AI Chat with Custom Data

A modern .NET 10 Blazor Web Application that enables intelligent conversations with your custom documents using semantic search and vector embeddings.

🚀 Quick Start

Prerequisites

• .NET 10 SDK
• Visual Studio 2022 or VS Code
• GitHub Models API Token (free tier available)

Setup

1. Add your documents

   • Place PDF files in wwwroot/Data/ directory
   • Example PDFs included: Example_Emergency_Survival_Kit.pdf, Example_GPS_Watch.pdf

2. Configure GitHub Models Token

   • Right-click project in Solution Explorer → "Manage User Secrets"
   • Add: { "GitHubModels:Token": "YOUR-GITHUB-TOKEN" }
   • Get token here

3. Run the application

   dotnet run

   • Open https://localhost:5001 and start chatting!

---

📊 How It Works

Data Flow:

1. User embeds message to vector representation
2. Semantic search finds top 5 relevant document chunks in SQLite
3. AI receives message + context chunks
4. Response with citations displayed to user

Ingestion Pipeline:

• PDFs loaded from /wwwroot/Data/
• Split into semantic chunks
• Embedded to 1536-dimensional vectors
• Stored in SQLite vector database

---

🏗️ Architecture

Component	Purpose
Chat.razor	Main chat interface
ChatHeader/Input/MessageList.razor	UI components
SemanticSearch.cs	Vector similarity search
DataIngestor.cs	Document ingestion & embedding
PDFDirectorySource.cs	PDF loader
IngestedChunk/Document.cs	Data models

---

⚙️ Configuration

appsettings.json

{
  "AiModels": {
    "ChatModel": "gpt-4o-mini",
    "EmbeddingModel": "text-embedding-3-small",
    "Endpoint": "https://models.inference.ai.azure.com"
  }
}

Supported Models

Model	Speed	Cost	Best For
gpt-4o-mini	⚡ Fast	$ Low	Current (default)
gpt-4o	Fast	$$ Medium	General use
o1	Slow	$$$ High	Complex reasoning
o1-mini	Medium	$$ Medium	Balanced reasoning
llama-2-70b	Fast	$ Free	Budget option

To switch models: Edit AiModels:ChatModel in appsettings.json

---

🎯 Key Features

✅ Semantic Search - Find relevant info using meaning, not keywords
✅ Auto Tool Calling - AI decides when to search documents
✅ Citations - Exact quotes with document sources
✅ Copy & Listen - One-click copy and text-to-speech
✅ Edit Questions - Modify and re-ask instantly
✅ Multi-turn Chat - Full conversation context
✅ Configurable Models - Easy model switching
✅ Local Vector DB - Fast, private search

---

🔒 Security

✅ API keys in User Secrets (not in source control)
✅ Semantic search limits AI to ingested documents
✅ System prompt prevents out-of-context responses
✅ Vector DB local (no cloud content storage)

⚠️ Only ingest trusted documents to prevent prompt injection

---

📁 Project Structure

AIChatBotS/
├── Components/Pages/Chat/
│   ├── Chat.razor, ChatHeader.razor, ChatInput.razor
│   └── ChatMessageList.razor, ChatMessageItem.razor
├── Services/
│   ├── SemanticSearch.cs
│   └── Ingestion/
│       ├── DataIngestor.cs, PDFDirectorySource.cs
│       └── IngestedChunk.cs, IngestedDocument.cs
├── wwwroot/Data/          (Add PDFs here)
├── Program.cs
├── appsettings.json
└── README.md

---

🔄 Extending

Add New Document Source

// Implement IIngestionSource interface
// Add to Program.cs:
await DataIngestor.IngestDataAsync(app.Services, new YourCustomSource());

Customize AI Behavior

• System Prompt: Edit Chat.razor SystemPrompt constant
• Button Actions: Edit ChatMessageItem.razor
• Embedding Model: Change EmbeddingModel in appsettings.json

---

📚 Resources

• GitHub Models Docs
• Microsoft.Extensions.AI
• Blazor Documentation
• Vector Embeddings Guide

---

📞 Troubleshooting

1. ❌ Chat not working → Verify GitHub token in User Secrets
2. ❌ No results → Ensure PDFs are in wwwroot/Data/
3. ❌ Model errors → Check model supports tool calling
4. ❌ Slow responses → Consider gpt-4o-mini for speed

---

Built with ❤️ using .NET 10 and Blazor