RAG实现：.Net + Ollama + Qdrant 实现文本向量化，实现简单RAG

本文使用.Net + Ollama + 向量数据库Qdrant，实现简单RAG，代码仅实现基本演示功能。

一、Ollama如何安装使用

请移步：https://blog.csdn.net/MrTraum/article/details/139240885

二、新建.NetCore WebApi项目，并导入Nuget包

本项目使用.Net8构建。

引入以下包：
Microsoft.Extensions.AI
LangChain
LangChain.Core
LangChain.DocumentLoaders.Pdf
Microsoft.Extensions.VectorData.Abstractions
OllamaSharp
Microsoft.SemanticKernel.Connectors.Qdrant

三、文本向量化+保存Qdrant

向量化模型使用：Qwen3-Embedding-0.6B
文本从Pdf导入，使用 IFormFile 接收

public class Content
{[VectorStoreKey]public Guid Key { get; set; }[VectorStoreData]public string Text { get; set; }[VectorStoreData]public string Source { get; set; }[VectorStoreVector(1024)]public ReadOnlyMemory<float> Vector { get; set; }
}

IDocumentLoader wordLoader = new PdfPigPdfLoader();
var dataSource = DataSource.FromStream(file.File.OpenReadStream());
var documents = await wordLoader.LoadAsync(dataSource: dataSource
);
//使用LangChain进行大文本分块
//此处采用递归分块算法
ITextSplitter recursiveSplitter = new RecursiveCharacterTextSplitter(chunkSize: 500,chunkOverlap: 128,separators: new[] { "\n\n", "\n", "。", "！", "？", "；", " ", "" } // 分隔符优先级
);
var splitDocuments = recursiveSplitter.SplitDocuments(documents);var vectorStore = new QdrantVectorStore(new QdrantClient("192.168.4.9", 6334), true);
var ragVectorRecordCollection = vectorStore.GetCollection<Guid, Content>("Contents");
await ragVectorRecordCollection.EnsureCollectionExistsAsync();var uri = new Uri(OllamaUrl);
var ollama = new OllamaApiClient(uri, "dengcao/Qwen3-Embedding-0.6B:Q8_0");
List<Content> vectorRecords = new List<Content>();
//调用大模型进行文本向量化
var embeddings = await ollama.GenerateAndZipAsync(splitDocuments.Select(t => t.PageContent));
foreach (var doc in splitDocuments)
{ReadOnlyMemory<float> embedding = null;foreach (var item in embeddings){if (item.Value == doc.PageContent){embedding = item.Embedding.Vector;}}Content ctx = new Content(){Key = Guid.NewGuid(),Text = doc.PageContent,Source = doc.Metadata.TryGetValue("source", out var src) ? src?.ToString() : "unknown",Vector = embedding};vectorRecords.Add(ctx);
}
//保存如Qdrant
await ragVectorRecordCollection.UpsertAsync(vectorRecords, CancellationToken.None);

四、搜索向量数据库，将搜索到的结果交给大模型整理

        [HttpGet]public async Task<string> Query([FromQuery] string query, [FromQuery] int top = 10){var vectorStore = new QdrantVectorStore(new QdrantClient("192.168.4.9", 6334), true);var ragVectorRecordCollection = vectorStore.GetCollection<Guid, Content>("Contents");await ragVectorRecordCollection.EnsureCollectionExistsAsync();var uri = new Uri(OllamaUrl);var ollama = new OllamaApiClient(uri, "dengcao/Qwen3-Embedding-0.6B:Q8_0");var searchVector = await ollama.GenerateVectorAsync(query);//以下代码，写不写都可以。默认会根据VectorStoreVector标签匹配VectorSearchOptions<Content> vectorSearchOptions = new VectorSearchOptions<Content>(){VectorProperty = m => m.Vector,};//查询向量数据库var searchResult = ragVectorRecordCollection.SearchAsync(searchVector, top, vectorSearchOptions);var resultsList = new List<VectorSearchResult<Content>>();await foreach (var result in searchResult){resultsList.Add(result);}// 如果没有找到结果if (!resultsList.Any()){return "没有找到相关的信息。";}// 构建精确控制的提示词var promptBuilder = new StringBuilder();// 角色定义（严格模式）promptBuilder.AppendLine(@"你是一个知识检索系统，必须遵守以下规则：
1. 回答范围：仅使用下方<!--reference-->标记的内容
2. 禁止行为：禁止联想、推测或添加任何外部知识
3. 免责声明：当参考内容不匹配时必须声明");// 结构化参考内容promptBuilder.AppendLine("<!--reference-->");foreach (var item in resultsList.OrderByDescending(x => x.Score)){promptBuilder.AppendLine($"| 内容：{item.Record.Text?.Trim()}");promptBuilder.AppendLine($"| 元数据：{item.Record.Source?.Trim()}");promptBuilder.AppendLine($"| 相似度：{item.Score:F4}");promptBuilder.AppendLine("|--"); }promptBuilder.AppendLine("<!--end-reference-->");// 问题与约束条件promptBuilder.AppendLine($"用户问题：{query?.Trim()}");promptBuilder.AppendLine(@"响应要求：1. 相似度高的条目优先2. 描述中必须包含问题关键词的直系近义词3. 采用以下响应模板：- 当匹配成功时：『根据参考内容，搜索到以下内容。』- 当匹配失败时：『根据提供的信息，未找到明确匹配的内容。』");// 调用大模型整理结果var chatRequest = new ChatRequest{Model = ModelName,Stream = false,Think = false,//禁止思考模式Messages = new List<Message>(){new Message { Role = OllamaSharp.Models.Chat.ChatRole.System, Content = promptBuilder.ToString() },new Message { Role = OllamaSharp.Models.Chat.ChatRole.User, Content = query }}};var modelResponse = await ollama.ChatAsync(chatRequest).StreamToEndAsync();return modelResponse.Message.Content;}