mirror of https://github.com/abpframework/abp.git
maliming
10 months ago
83 changed files with 1780 additions and 236 deletions
@ -0,0 +1,158 @@ |
|||
# My First Look and Experience with Google AntiGravity |
|||
|
|||
## Is Google AntiGravity Going to Replace Your Main Code Editor? |
|||
|
|||
Today, I tried the new code-editor AntiGravity by Google. *"It's beyond a code-editor*" by Google 🙄 |
|||
When I first launch it, I see the UI is almost same as Cursor. They're both based on Visual Studio Code. |
|||
That's why it was not hard to find what I'm looking for. |
|||
|
|||
First of all, the main difference as I see from the Cursor is; when I type a prompt in the agent section **AntiGravity first creates a Task List** (like a road-map) and whenever it finishes a task, it checks the corresponding task. Actually Cursor has a similar functionality but AntiGravity took it one step further. |
|||
|
|||
Second thing which was good to me; AntiGravity uses [Nano Banana 🍌](https://gemini.google/tr/overview/image-generation/). This is Google's AI image generation model... Why it's important because when you create an app, you don't need to search for graphics, deal with image licenses. **AntiGravity generates images automatically and no license is required!** |
|||
|
|||
Third exciting feature for me; **AntiGravity is integrated with Google Chrome and can communicate with the running website**. When I first run my web project, it installed a browser extension which can see and interact with my website. It can see the results, click somewhere else on the page, scroll, fill up the forms, amazing 😵 |
|||
|
|||
Another feature I loved is that **you can enter a new prompt even while AntiGravity is still generating a response** 🧐. It instantly prioritizes the latest input and adjusts the ongoing process if needed. But in Cursor, if you add a prompt before the cursor finishes, it simply queues it and runs it later 😔. |
|||
|
|||
And lastly, **AntiGravity is working very good with Gemini 3**. |
|||
|
|||
Well, everything was not so perfect 😥 When I tried AntiGravity, couple of times it stucked AI generation and Agent stopped. I faced errors like this 👇 |
|||
|
|||
 |
|||
|
|||
|
|||
|
|||
## Debugging .NET Projects via AntiGravity |
|||
|
|||
⚠ There's a crucial development issue with AntiGravity (and also for Cursor, Windsurf etc...) 🤕 you **cannot debug your .NET application with AntiGravity 🥺.** *This is Microsoft's policy!* Microsoft doesn't allow debugging for 3rd party IDEs and shows the below error... That's why I cannot say it's a downside of AntiGravity. You need to use Microsft's original VS Code, Visual Studio or Rider for debugging. But wait a while there's a workaround for this, I'll let you know in the next section. |
|||
|
|||
|
|||
|
|||
 |
|||
|
|||
### What does this error mean? |
|||
|
|||
AntiGravity, Cursor, Windsurf etc... are using Visual Studio Code and the C# extension for VS Code includes the Microsoft .NET Core Debugger "*vsdbg*". |
|||
VS Code is open-source but "*vsdbg*" is not open-source! It's working only with Visual Studio Code, Visual Studio and Visual Studio for Mac. This is clearly stated at [Microsoft's this link](https://github.com/dotnet/vscode-csharp/blob/main/docs/debugger/Microsoft-.NET-Core-Debugger-licensing-and-Microsoft-Visual-Studio-Code.md). |
|||
|
|||
### Ok! How to resolve debugging issue with AntiGravity? and Cursor and Windsurf... |
|||
|
|||
There's a free C# debugger extension for Visual Studio Code based IDEs that supports AntiGravity, Cursor and Windsurf. The extension name is **C#**. |
|||
You can download this free C# debugger extension at 👉 [open-vsx.org/extension/muhammad-sammy/csharp/](https://open-vsx.org/extension/muhammad-sammy/csharp/). |
|||
For AntiGravity open Extension window (*Ctrl + Shift + X*) and search for `C#`, there you'll see this extension. |
|||
|
|||
 |
|||
|
|||
After installing, I restarted AntiGravity and now I can see the red circle which allows me to add breakpoint on C# code. |
|||
|
|||
 |
|||
|
|||
### Another Extension For Debugging .NET Apps on VS Code |
|||
|
|||
Recently I heard about DotRush extension from the folks. As they say DotRush works slightly faster and support Razor pages (.cshtml files). |
|||
Here's the link for DotRush https://github.com/JaneySprings/DotRush |
|||
|
|||
### Finding Website Running Port |
|||
|
|||
When you run the web project via C# debugger extension, normally it's not using the `launch.json` therefore the website port is not the one when you start from Visual Studio / Rider... So what's my website's port which I just run now? Normally for ASP.NET Core **the default port is 5000**. You can try navigating to http://localhost:5000/. |
|||
Alternatively you can write the below code in `Program.cs` which prints the full address of your website in the logs. |
|||
If you do the steps which I showed you, you can debug your C# application via AntiGravity and other VS Code derivatives. |
|||
|
|||
 |
|||
|
|||
## How Much is AntiGravity? 💲 |
|||
|
|||
Currently there's only individual plan is available for personal accounts and that's free 👏! The contents of Team and Enterprise plans and prices are not announced yet. But **Gemini 3 is not free**! I used it with my company's Google Workspace account which we normally pay for Gemini. |
|||
|
|||
 |
|||
|
|||
## More About AntiGravity |
|||
|
|||
There have been many AI assisted IDEs like [Windsurf](https://windsurf.com/), [Cursor](https://cursor.com/), [Zed](https://zed.dev/), [Replit](https://replit.com/) and [Fleet](https://www.jetbrains.com/fleet/). But this time it's different, this is backed by Google. |
|||
As you see from the below image AntiGravity, uses a standard grid layout as others based on VS Code editor. |
|||
It's very similar to Cursor, Visual Studio, Rider. |
|||
|
|||
 |
|||
|
|||
## Supported LLMs 🧠 |
|||
|
|||
Antigravity offers the below models which supports reasoning: Gemini 3 Pro, Claude Sonnet 4.5, GPT-OSS |
|||
|
|||
 |
|||
|
|||
Antigravity uses other models for supportive tasks in the background: |
|||
|
|||
- **Nano banana**: This is used to generate images. |
|||
- **Gemini 2.5 Pro UI Checkpoint**: It's for the browser subagent to trigger browser action such as clicking, scrolling, or filling in input. |
|||
- **Gemini 2.5 Flash**: For checkpointing and context summarization, this is used. |
|||
- **Gemini 2.5 Flash Lite**: And when it's need to make a semantic search in your code-base, this is used. |
|||
|
|||
## AntiGravity Can See Your Website |
|||
|
|||
This makes a big difference from traditional IDEs. AntiGravity's browser agent is taking screenshots of your pages when it needs to check. This is achieved by a Chrome Extension as a tool to the agent, and you can also prompt the agent to take a screenshot of a page. It can iterate on website designs and implementations, it can perform UI Testing, it can monitor dashboards, it can automate routine tasks like rerunning CI. |
|||
This is the link for the extension 👉 [chromewebstore.google.com/detail/antigravity-browser-exten/eeijfnjmjelapkebgockoeaadonbchdd](https://chromewebstore.google.com/detail/antigravity-browser-exten/eeijfnjmjelapkebgockoeaadonbchdd). AntiGravity will install this extension automatically on the first run. |
|||
|
|||
 |
|||
|
|||
 |
|||
|
|||
## MCP Integration |
|||
|
|||
### When Do We Need MCP in a Code Editor? |
|||
|
|||
Simply if we want to connect to a 3rd party service to complete our task we need MCP. So AntiGravity can connect to your DB and write proper SQL queries or it can pull in recent build logs from Netlify or Heroku. Also you can ask AntiGravity to to connect GitHub for finding the best authentication pattern. |
|||
|
|||
### AntiGravity Supports These MCP Servers |
|||
|
|||
Airweave, AlloyDB for PostgreSQL, Atlassian, BigQuery, Cloud SQL for PostgreSQL, Cloud SQL for MySQL, Cloud SQL for SQL Server, Dart, Dataplex, Figma Dev Mode MCP, Firebase, GitHub, Harness, Heroku, Linear, Locofy, Looker, MCP Toolbox for Databases, MongoDB, Neon, Netlify, Notion, PayPal, Perplexity Ask, Pinecone, Prisma, Redis, Sequential Thinking, SonarQube, Spanner, Stripe and Supabase. |
|||
|
|||
 |
|||
|
|||
## Agent Settings ⚙️ |
|||
|
|||
The major settings of Agent are: |
|||
|
|||
- **Agent Auto Fix Lints**: I enabled this setting because I want the Agent automatically fixes its own mistakes for invalid syntax, bad formatting, unused variables, unreachable code or following coding standards... It makes extra tool calls that's why little bit expensive 🥴. |
|||
- **Auto Execution**: Sometimes Agent tries to build application or writing test code and running it, in these cases it executes command. I choose "Turbo" 🤜 With this option, Agent always runs the terminal command and controls my browser. |
|||
- **Review Policy**: How much control you are giving to agent 🙎. I choose "Always Proceed" 👌 because I mostly trust AI 😀. The Agent will never ask for review. |
|||
|
|||
 |
|||
|
|||
## Differences Between Cursor and AntiGravity |
|||
|
|||
While Cursor was the champion of AI code editors, **Antigravity brings a different philosophy**. |
|||
|
|||
### 1. "Agent-First 🤖" vs "You-First 🤠" |
|||
|
|||
- **Cursor:** It acts like an assistant; it predicts your next move, auto-completes your thoughts, and helps you refactor while you type. You are still the driver; Cursor just drives the car at 200 km/h. |
|||
- **Antigravity:** Antigravity is built to let you manage coding tasks. It is "Agent-First." You don't just type code; you assign tasks to autonomous agents (e.g., "Fix the bug in the login flow and verify it in the browser"). It behaves more like a junior developer that you supervise. |
|||
|
|||
### 2. The Interface |
|||
|
|||
- **Cursor:** Looks and feels exactly like **VS Code**. If you know VS Code, you know Cursor. |
|||
|
|||
- **Antigravity:** Introduces 2 major layouts: |
|||
- **Editor View:** Similar to a standard IDE |
|||
- **Manager View:** A dashboard where you see multiple "Agents" working in parallel. You can watch them plan, execute, and test tasks asynchronously. |
|||
|
|||
### 3. Verification & Trust |
|||
|
|||
- **Cursor:** You verify by reading the code diffs it suggests. |
|||
- **Antigravity:** Introduces **Artifacts**... Since the agents work autonomously, they generate proof-of-work documents, screenshots of the app running, browser logs and execution plans. So you can verify what they did without necessarily reading every line of code immediately. |
|||
|
|||
### 4. Capabilities |
|||
|
|||
- **Cursor:** Best-in-class **Autocomplete** ("Tab" feature) and **Composer** (multi-file editing). It excels at "Vibe Coding". It's getting into a flow state where the AI writes the boilerplate and you direct the logic. |
|||
- **Antigravity:** Is good at **Autonomous Execution**. It has a built-in browser and terminal that the *Agent* controls. The Agent can write code, run the server, open the browser, see the error, and fix it 😎 |
|||
|
|||
### 5. AI Models (Brains 🧠) |
|||
|
|||
- **Cursor:** Model Agnostic. You can switch between **Claude 3.5 Sonnet** *-mostly the community uses this-*, GPT-4o, and others. |
|||
- **Antigravity:** Built deeply around **Gemini 3 Pro**. It leverages Gemini's massive context window (1M+ tokens) to understand huge mono repos without needing as much "RAG" as Cursor. |
|||
|
|||
|
|||
|
|||
## Try It Yourself Now 🤝 |
|||
|
|||
If you are ready to experience the new AI code editor by Google, download and use 👇 |
|||
[**Launch Google AntiGravity**](https://antigravity.google/) |
|||
|
After Width: | Height: | Size: 70 KiB |
|
After Width: | Height: | Size: 183 KiB |
|
After Width: | Height: | Size: 62 KiB |
|
After Width: | Height: | Size: 1.1 MiB |
|
After Width: | Height: | Size: 90 KiB |
|
After Width: | Height: | Size: 13 KiB |
|
After Width: | Height: | Size: 102 KiB |
|
After Width: | Height: | Size: 293 KiB |
|
After Width: | Height: | Size: 76 KiB |
|
After Width: | Height: | Size: 46 KiB |
|
After Width: | Height: | Size: 13 KiB |
|
After Width: | Height: | Size: 13 KiB |
|
After Width: | Height: | Size: 275 KiB |
|
After Width: | Height: | Size: 49 KiB |
|
After Width: | Height: | Size: 154 KiB |
|
After Width: | Height: | Size: 5.4 KiB |
|
After Width: | Height: | Size: 6.3 KiB |
|
After Width: | Height: | Size: 6.0 KiB |
|
After Width: | Height: | Size: 4.8 KiB |
|
After Width: | Height: | Size: 7.2 KiB |
|
After Width: | Height: | Size: 3.1 KiB |
|
After Width: | Height: | Size: 6.6 KiB |
@ -0,0 +1,414 @@ |
|||
# Building Production-Ready LLM Applications with .NET: A Practical Guide |
|||
|
|||
Large Language Models (LLMs) have evolved rapidly, and integrating them into production .NET applications requires staying current with the latest approaches. In this article, I'll share practical tips and patterns I've learned while building LLM-powered systems, covering everything from API changes in GPT-5 to implementing efficient RAG (Retrieval Augmented Generation) architectures. |
|||
|
|||
Whether you're building a chatbot, a knowledge base assistant, or integrating AI into your enterprise applications, these production-tested insights will help you avoid common pitfalls and build more reliable systems. |
|||
|
|||
## The Temperature Paradigm Shift: GPT-5 Changes Everything |
|||
|
|||
If you've been working with GPT-4 or earlier models, you're familiar with the `temperature` and `top_p` parameters for controlling response randomness. **Here's the critical update**: GPT-5 no longer supports these parameters! |
|||
|
|||
### The Old Way (GPT-4) |
|||
```csharp |
|||
var chatRequest = new ChatOptions |
|||
{ |
|||
Temperature = 0.7, // ✅ Worked with GPT-4 |
|||
TopP = 0.9 // ✅ Worked with GPT-4 |
|||
}; |
|||
``` |
|||
|
|||
### The New Way (GPT-5) |
|||
```csharp |
|||
var chatRequest = new ChatOptions |
|||
{ |
|||
RawRepresentationFactory = (client => new ChatCompletionOptions() |
|||
{ |
|||
#pragma warning disable OPENAI001 |
|||
ReasoningEffortLevel = "minimal", |
|||
#pragma warning restore OPENAI001 |
|||
}) |
|||
}; |
|||
``` |
|||
|
|||
**Why the change?** GPT-5 incorporates an internal reasoning and verification process. Instead of controlling randomness, you now specify how much computational effort the model should invest in reasoning through the problem. |
|||
|
|||
 |
|||
|
|||
### Choosing the Right Reasoning Level |
|||
|
|||
- **Low**: Quick responses for simple queries (e.g., "What's the capital of France?") |
|||
- **Medium**: Balanced approach for most use cases |
|||
- **High**: Complex reasoning tasks (e.g., code generation, multi-step problem solving) |
|||
|
|||
> **Pro Tip**: Reasoning tokens are included in your API costs. Use "High" only when necessary to optimize your budget. |
|||
|
|||
## System Prompts: The "Lost in the Middle" Problem |
|||
|
|||
Here's a critical insight that can save you hours of debugging: **Important rules must be repeated at the END of your prompt!** |
|||
|
|||
### ❌ What Doesn't Work |
|||
``` |
|||
You are a helpful assistant. |
|||
RULE: Never share passwords or sensitive information. |
|||
|
|||
[User Input] |
|||
``` |
|||
|
|||
### ✅ What Actually Works |
|||
``` |
|||
You are a helpful assistant. |
|||
RULE: Never share passwords or sensitive information. |
|||
|
|||
[User Input] |
|||
|
|||
⚠️ REMINDER: Apply the rules above strictly, ESPECIALLY regarding passwords. |
|||
``` |
|||
|
|||
**Why?** LLMs suffer from the "Lost in the Middle" phenomenon—they pay more attention to the beginning and end of the context window. Critical instructions buried in the middle are often ignored. |
|||
|
|||
## RAG Architecture: The Parent-Child Pattern |
|||
|
|||
Retrieval Augmented Generation (RAG) is essential for grounding LLM responses in your own data. The most effective pattern I've found is the **Parent-Child approach**. |
|||
|
|||
 |
|||
|
|||
### How It Works |
|||
|
|||
1. **Split documents into hierarchies**: |
|||
- **Parent chunks**: Large sections (1000-2000 tokens) for context |
|||
- **Child chunks**: Small segments (200-500 tokens) for precise retrieval |
|||
|
|||
2. **Store both in vector database** with references |
|||
|
|||
3. **Query flow**: |
|||
- Search using child chunks (higher precision) |
|||
- Return parent chunks to LLM (richer context) |
|||
|
|||
### The Overlap Strategy |
|||
|
|||
Always use overlapping chunks to prevent information loss at boundaries! |
|||
|
|||
``` |
|||
Chunk 1: Token 0-500 |
|||
Chunk 2: Token 400-900 ← 100 token overlap |
|||
Chunk 3: Token 800-1300 ← 100 token overlap |
|||
``` |
|||
|
|||
**Standard recommendation**: 10-20% overlap (for 500 tokens, use 50-100 token overlap) |
|||
|
|||
### Implementation with Semantic Kernel |
|||
|
|||
```csharp |
|||
using Microsoft.SemanticKernel.Text; |
|||
|
|||
var chunks = TextChunker.SplitPlainTextParagraphs( |
|||
documentText, |
|||
maxTokensPerParagraph: 500, |
|||
overlapTokens: 50 |
|||
); |
|||
|
|||
foreach (var chunk in chunks) |
|||
{ |
|||
var embedding = await embeddingService.GenerateEmbeddingAsync(chunk); |
|||
await vectorDb.StoreAsync(chunk, embedding); |
|||
} |
|||
``` |
|||
|
|||
## PostgreSQL + pgvector: The Pragmatic Choice |
|||
|
|||
For .NET developers, choosing a vector database can be overwhelming. After evaluating multiple options, **PostgreSQL with pgvector** is the most practical choice for most scenarios. |
|||
|
|||
 |
|||
|
|||
### Why pgvector? |
|||
|
|||
✅ **Use existing SQL knowledge** - No new query language to learn |
|||
✅ **EF Core integration** - Works with your existing data access layer |
|||
✅ **JOIN with metadata** - Combine vector search with traditional queries |
|||
✅ **WHERE clause filtering** - Filter by tenant, user, date, etc. |
|||
✅ **ACID compliance** - Transaction support for data consistency |
|||
✅ **No separate infrastructure** - One database for everything |
|||
|
|||
### Setting Up pgvector with EF Core |
|||
|
|||
First, install the NuGet package: |
|||
|
|||
```bash |
|||
dotnet add package Pgvector.EntityFrameworkCore |
|||
``` |
|||
|
|||
Define your entity: |
|||
|
|||
```csharp |
|||
using Pgvector; |
|||
using Pgvector.EntityFrameworkCore; |
|||
|
|||
public class DocumentChunk |
|||
{ |
|||
public Guid Id { get; set; } |
|||
public string Content { get; set; } |
|||
public Vector Embedding { get; set; } // 👈 pgvector type |
|||
public Guid ParentChunkId { get; set; } |
|||
public DateTime CreatedAt { get; set; } |
|||
} |
|||
``` |
|||
|
|||
Configure in DbContext: |
|||
|
|||
```csharp |
|||
protected override void OnModelCreating(ModelBuilder builder) |
|||
{ |
|||
builder.HasPostgresExtension("vector"); |
|||
|
|||
builder.Entity<DocumentChunk>() |
|||
.Property(e => e.Embedding) |
|||
.HasColumnType("vector(1536)"); // 👈 OpenAI embedding dimension |
|||
|
|||
builder.Entity<DocumentChunk>() |
|||
.HasIndex(e => e.Embedding) |
|||
.HasMethod("hnsw") // 👈 Fast approximate search |
|||
.HasOperators("vector_cosine_ops"); |
|||
} |
|||
``` |
|||
|
|||
### Performing Vector Search |
|||
|
|||
```csharp |
|||
using Pgvector.EntityFrameworkCore; |
|||
|
|||
public async Task<List<DocumentChunk>> SearchAsync(string query) |
|||
{ |
|||
// 1. Convert query to embedding |
|||
var queryVector = await _embeddingService.GetEmbeddingAsync(query); |
|||
|
|||
// 2. Search |
|||
return await _context.DocumentChunks |
|||
.OrderBy(c => c.Embedding.L2Distance(queryVector)) // 👈 Lower is better |
|||
.Take(5) |
|||
.ToListAsync(); |
|||
} |
|||
``` |
|||
|
|||
**Source**: [Pgvector.NET on GitHub](https://github.com/pgvector/pgvector-dotnet?tab=readme-ov-file#entity-framework-core) |
|||
|
|||
## Smart Tool Usage: Make RAG a Tool, Not a Tax |
|||
|
|||
A common mistake is calling RAG on every single user message. This wastes tokens and money. Instead, **make RAG a tool** and let the LLM decide when to use it. |
|||
|
|||
### ❌ Expensive Approach |
|||
```csharp |
|||
// Always call RAG, even for "Hello" |
|||
var context = await PerformRAG(userMessage); |
|||
var response = await chatClient.CompleteAsync($"{context}\n\n{userMessage}"); |
|||
``` |
|||
|
|||
### ✅ Smart Approach |
|||
```csharp |
|||
[KernelFunction] |
|||
[Description("Search the company knowledge base for information")] |
|||
public async Task<string> SearchKnowledgeBase( |
|||
[Description("The search query")] string query) |
|||
{ |
|||
var results = await _vectorDb.SearchAsync(query); |
|||
return string.Join("\n---\n", results.Select(r => r.Content)); |
|||
} |
|||
``` |
|||
|
|||
The LLM will call `SearchKnowledgeBase` only when needed: |
|||
- "Hello" → No tool call |
|||
- "What was our 2024 revenue?" → Calls tool |
|||
- "Tell me a joke" → No tool call |
|||
|
|||
## Multilingual RAG: Query Translation Strategy |
|||
|
|||
When your documents are in one language (e.g., English) but users query in another (e.g., Turkish), you need a translation strategy. |
|||
|
|||
 |
|||
|
|||
### Solution Options |
|||
|
|||
**Option 1**: Use an LLM that automatically calls tools in English |
|||
- Many modern LLMs can do this if properly instructed |
|||
|
|||
**Option 2**: Tool chain approach |
|||
```csharp |
|||
[KernelFunction] |
|||
[Description("Translate text to English")] |
|||
public async Task<string> TranslateToEnglish(string text) |
|||
{ |
|||
// Translation logic |
|||
} |
|||
|
|||
[KernelFunction] |
|||
[Description("Search knowledge base (English only)")] |
|||
public async Task<string> SearchKnowledgeBase(string englishQuery) |
|||
{ |
|||
// Search logic |
|||
} |
|||
``` |
|||
|
|||
The LLM will: |
|||
1. Call `TranslateToEnglish("2024 geliri nedir?")` |
|||
2. Get "What was 2024 revenue?" |
|||
3. Call `SearchKnowledgeBase("What was 2024 revenue?")` |
|||
4. Return results and respond in Turkish |
|||
|
|||
## Model Context Protocol (MCP): Beyond In-Process Tools |
|||
|
|||
Microsoft and Anthropic recently released official C# SDKs for the Model Context Protocol (MCP). This is a game-changer for tool reusability. |
|||
|
|||
 |
|||
|
|||
### MCP vs. Semantic Kernel Plugins |
|||
|
|||
| Feature | SK Plugins | MCP Servers | |
|||
|---------|-----------|-------------| |
|||
| **Process** | In-process | Out-of-process (stdio/http) | |
|||
| **Reusability** | Application-specific | Cross-application | |
|||
| **Examples** | Used within your app | VS Code Copilot, Claude Desktop | |
|||
|
|||
### Creating an MCP Server |
|||
|
|||
```csharp |
|||
using Microsoft.Extensions.Hosting; |
|||
using ModelContextProtocol.Extensions.Hosting; |
|||
|
|||
var builder = Host.CreateEmptyApplicationBuilder(settings: null); |
|||
|
|||
builder.Services.AddMcpServer() |
|||
.WithStdioServerTransport() |
|||
.WithToolsFromAssembly(); |
|||
|
|||
await builder.Build().RunAsync(); |
|||
``` |
|||
|
|||
Define your tools: |
|||
|
|||
```csharp |
|||
[McpServerToolType] |
|||
public static class FileSystemTools |
|||
{ |
|||
[McpServerTool, Description("Read a file from the file system")] |
|||
public static async Task<string> ReadFile(string path) |
|||
{ |
|||
// ⚠️ SECURITY: Always validate paths! |
|||
if (!IsPathSafe(path)) |
|||
throw new SecurityException("Invalid path"); |
|||
|
|||
return await File.ReadAllTextAsync(path); |
|||
} |
|||
|
|||
private static bool IsPathSafe(string path) |
|||
{ |
|||
// Implement path traversal prevention |
|||
var fullPath = Path.GetFullPath(path); |
|||
return fullPath.StartsWith(AllowedDirectory); |
|||
} |
|||
} |
|||
``` |
|||
|
|||
Your MCP server can now be used by VS Code Copilot, Claude Desktop, or any other MCP client! |
|||
|
|||
## Chat History Management: Truncation + RAG Hybrid |
|||
|
|||
For long conversations, storing all history in the context window becomes impractical. Here's the pattern that works: |
|||
|
|||
 |
|||
|
|||
### ❌ Lossy Approach |
|||
``` |
|||
First 50 messages → Summarize with LLM → Single summary message |
|||
``` |
|||
**Problem**: Detail loss (fidelity loss) |
|||
|
|||
### ✅ Hybrid Approach |
|||
1. **Recent messages** (last 5-10): Keep in prompt for immediate context |
|||
2. **Older messages**: Store in vector database as a tool |
|||
|
|||
```csharp |
|||
[KernelFunction] |
|||
[Description("Search conversation history for past discussions")] |
|||
public async Task<string> SearchChatHistory( |
|||
[Description("What to search for")] string query) |
|||
{ |
|||
var relevantMessages = await _vectorDb.SearchAsync(query); |
|||
return string.Join("\n", relevantMessages.Select(m => |
|||
$"[{m.Timestamp}] {m.Role}: {m.Content}")); |
|||
} |
|||
``` |
|||
|
|||
The LLM retrieves only relevant past context when needed, avoiding summary-induced information loss. |
|||
|
|||
## RAG vs. Fine-Tuning: Choose Wisely |
|||
|
|||
A common misconception is using fine-tuning for knowledge injection. Here's when to use each: |
|||
|
|||
| Purpose | RAG | Fine-Tuning | |
|||
|---------|-----|-------------| |
|||
| **Goal** | Memory (provide facts) | Behavior (teach style) | |
|||
| **Updates** | Dynamic (add docs anytime) | Static (requires retraining) | |
|||
| **Cost** | Low dev, higher inference | High dev, lower inference | |
|||
| **Hallucination** | Reduces | Doesn't reduce | |
|||
| **Use Case** | Company docs, FAQs | Brand voice, specific format | |
|||
|
|||
**Common mistake**: "Let's fine-tune on our company documents" ❌ |
|||
**Better approach**: Use RAG! ✅ |
|||
|
|||
Fine-tuning is for teaching the model *how* to respond, not *what* to know. |
|||
|
|||
**Source**: [Oracle - RAG vs Fine-Tuning](https://www.oracle.com/artificial-intelligence/generative-ai/retrieval-augmented-generation-rag/rag-fine-tuning/) |
|||
|
|||
## Bonus: Why SVG is Superior for LLM-Generated Images |
|||
|
|||
When using LLMs to generate diagrams and visualizations, always request SVG format instead of PNG or JPG. |
|||
|
|||
### Why SVG? |
|||
|
|||
✅ **Text-based** → LLMs produce better results |
|||
✅ **Lower cost** → Fewer tokens than base64-encoded images |
|||
✅ **Editable** → Easy to modify after generation |
|||
✅ **Scalable** → Perfect quality at any size |
|||
✅ **Version control friendly** → Works great in Git |
|||
|
|||
### Example Prompt |
|||
|
|||
``` |
|||
Create an architecture diagram showing PostgreSQL with pgvector integration. |
|||
Format: SVG, 800x400 pixels. Show: .NET Application → EF Core → PostgreSQL → Vector Search. |
|||
Use arrows to connect stages. Color scheme: Blue tones. |
|||
``` |
|||
|
|||
 |
|||
|
|||
All diagrams in this article were generated as SVG, resulting in excellent quality and lower token costs! |
|||
|
|||
> **Pro Tip**: If you don't need photographs or complex renders, always choose SVG. |
|||
|
|||
## Architecture Roadmap: Putting It All Together |
|||
|
|||
Here's the recommended stack for building production LLM applications with .NET: |
|||
|
|||
1. **Orchestration**: Microsoft.Extensions.AI + Semantic Kernel (when needed) |
|||
2. **Vector Database**: PostgreSQL + Pgvector.EntityFrameworkCore |
|||
3. **RAG Pattern**: Parent-Child chunks with 10-20% overlap |
|||
4. **Tools**: MCP servers for reusability |
|||
5. **Reasoning**: ReasoningEffortLevel instead of temperature |
|||
6. **Prompting**: Critical rules at the end |
|||
7. **Cost Optimization**: Make RAG a tool, not automatic |
|||
|
|||
## Key Takeaways |
|||
|
|||
Let me summarize the most important production tips: |
|||
|
|||
1. **Temperature is gone** → Use `ReasoningEffortLevel` with GPT-5 |
|||
2. **Rules at the end** → Combat "Lost in the Middle" |
|||
3. **RAG as a tool** → Reduce costs significantly |
|||
4. **Parent-Child pattern** → Search small, respond with large |
|||
5. **Always use overlap** → 10-20% is the standard |
|||
6. **pgvector for most cases** → Unless you have billions of vectors |
|||
7. **MCP for reusability** → One codebase, works everywhere |
|||
8. **SVG for diagrams** → Better results, lower cost |
|||
9. **Hybrid chat history** → Recent in prompt, old in vector DB |
|||
10. **RAG > Fine-tuning** → For knowledge, not behavior |
|||
|
|||
Happy coding! 🚀 |
|||
@ -0,0 +1 @@ |
|||
Learn how to build production-ready LLM applications with .NET. This comprehensive guide covers GPT-5 API changes, advanced RAG architectures with parent-child patterns, PostgreSQL pgvector integration, smart tool usage strategies, multilingual query handling, Model Context Protocol (MCP) for cross-application tool reusability, and chat history management techniques for enterprise applications. |
|||
|
Before Width: | Height: | Size: 407 KiB After Width: | Height: | Size: 559 KiB |
|
Before Width: | Height: | Size: 10 KiB After Width: | Height: | Size: 25 KiB |
|
Before Width: | Height: | Size: 41 KiB After Width: | Height: | Size: 41 KiB |
@ -0,0 +1,38 @@ |
|||
using System.Linq; |
|||
using System.Threading.Tasks; |
|||
using Volo.Abp.Authorization.Permissions; |
|||
|
|||
namespace Volo.Abp.Authorization.TestServices; |
|||
|
|||
public class TestProhibitedPermissionValueProvider1 : PermissionValueProvider |
|||
{ |
|||
public TestProhibitedPermissionValueProvider1(IPermissionStore permissionStore) : base(permissionStore) |
|||
{ |
|||
} |
|||
|
|||
public override string Name => "TestProhibitedPermissionValueProvider1"; |
|||
|
|||
public override Task<PermissionGrantResult> CheckAsync(PermissionValueCheckContext context) |
|||
{ |
|||
var result = PermissionGrantResult.Undefined; |
|||
if (context.Permission.Name == "MyPermission8" || context.Permission.Name == "MyPermission9") |
|||
{ |
|||
result = PermissionGrantResult.Granted; |
|||
} |
|||
|
|||
return Task.FromResult(result); |
|||
} |
|||
|
|||
public override Task<MultiplePermissionGrantResult> CheckAsync(PermissionValuesCheckContext context) |
|||
{ |
|||
var result = new MultiplePermissionGrantResult(); |
|||
foreach (var name in context.Permissions.Select(x => x.Name)) |
|||
{ |
|||
result.Result.Add(name, name == "MyPermission8" || name == "MyPermission9" |
|||
? PermissionGrantResult.Granted |
|||
: PermissionGrantResult.Undefined); |
|||
} |
|||
|
|||
return Task.FromResult(result); |
|||
} |
|||
} |
|||
@ -0,0 +1,38 @@ |
|||
using System.Linq; |
|||
using System.Threading.Tasks; |
|||
using Volo.Abp.Authorization.Permissions; |
|||
|
|||
namespace Volo.Abp.Authorization.TestServices; |
|||
|
|||
public class TestProhibitedPermissionValueProvider2 : PermissionValueProvider |
|||
{ |
|||
public TestProhibitedPermissionValueProvider2(IPermissionStore permissionStore) : base(permissionStore) |
|||
{ |
|||
} |
|||
|
|||
public override string Name => "TestProhibitedPermissionValueProvider2"; |
|||
|
|||
public override Task<PermissionGrantResult> CheckAsync(PermissionValueCheckContext context) |
|||
{ |
|||
var result = PermissionGrantResult.Undefined; |
|||
if (context.Permission.Name == "MyPermission8" || context.Permission.Name == "MyPermission9") |
|||
{ |
|||
result = PermissionGrantResult.Prohibited; |
|||
} |
|||
|
|||
return Task.FromResult(result); |
|||
} |
|||
|
|||
public override Task<MultiplePermissionGrantResult> CheckAsync(PermissionValuesCheckContext context) |
|||
{ |
|||
var result = new MultiplePermissionGrantResult(); |
|||
foreach (var name in context.Permissions.Select(x => x.Name)) |
|||
{ |
|||
result.Result.Add(name, name == "MyPermission8" || name == "MyPermission9" |
|||
? PermissionGrantResult.Prohibited |
|||
: PermissionGrantResult.Undefined); |
|||
} |
|||
|
|||
return Task.FromResult(result); |
|||
} |
|||
} |
|||
Loading…
Reference in new issue