.IsNullOrWhiteSpace(prompt) ? GeneralLLMService.DefaultAltPhotoPrompt : prompt;
+
             if (channel == null || string.IsNullOrWhiteSpace(channel.Gateway) || string.IsNullOrWhiteSpace(channel.ApiKey)) {
                 _logger.LogError("{ServiceName}: Channel, Gateway or ApiKey is not configured.", ServiceName);
                 return $"Error: {ServiceName} channel/gateway/apikey is not configured.";
@@ -1488,8 +1490,6 @@ public async Task<string> AnalyzeImageAsync(string photoPath, string modelName,
                 var tg_img_arr = tg_img_data.ToArray();
                 var base64Image = Convert.ToBase64String(tg_img_arr);
 
-                var prompt = $"请根据这张图片生成一句准确、详尽的中文alt文本，说明画面中重要的元素、场景和含义，避免使用'图中显示'或'这是一张图片'这类通用表达。";
-
                 var messages = new List<ChatMessage> {
                     new UserChatMessage(new List<ChatMessageContentPart>() {
                         ChatMessageContentPart.CreateTextPart(prompt),

TelegramSearchBot.Test/Manage/EditOCRConfServiceTests.cs

@@ -0,0 +1,115 @@
+using System;
+using System.Collections.Generic;
+using System.Linq;
+using System.Threading.Tasks;
+using Microsoft.EntityFrameworkCore;
+using Microsoft.Extensions.Logging;
+using Moq;
+using StackExchange.Redis;
+using Telegram.Bot.Types.ReplyMarkups;
+using TelegramSearchBot.Model.AI;
+using TelegramSearchBot.Model;
+using TelegramSearchBot.Model.Data;
+using TelegramSearchBot.Service.Manage;
+using Xunit;
+
+namespace TelegramSearchBot.Test.Manage {
+    public class EditOCRConfServiceTests {
+        private readonly DataDbContext _dbContext;
+        private readonly Mock<IConnectionMultiplexer> _redisMock;
+        private readonly Mock<IDatabase> _dbMock;
+        private readonly Dictionary<string, RedisValue> _redisStore = new();
+        private readonly EditOCRConfService _service;
+
+        public EditOCRConfServiceTests() {
+            var options = new DbContextOptionsBuilder<DataDbContext>()
+                .UseInMemoryDatabase(Guid.NewGuid().ToString())
+                .Options;
+            _dbContext = new DataDbContext(options);
+
+            _redisMock = new Mock<IConnectionMultiplexer>();
+            _dbMock = new Mock<IDatabase>();
+            _redisMock.Setup(r => r.GetDatabase(It.IsAny<int>(), It.IsAny<object>()))
+                .Returns(_dbMock.Object);
+
+            _dbMock.Setup(d => d.StringGetAsync(It.IsAny<RedisKey>(), It.IsAny<CommandFlags>()))
+                .ReturnsAsync((RedisKey key, CommandFlags flags) =>
+                    _redisStore.TryGetValue(key.ToString(), out var value) ? value : RedisValue.Null);
+
+            _dbMock.Setup(d => d.StringSetAsync(
+                    It.IsAny<RedisKey>(),
+                    It.IsAny<RedisValue>(),
+                    It.IsAny<TimeSpan?>(),
+                    It.IsAny<When>()))
+                .ReturnsAsync((RedisKey key, RedisValue value, TimeSpan? expiry, When when) => {
+                    _redisStore[key.ToString()] = value;
+                    return true;
+                });
+
+            _dbMock.Setup(d => d.StringSetAsync(
+                    It.IsAny<RedisKey>(),
+                    It.IsAny<RedisValue>(),
+                    It.IsAny<TimeSpan?>(),
+                    It.IsAny<When>(),
+                    It.IsAny<CommandFlags>()))
+                .ReturnsAsync((RedisKey key, RedisValue value, TimeSpan? expiry, When when, CommandFlags flags) => {
+                    _redisStore[key.ToString()] = value;
+                    return true;
+                });
+
+            _dbMock.Setup(d => d.StringSetAsync(
+                    It.IsAny<RedisKey>(),
+                    It.IsAny<RedisValue>(),
+                    It.IsAny<TimeSpan?>(),
+                    It.IsAny<bool>(),
+                    It.IsAny<When>(),
+                    It.IsAny<CommandFlags>()))
+                .ReturnsAsync((RedisKey key, RedisValue value, TimeSpan? expiry, bool keepTtl, When when, CommandFlags flags) => {
+                    _redisStore[key.ToString()] = value;
+                    return true;
+                });
+
+            _dbMock.Setup(d => d.KeyDeleteAsync(It.IsAny<RedisKey>(), It.IsAny<CommandFlags>()))
+                .ReturnsAsync((RedisKey key, CommandFlags flags) => _redisStore.Remove(key.ToString()));
+
+            var loggerMock = new Mock<ILogger<EditOCRConfService>>();
+            _service = new EditOCRConfService(_dbContext, _redisMock.Object, loggerMock.Object);
+        }
+
+        [Fact]
+        public async Task ExecuteAsync_UnrelatedCommandWithoutActiveSession_ReturnsFalse() {
+            var (status, message) = await _service.ExecuteAsync("你好", 12345);
+
+            Assert.False(status);
+            Assert.Equal(string.Empty, message);
+        }
+
+        [Fact]
+        public async Task ExecuteAsync_SwitchEngineFlow_UsesReplyKeyboardAndTransitionsState() {
+            _dbContext.LLMChannels.Add(new LLMChannel {
+                Name = "vision-channel",
+                Gateway = "https://example.com",
+                ApiKey = "key",
+                Provider = LLMProvider.OpenAI,
+                Parallel = 1,
+                Priority = 1
+            });
+            await _dbContext.SaveChangesAsync();
+
+            _redisStore["ocrconf:100:state"] = OCRConfState.MainMenu.GetDescription();
+            var mainMenuMarkup = await _service.GetReplyMarkupAsync(100);
+            var mainMenuKeyboard = Assert.IsType<ReplyKeyboardMarkup>(mainMenuMarkup);
+            Assert.Contains(mainMenuKeyboard.Keyboard, row => row.Any(button => button.Text == "切换OCR引擎"));
+
+            var (switched, engineMessage) = await _service.ExecuteAsync("切换OCR引擎", 100);
+            Assert.True(switched);
+            Assert.Contains("请选择 OCR 引擎", engineMessage);
+
+            _redisStore["ocrconf:100:state"] = OCRConfState.SelectingEngine.GetDescription();
+            var engineMarkup = await _service.GetReplyMarkupAsync(100);
+            var engineKeyboard = Assert.IsType<ReplyKeyboardMarkup>(engineMarkup);
+            Assert.Contains(engineKeyboard.Keyboard, row => row.Any(button => button.Text == "PaddleOCR"));
+            Assert.Contains(engineKeyboard.Keyboard, row => row.Any(button => button.Text == "LLM"));
+        }
+    }
+}

TelegramSearchBot.Test/Service/AI/OCR/LLMOCRServiceTests.cs

@@ -0,0 +1,43 @@
+using System.IO;
+using System.Threading;
+using System.Threading.Tasks;
+using Microsoft.Extensions.Logging;
+using Moq;
+using SkiaSharp;
+using TelegramSearchBot.Interface.AI.LLM;
+using TelegramSearchBot.Service.AI.LLM;
+using TelegramSearchBot.Service.AI.OCR;
+using Xunit;
+
+namespace TelegramSearchBot.Test.Service.AI.OCR {
+    public class LLMOCRServiceTests {
+        [Fact]
+        public async Task ExecuteAsync_UsesVisionOcrPrompt() {
+            var llmServiceMock = new Mock<IGeneralLLMService>();
+            llmServiceMock
+                .Setup(s => s.AnalyzeImageAsync(
+                    It.Is<string>(path => path.EndsWith(".jpg")),
+                    0,
+                    GeneralLLMService.DefaultVisionOcrPrompt,
+                    It.IsAny<CancellationToken>()))
+                .ReturnsAsync("recognized text");
+
+            var loggerMock = new Mock<ILogger<LLMOCRService>>();
+            var service = new LLMOCRService(llmServiceMock.Object, loggerMock.Object);
+
+            using var bitmap = new SKBitmap(8, 8);
+            using var image = SKImage.FromBitmap(bitmap);
+            using var data = image.Encode(SKEncodedImageFormat.Png, 100);
+            using var stream = new MemoryStream(data.ToArray());
+
+            var result = await service.ExecuteAsync(stream);
+
+            Assert.Equal("recognized text", result);
+            llmServiceMock.Verify(s => s.AnalyzeImageAsync(
+                It.Is<string>(path => path.EndsWith(".jpg")),
+                0,
+                GeneralLLMService.DefaultVisionOcrPrompt,
+                It.IsAny<CancellationToken>()), Times.Once);
+        }
+    }
+}