# 为您的博客建设“律师GPT”-第8部分:高级地物和制作部署

<!--category-- AI, LLM, Deployment, Production, C#, AI-Article, mostlylucid.blogllm -->
<datetime class="hidden">2025-11-12T22:45</datetime>

警告:这些是“加入”的草稿。

可能很多下面的东西是行不通的; 我制作了这些作为给ME的操作方法, 然后做所有步骤,让样本应用起作用...你一直偷偷摸摸地看到它们!它们很可能在12月中旬就绪。

<img src="https://media1.tenor.com/m/_rQc7PIEqwQAAAAd/cat-hello-cat-peek.gif" height="300px" />
## 一. 导言 导言 导言 导言 导言 导言 一,导言 导言 导言 导言 导言 导言

欢迎来到第八部分 最后一部分![我们建造了一个完整的](https://www.anthropic.com/index/contextual-retrieval)RAG 区域包

> - 基础写作助理从头开始。

现在让我们加上一个能使其生产就绪的抛光: 自动连接、部署、配置管理 以及现实世界的使用模式。

[TOC]

## 注:这是我对人工智能(协助起草)和我自己编辑的实验的一部分。

同一个声音,同样的务实;只是更快的手指。

### 我们在这里使用一个工作原型 把它变成你每天都会使用的东西。

```csharp
namespace Mostlylucid.BlogLLM.Core.Services
{
    public interface ILinkSuggestionService
    {
        Task<List<LinkSuggestion>> SuggestLinksAsync(string text);
        string InsertLinks(string text, List<LinkSuggestion> acceptedLinks);
    }

    public class LinkSuggestion
    {
        public string Phrase { get; set; } = string.Empty;
        public string TargetSlug { get; set; } = string.Empty;
        public string TargetTitle { get; set; } = string.Empty;
        public float RelevanceScore { get; set; }
        public int Position { get; set; }
    }

    public class LinkSuggestionService : ILinkSuggestionService
    {
        private readonly BatchEmbeddingService _embedder;
        private readonly QdrantVectorStore _vectorStore;
        private readonly ILogger<LinkSuggestionService> _logger;

        public LinkSuggestionService(
            BatchEmbeddingService embedder,
            QdrantVectorStore vectorStore,
            ILogger<LinkSuggestionService> logger)
        {
            _embedder = embedder;
            _vectorStore = vectorStore;
            _logger = logger;
        }

        public async Task<List<LinkSuggestion>> SuggestLinksAsync(string text)
        {
            var suggestions = new List<LinkSuggestion>();

            // Extract key phrases (noun phrases, technical terms)
            var phrases = ExtractKeyPhrases(text);

            _logger.LogInformation("Extracted {Count} key phrases for linking", phrases.Count);

            foreach (var phrase in phrases)
            {
                // Search for related posts
                var embedding = _embedder.GenerateEmbedding(phrase.Text);
                var results = await _vectorStore.SearchAsync(
                    queryEmbedding: embedding,
                    limit: 3,
                    scoreThreshold: 0.75f  // High threshold for links
                );

                if (results.Any())
                {
                    var topResult = results.First();

                    // Don't link to current post
                    if (!IsCurrentPost(topResult.BlogPostSlug, text))
                    {
                        suggestions.Add(new LinkSuggestion
                        {
                            Phrase = phrase.Text,
                            TargetSlug = topResult.BlogPostSlug,
                            TargetTitle = topResult.BlogPostTitle,
                            RelevanceScore = topResult.Score,
                            Position = phrase.Position
                        });

                        _logger.LogDebug("Link suggestion: '{Phrase}' -> '{Target}' (score: {Score:F3})",
                            phrase.Text, topResult.BlogPostTitle, topResult.Score);
                    }
                }
            }

            // Remove duplicates and low-value links
            return DeduplicateAndFilter(suggestions);
        }

        public string InsertLinks(string text, List<LinkSuggestion> acceptedLinks)
        {
            // Sort by position (descending) to maintain positions as we insert
            var sorted = acceptedLinks.OrderByDescending(l => l.Position).ToList();

            foreach (var link in sorted)
            {
                var before = text.Substring(0, link.Position);
                var phrase = link.Phrase;
                var after = text.Substring(link.Position + phrase.Length);

                // Check if already a link
                if (IsAlreadyLinked(before, phrase, after))
                {
                    continue;
                }

                var markdownLink = $"[{phrase}](/blog/{link.TargetSlug})";
                text = before + markdownLink + after;

                _logger.LogInformation("Inserted link: {Phrase} -> /blog/{Slug}",
                    phrase, link.TargetSlug);
            }

            return text;
        }

        private List<KeyPhrase> ExtractKeyPhrases(string text)
        {
            var phrases = new List<KeyPhrase>();

            // Simple regex-based extraction
            // In production, use NLP library like Stanford.NLP or Azure Cognitive Services

            // Technical terms (CamelCase, dot notation)
            var technicalTerms = Regex.Matches(text,
                @"\b([A-Z][a-z]+([A-Z][a-z]+)+|[A-Z]\w+\.\w+)\b");

            foreach (Match match in technicalTerms)
            {
                phrases.Add(new KeyPhrase
                {
                    Text = match.Value,
                    Position = match.Index
                });
            }

            // Multi-word phrases in quotes or code backticks
            var quotedPhrases = Regex.Matches(text, @"[`""]([^`""]{10,50})[`""]");

            foreach (Match match in quotedPhrases)
            {
                phrases.Add(new KeyPhrase
                {
                    Text = match.Groups[1].Value,
                    Position = match.Index
                });
            }

            return phrases;
        }

        private List<LinkSuggestion> DeduplicateAndFilter(List<LinkSuggestion> suggestions)
        {
            // Remove duplicate phrases (keep highest score)
            var deduped = suggestions
                .GroupBy(s => s.Phrase.ToLowerInvariant())
                .Select(g => g.OrderByDescending(s => s.RelevanceScore).First())
                .ToList();

            // Limit links per post
            return deduped
                .OrderByDescending(s => s.RelevanceScore)
                .Take(5)  // Max 5 auto-links per draft
                .ToList();
        }

        private bool IsCurrentPost(string slug, string text)
        {
            // Simple heuristic - check if slug appears in text
            // In production, track actual post being edited
            return text.ToLowerInvariant().Contains(slug.ToLowerInvariant());
        }

        private bool IsAlreadyLinked(string before, string phrase, string after)
        {
            // Check if phrase is already inside markdown link
            var lookback = before.TakeLast(20).ToString() ?? "";
            var lookahead = new string(after.Take(20).ToArray());

            return lookback.Contains("[") || lookahead.StartsWith("]");
        }
    }

    public class KeyPhrase
    {
        public string Text { get; set; } = string.Empty;
        public int Position { get; set; }
    }
}
```

### 让我们完成强大的!

```csharp
[RelayCommand]
private async Task SuggestLinks()
{
    IsProcessing = true;
    StatusMessage = "Analyzing text for link opportunities...";

    try
    {
        var suggestions = await _linkService.SuggestLinksAsync(EditorText);

        // Show suggestions in UI
        LinkSuggestions.Clear();
        foreach (var suggestion in suggestions)
        {
            LinkSuggestions.Add(new LinkSuggestionViewModel(suggestion));
        }

        StatusMessage = $"Found {suggestions.Count} link opportunities";
    }
    finally
    {
        IsProcessing = false;
    }
}

[RelayCommand]
private void AcceptAllLinks()
{
    var acceptedLinks = LinkSuggestions
        .Where(vm => vm.IsAccepted)
        .Select(vm => vm.Suggestion)
        .ToList();

    EditorText = _linkService.InsertLinks(EditorText, acceptedLinks);

    StatusMessage = $"Inserted {acceptedLinks.Count} links";
}
```

## 与相关员额自动链接

最宝贵的特点之一是自动建议与相关员额挂钩。

```csharp
namespace Mostlylucid.BlogLLM.Configuration
{
    public class BlogLLMConfig
    {
        public EmbeddingConfig Embedding { get; set; } = new();
        public VectorStoreConfig VectorStore { get; set; } = new();
        public LLMConfig LLM { get; set; } = new();
        public UIConfig UI { get; set; } = new();
    }

    public class EmbeddingConfig
    {
        public string ModelPath { get; set; } = string.Empty;
        public string TokenizerPath { get; set; } = string.Empty;
        public bool UseGpu { get; set; } = true;
        public int BatchSize { get; set; } = 32;
    }

    public class VectorStoreConfig
    {
        public string Type { get; set; } = "Qdrant";  // or "pgvector"
        public string Host { get; set; } = "localhost";
        public int Port { get; set; } = 6334;
        public string CollectionName { get; set; } = "blog_embeddings";
    }

    public class LLMConfig
    {
        public string ModelPath { get; set; } = string.Empty;
        public int ContextSize { get; set; } = 4096;
        public int GpuLayers { get; set; } = 35;
        public float DefaultTemperature { get; set; } = 0.7f;
        public int MaxTokens { get; set; } = 500;
        public bool EnableStreaming { get; set; } = true;
    }

    public class UIConfig
    {
        public int AutoSaveIntervalSeconds { get; set; } = 60;
        public bool EnableAutoLinking { get; set; } = true;
        public int SuggestionDebounceMs { get; set; } = 500;
        public int MaxRecentFiles { get; set; } = 10;
    }
}
```

### 链接探测处

```json
{
  "BlogLLM": {
    "Embedding": {
      "ModelPath": "C:\\models\\bge-base-en-onnx\\model.onnx",
      "TokenizerPath": "C:\\models\\bge-base-en-onnx\\tokenizer.json",
      "UseGpu": true,
      "BatchSize": 32
    },
    "VectorStore": {
      "Type": "Qdrant",
      "Host": "localhost",
      "Port": 6334,
      "CollectionName": "blog_embeddings"
    },
    "LLM": {
      "ModelPath": "C:\\models\\mistral-7b\\mistral-7b-instruct-v0.2.Q5_K_M.gguf",
      "ContextSize": 4096,
      "GpuLayers": 35,
      "DefaultTemperature": 0.7,
      "MaxTokens": 500,
      "EnableStreaming": true
    },
    "UI": {
      "AutoSaveIntervalSeconds": 60,
      "EnableAutoLinking": true,
      "SuggestionDebounceMs": 500,
      "MaxRecentFiles": 10
    }
  },
  "Logging": {
    "LogLevel": {
      "Default": "Information",
      "Mostlylucid.BlogLLM": "Debug"
    }
  }
}
```

### UI 整合

```csharp
public class App : Application
{
    public override void OnFrameworkInitializationCompleted()
    {
        var services = new ServiceCollection();

        // Load configuration
        var configuration = new ConfigurationBuilder()
            .SetBasePath(Directory.GetCurrentDirectory())
            .AddJsonFile("appsettings.json", optional: false)
            .AddJsonFile($"appsettings.{Environment.GetEnvironmentVariable("ENVIRONMENT")}.json", optional: true)
            .AddEnvironmentVariables()
            .Build();

        // Bind configuration
        var config = new BlogLLMConfig();
        configuration.GetSection("BlogLLM").Bind(config);

        // Register as singleton
        services.AddSingleton(config);

        // Register services using config
        services.AddSingleton(sp => new BatchEmbeddingService(
            config.Embedding.ModelPath,
            config.Embedding.TokenizerPath,
            config.Embedding.UseGpu
        ));

        // ... rest of service registration
    }
}
```

## 配置管理

### 生产准备配置系统:

```bash
# Publish as single-file executable
dotnet publish Mostlylucid.BlogLLM.Client/Mostlylucid.BlogLLM.Client.csproj \
    -c Release \
    -r win-x64 \
    --self-contained true \
    -p:PublishSingleFile=true \
    -p:IncludeNativeLibrariesForSelfExtract=true \
    -o ./publish/win-x64

# Result: Single .exe file with all dependencies
```

### 缩略语.json

```xml
<?xml version="1.0" encoding="UTF-8"?>
<!-- Using WiX Toolset: https://wixtoolset.org/ -->
<Wix xmlns="http://schemas.microsoft.com/wix/2006/wi">
    <Product Id="*" Name="Blog Writing Assistant" Language="1033"
             Version="1.0.0.0" Manufacturer="YourName" UpgradeCode="PUT-GUID-HERE">

        <Package InstallerVersion="200" Compressed="yes" InstallScope="perMachine" />

        <MediaTemplate EmbedCab="yes" />

        <Directory Id="TARGETDIR" Name="SourceDir">
            <Directory Id="ProgramFiles64Folder">
                <Directory Id="INSTALLFOLDER" Name="BlogLLM" />
            </Directory>
            <Directory Id="ProgramMenuFolder">
                <Directory Id="ApplicationProgramsFolder" Name="Blog Writing Assistant"/>
            </Directory>
        </Directory>

        <DirectoryRef Id="INSTALLFOLDER">
            <Component Id="MainExecutable" Guid="PUT-GUID-HERE">
                <File Id="BlogLLMExe" Source="$(var.PublishDir)\BlogLLM.exe" KeyPath="yes" />
            </Component>
            <Component Id="ConfigFile" Guid="PUT-GUID-HERE">
                <File Id="AppSettings" Source="$(var.PublishDir)\appsettings.json" />
            </Component>
        </DirectoryRef>

        <DirectoryRef Id="ApplicationProgramsFolder">
            <Component Id="ApplicationShortcut" Guid="PUT-GUID-HERE">
                <Shortcut Id="ApplicationStartMenuShortcut"
                         Name="Blog Writing Assistant"
                         Target="[INSTALLFOLDER]BlogLLM.exe"
                         WorkingDirectory="INSTALLFOLDER"/>
                <RemoveFolder Id="ApplicationProgramsFolder" On="uninstall"/>
                <RegistryValue Root="HKCU" Key="Software\BlogLLM" Name="installed" Type="integer" Value="1" KeyPath="yes"/>
            </Component>
        </DirectoryRef>

        <Feature Id="ProductFeature" Title="Blog Writing Assistant" Level="1">
            <ComponentRef Id="MainExecutable" />
            <ComponentRef Id="ConfigFile" />
            <ComponentRef Id="ApplicationShortcut" />
        </Feature>
    </Product>
</Wix>
```

### 装入配置

```powershell
# download-models.ps1
param(
    [string]$ModelsPath = "C:\models"
)

Write-Host "Downloading models to $ModelsPath..." -ForegroundColor Green

# Create directories
New-Item -ItemType Directory -Force -Path "$ModelsPath\bge-base-en-onnx" | Out-Null
New-Item -ItemType Directory -Force -Path "$ModelsPath\mistral-7b" | Out-Null

# Install huggingface-cli if needed
$hfCli = Get-Command huggingface-cli -ErrorAction SilentlyContinue
if (-not $hfCli) {
    Write-Host "Installing huggingface-cli..." -ForegroundColor Yellow
    pip install huggingface-hub
}

# Download embedding model
Write-Host "Downloading BGE embedding model..." -ForegroundColor Green
huggingface-cli download BAAI/bge-base-en-v1.5-onnx `
    --local-dir "$ModelsPath\bge-base-en-onnx" `
    --local-dir-use-symlinks False

# Download LLM
Write-Host "Downloading Mistral 7B model..." -ForegroundColor Green
huggingface-cli download TheBloke/Mistral-7B-Instruct-v0.2-GGUF `
    mistral-7b-instruct-v0.2.Q5_K_M.gguf `
    --local-dir "$ModelsPath\mistral-7b" `
    --local-dir-use-symlinks False

Write-Host "Download complete!" -ForegroundColor Green
Write-Host "Update appsettings.json with these paths:"
Write-Host "  Embedding: $ModelsPath\bge-base-en-onnx\model.onnx"
Write-Host "  LLM: $ModelsPath\mistral-7b\mistral-7b-instruct-v0.2.Q5_K_M.gguf"
```

## 部署战略

### 独立可执行文件

```csharp
public class UsageTracker
{
    private readonly string _usageFilePath;

    public UsageTracker(string dataPath)
    {
        _usageFilePath = Path.Combine(dataPath, "usage.json");
    }

    public void TrackSuggestionAccepted(string suggestionType, int length)
    {
        var usage = LoadUsage();
        usage.SuggestionsAccepted++;
        usage.TotalCharactersGenerated += length;
        usage.LastUsed = DateTime.UtcNow;

        SaveUsage(usage);
    }

    public void TrackLinkInserted(string targetPost)
    {
        var usage = LoadUsage();
        usage.LinksInserted++;

        if (!usage.FrequentlyLinkedPosts.ContainsKey(targetPost))
        {
            usage.FrequentlyLinkedPosts[targetPost] = 0;
        }
        usage.FrequentlyLinkedPosts[targetPost]++;

        SaveUsage(usage);
    }

    public UsageStats GetStats()
    {
        return LoadUsage();
    }

    private UsageStats LoadUsage()
    {
        if (!File.Exists(_usageFilePath))
        {
            return new UsageStats();
        }

        var json = File.ReadAllText(_usageFilePath);
        return JsonSerializer.Deserialize<UsageStats>(json) ?? new UsageStats();
    }

    private void SaveUsage(UsageStats stats)
    {
        var json = JsonSerializer.Serialize(stats, new JsonSerializerOptions
        {
            WriteIndented = true
        });

        File.WriteAllText(_usageFilePath, json);
    }
}

public class UsageStats
{
    public int SuggestionsGenerated { get; set; }
    public int SuggestionsAccepted { get; set; }
    public int LinksInserted { get; set; }
    public int TotalCharactersGenerated { get; set; }
    public Dictionary<string, int> FrequentlyLinkedPosts { get; set; } = new();
    public DateTime LastUsed { get; set; }
    public DateTime FirstUsed { get; set; } = DateTime.UtcNow;
}
```

### 使用 WiX 安装器

```csharp
[RelayCommand]
private async Task ProvideFeedback(GenerationResult result, FeedbackType type)
{
    var feedback = new SuggestionFeedback
    {
        GeneratedText = result.GeneratedText,
        PromptType = result.PromptType,
        ContextTokens = result.ContextTokensUsed,
        FeedbackType = type,
        Timestamp = DateTime.UtcNow
    };

    await _feedbackService.RecordAsync(feedback);

    // Adjust parameters based on feedback
    if (type == FeedbackType.TooGeneric)
    {
        // Increase temperature for more creativity
        _config.LLM.DefaultTemperature = Math.Min(1.0f, _config.LLM.DefaultTemperature + 0.1f);
    }
    else if (type == FeedbackType.TooRambling)
    {
        // Decrease temperature for more focus
        _config.LLM.DefaultTemperature = Math.Max(0.3f, _config.LLM.DefaultTemperature - 0.1f);
    }
}

public enum FeedbackType
{
    Helpful,
    TooGeneric,
    TooRambling,
    WrongStyle,
    Perfect
}
```

## 模型下载脚本

### 不断改进

```csharp
// User opens app
// Loads yesterday's draft

[RelayCommand]
private async Task ContinueFromYesterday()
{
    // Load last saved draft
    var draft = await LoadLastDraft();
    EditorText = draft.Content;

    // Generate fresh suggestions based on overnight ingestion
    await RefreshSuggestions();
}
```

### 跟踪使用情况

1. **反馈循环**真实世界使用模式
2. **早 晨 例 例**书写流程
3. **外衬**:反复使用“建议结构”构建大纲
4. **起草**: 类型介绍,请AI建议继续
5. **守则示例**: 请求使用上下文的代码生成

### 链接

```csharp
[RelayCommand]
private async Task ProcessAllDrafts()
{
    var drafts = GetAllDraftFiles();

    foreach (var draft in drafts)
    {
        var content = await File.ReadAllTextAsync(draft);

        // Suggest links for each draft
        var links = await _linkService.SuggestLinksAsync(content);

        // Auto-accept high-confidence links
        var autoAccept = links.Where(l => l.RelevanceScore > 0.9f).ToList();

        var updated = _linkService.InsertLinks(content, autoAccept);
        await File.WriteAllTextAsync(draft, updated);

        _logger.LogInformation("Processed {Draft}: inserted {Count} links",
            Path.GetFileName(draft), autoAccept.Count);
    }
}
```

## :当草稿完成 80% 完成时运行自动链接

### 波兰语Name

**:在较弱部分使用“ 整合部分” 。**

```
Solution: Reduce GpuLayers or ContextSize in config:
{
  "LLM": {
    "GpuLayers": 20,  // Lower from 35
    "ContextSize": 2048  // Lower from 4096
  }
}
```

**批次处理**

```
Solution: Increase context tokens and adjust temperature:
{
  "LLM": {
    "DefaultTemperature": 0.8  // Higher = more creative
  }
}
And in code: request.MaxContextTokens = 3000  // More context
```

**排除麻烦指南**

```
Check:
1. Is model fully on GPU? (Check GpuLayers = 35)
2. Using Q5 or Q4 quantization? (Faster than Q8)
3. Is something else using GPU? (Check nvidia-smi)
```

**共同问题 共同问题**

```
Solution: Increase scoreThreshold:
scoreThreshold: 0.85f  // Higher threshold = only very relevant links
```

## 问题: "CUDA失忆"

### 问题: "建议太笼统"

1. **问题:"启动速度缓慢"**问题:“链接无关紧要”
2. **未来增强能力**V2 的想法
3. **多种语文支助**- 超越英语
4. **图像建议**- 从过去的文章中找到相关的图像
5. **SEO优化**- 建议元说明,关键词
6. **高血压检测**- 与现有内容核对
7. **声音一致性**- 专门培训你的写作风格
8. **协作特点**- 多作者支持

### 网络版本

- - Blazor WebAssembly客户端
- 移动应用程序
- - Avalonia在iOS/Android上工作! - 是的。 - Avalonia在iOS/Android上工作!
- 研究方向

## 微调嵌入模式

将检索分数用作培训信号

```mermaid
graph TB
    subgraph "Part 1: Architecture"
        A1[System Design]
        A2[Technology Choices]
    end

    subgraph "Part 2: GPU Setup"
        B1[CUDA Installation]
        B2[cuDNN Setup]
        B3[Testing]
    end

    subgraph "Part 3: Embeddings"
        C1[Embedding Models]
        C2[Vector Databases]
        C3[Semantic Search]
    end

    subgraph "Part 4: Ingestion"
        D1[Markdown Parsing]
        D2[Chunking]
        D3[Embedding Generation]
        D4[Vector Storage]
    end

    subgraph "Part 5: UI"
        E1[Avalonia Client]
        E2[Editor Component]
        E3[Suggestions Panel]
    end

    subgraph "Part 6: LLM"
        F1[LLamaSharp]
        F2[Model Loading]
        F3[Inference]
    end

    subgraph "Part 7: Generation"
        G1[Context Building]
        G2[Prompt Engineering]
        G3[Content Generation]
    end

    subgraph "Part 8: Production"
        H1[Auto-linking]
        H2[Configuration]
        H3[Deployment]
        H4[Monitoring]
    end

    A1 --> B1
    B3 --> C1
    C3 --> D1
    D4 --> E1
    E3 --> F1
    F3 --> G1
    G3 --> H1

    class A1,F3 architecture
    class D4,E1 data
    class G3,H3 production

    classDef architecture stroke:#333
    classDef data stroke:#333
    classDef production stroke:#333,stroke-width:4px
```

## 与大型模型(13B, 30B)在云层 GPU 上实验

实施多式RAG(代码+图表+文本)

1. **完整图片**让我们想象一下我们所建造的一切:
2. **结论 结论 结论 结论 结论**我们成功了!
3. **超过8个完整的部件, 我们建立了一个完整的, 制作准备就绪的写作助理:**第一部分 第一部分
4. **:设计架构和选择技术**第二部分 第二部分
5. **建立 CUDA 和 GPU 加速度**第三部分 第三部分
6. **:已实施嵌入和矢量搜索**第四部分 第四部分
7. **:建造输入管道**第5部分 第五部分
8. **: 创建了一个美丽的 Windows 客户端**第6部分 第六部分

**:当地LLM综合推理**:

- 第7部分 第七部分
- :精密精密电讯
- 第8部分
- :加油和部署
- 这有什么特别的?
- 100%地方和私人

**无APPI费用**:

- 快速 GPU 加速加速推断
- 基于你的实际内容
- 可用于生产代码
- 跨平台潜力
- 现实世界的影响

快速博客写作

一贯的风格和声音

## 改善内部连接

旧内容的再使用

- 出版障碍降低
- 这不仅仅是一个辅导项目 - 这是一个真正有用的工具 帮助您写作更好,更快, 同时保持与您现有工作的一致性。
- 律师如何利用受过判例法培训的LLMs起草更好的简报,
- 谢谢!
- 感谢大家继续关注这整个系列。

**我希望你已经学到了:**:

1. RAG 系统如何运作
2. 以 C # 表示的 GPU 加速 AI
3. 矢量数据库和嵌入
4. 当地LLM 部署

## 生产应用程序建筑

### 下一步步骤

- [克隆回波(即将到来)](/blog/building-a-lawyer-gpt-for-your-blog-part1)
- [下载模式](/blog/building-a-lawyer-gpt-for-your-blog-part2)
- [运行输入管道](/blog/building-a-lawyer-gpt-for-your-blog-part3)
- [以AI援助开始写作!](/blog/building-a-lawyer-gpt-for-your-blog-part4)
- [资源资源资源 资源资源资源 资源资源 资源资源](/blog/building-a-lawyer-gpt-for-your-blog-part5)
- [所有部分](/blog/building-a-lawyer-gpt-for-your-blog-part6)
- [第1部分:导言和建筑](/blog/building-a-lawyer-gpt-for-your-blog-part7)
- [第2部分: GPU 设置和 CUDA](/blog/building-a-lawyer-gpt-for-your-blog-part8)

### 第3部分:嵌入和矢量数据库

- [第4部分:摄入管](/blog/building-a-lawyer-gpt-for-your-blog-part1)
- [第5部分:Windows客户](/blog/building-a-lawyer-gpt-for-your-blog-part2)
- [第6部分:地方LLM整合](/blog/building-a-lawyer-gpt-for-your-blog-part3)
- [第7部分:内容生成](/blog/building-a-lawyer-gpt-for-your-blog-part4)
- [第8部分:生产部署](/blog/building-a-lawyer-gpt-for-your-blog-part5)
- [系列完成!](/blog/building-a-lawyer-gpt-for-your-blog-part6)
- [第1部分:导言和建筑](/blog/building-a-lawyer-gpt-for-your-blog-part7)
- **第2部分: GPU 设置和 CUDA**第3部分:嵌入和矢量数据库

### 第4部分:摄入管

- [第5部分:Windows客户](https://github.com/SciSharp/LLamaSharp)
- [第6部分:地方LLM整合](https://qdrant.tech/)
- [第7部分:内容生成](https://avaloniaui.net/)
- [第8部分:生产部署](https://onnxruntime.ai/)
- [(本员额)](https://wixtoolset.org/)

Happy writing with your new AI assistant! 🚀