TokenRateGate.Extensions 0.9.1

TokenRateGate

Stop getting HTTP 429 "Rate limit exceeded" errors from Azure OpenAI, OpenAI, and Anthropic Claude APIs.

TokenRateGate is a .NET library that prevents rate limit errors by intelligently managing your token and request budgets. It tracks both TPM (Tokens-Per-Minute) and RPM (Requests-Per-Minute) limits, automatically queues requests when capacity is full, and ensures you never hit the dreaded 429 error again.

The Problem

Getting this error when calling LLM APIs?

HTTP 429: Too Many Requests
Rate limit is exceeded. Try again in X seconds.

This happens when you exceed your API provider's limits:

• Azure OpenAI / Azure AI Foundry: TPM (Tokens-Per-Minute) limits based on your deployment tier
• OpenAI API: Both TPM and RPM (Requests-Per-Minute) limits
• Anthropic Claude: Rate limits based on usage tier

TokenRateGate prevents these errors by managing your token budget and queueing requests before they hit the API.

Features

• Prevents HTTP 429 Errors: Enforces TPM and RPM limits before making API calls
• Dual Limiting: Tracks both token usage (TPM) and request count (RPM) - whichever is more restrictive applies
• Smart Queuing: Automatically queues requests when capacity is full, with configurable timeouts
• Safety Buffer: Configurable buffer to avoid hitting exact limits (prevents edge-case 429s)
• Accurate Tracking: Records actual token usage from API responses for precise capacity management
• Multiple Providers: Built-in support for OpenAI and Azure OpenAI SDKs
• Multi-Tenant: Factory pattern for managing different rate limits per tenant/model/tier
• Real-Time Monitoring: Detailed usage statistics and capacity tracking
• Dependency Injection: First-class DI support with configuration binding
• High Performance: Optimized for throughput with minimal overhead
• Thread-Safe: Concurrent request handling with proper synchronization

Installation

Option 1: Complete Solution (Recommended)

For OpenAI or Azure OpenAI users:

dotnet add package TokenRateGate

Includes: Everything - base engine, DI, OpenAI integration, Azure integration.

Option 2: Custom LLM Providers

For Anthropic Claude, Google Gemini, or custom APIs:

dotnet add package TokenRateGate.Base

Includes: Core engine, DI support, character-based token estimation. Use when: Building custom integrations without OpenAI/Azure SDKs.

Option 3: Add Integrations Individually

# Base + OpenAI only
dotnet add package TokenRateGate.Base
dotnet add package TokenRateGate.OpenAI

# Base + Azure only
dotnet add package TokenRateGate.Base
dotnet add package TokenRateGate.Azure

Quick Start

1. Dependency Injection Setup (Recommended)

The easiest way to use TokenRateGate is with dependency injection:

using TokenRateGate.Extensions.DependencyInjection;

// In your Program.cs or Startup.cs
var builder = WebApplication.CreateBuilder(args);

// Register TokenRateGate with configuration
builder.Services.AddTokenRateGate(options =>
{
    options.TokenLimit = 150000;              // 150K tokens per minute (Azure Standard tier)
    options.WindowSeconds = 60;               // 60-second sliding window
    options.SafetyBufferPercentage = 0.05;    // 5% safety buffer (avoids hitting exact limit)
    options.MaxConcurrentRequests = 10;       // Limit concurrent API calls
    options.MaxRequestsPerMinute = 100;       // Optional: Also enforce RPM limit
});

var app = builder.Build();

Or bind from configuration:

// appsettings.json
{
  "TokenRateGate": {
    "TokenLimit": 500000,
    "WindowSeconds": 60,
    "MaxConcurrentRequests": 10
  }
}

builder.Services.AddTokenRateGate(
    builder.Configuration.GetSection("TokenRateGate"));

2. Using with Custom LLM Providers

For APIs without built-in token estimation (Anthropic Claude, Google Gemini, custom LLMs):

using TokenRateGate.Core;
using TokenRateGate.Core.TokenEstimation;
using TokenRateGate.Abstractions;

public class CustomLlmService
{
    private readonly ITokenRateGate _rateGate;
    private readonly ITokenEstimator _tokenEstimator;
    private readonly ILogger<CustomLlmService> _logger;

    public CustomLlmService(ITokenRateGate rateGate, ILogger<CustomLlmService> logger)
    {
        _rateGate = rateGate;
        _logger = logger;

        // Use character-based estimation (4 chars ≈ 1 token for most LLMs)
        _tokenEstimator = new CharacterBasedTokenEstimator();
    }

    public async Task<string> CallCustomLlmAsync(string prompt)
    {
        // Estimate tokens using character-based estimator
        int estimatedInputTokens = _tokenEstimator.EstimateTokens(prompt);
        int estimatedOutputTokens = 1000; // Your estimated response size

        // Reserve capacity before calling the LLM
        await using var reservation = await _rateGate.ReserveTokensAsync(
            estimatedInputTokens,
            estimatedOutputTokens);

        _logger.LogInformation("Reserved {Tokens} tokens", reservation.ReservedTokens);

        // Make your custom LLM API call
        var response = await CallYourCustomApiAsync(prompt);

        // Record actual usage from the response (IMPORTANT for accurate tracking)
        var actualTotalTokens = response.Usage.TotalTokens;
        reservation.RecordActualUsage(actualTotalTokens);

        _logger.LogInformation("Actual usage: {Tokens} tokens", actualTotalTokens);

        return response.Content;
    }
}

Using CharacterBasedTokenEstimator:

// Default: 4 characters per token
var estimator = new CharacterBasedTokenEstimator();
int tokens = estimator.EstimateTokens("Hello world!");  // ≈ 3 tokens

// Custom ratio for different languages
var chineseEstimator = new CharacterBasedTokenEstimator(charactersPerToken: 2.0);
int chineseTokens = chineseEstimator.EstimateTokens("你好世界");  // Better for non-Latin scripts

// Estimate multiple texts
var messages = new[] { "System prompt", "User message", "Assistant response" };
int totalTokens = estimator.EstimateTokens(messages);

3. Using with OpenAI SDK (Recommended for OpenAI Users)

TokenRateGate integrates seamlessly with the OpenAI SDK.

Note: Logging is optional. The WithRateLimit() extension method accepts an optional ILoggerFactory parameter for diagnostics. If not provided, logging is disabled.

using OpenAI.Chat;
using TokenRateGate.OpenAI;
using TokenRateGate.Abstractions;

public class ChatService
{
    private readonly ITokenRateGate _rateGate;
    private readonly string _apiKey;

    public ChatService(
        ITokenRateGate rateGate,
        IConfiguration configuration)
    {
        _rateGate = rateGate;
        _apiKey = configuration["OpenAI:ApiKey"];
    }

    public async Task<string> AskQuestionAsync(string question)
    {
        // Create OpenAI client and wrap with rate limiting
        var client = new ChatClient("gpt-4", _apiKey);
        var rateLimitedClient = client.WithRateLimit(_rateGate, "gpt-4");

        // Make rate-limited API call - automatic token tracking!
        var messages = new[] { new UserChatMessage(question) };
        var response = await rateLimitedClient.CompleteChatAsync(messages);

        return response.Content[0].Text;
    }

    public async Task<string> AskQuestionStreamingAsync(string question)
    {
        var client = new ChatClient("gpt-4", _apiKey);
        var rateLimitedClient = client.WithRateLimit(_rateGate, "gpt-4");

        var messages = new[] { new UserChatMessage(question) };
        var result = new StringBuilder();

        // Streaming support with automatic token tracking
        await foreach (var chunk in rateLimitedClient.CompleteChatStreamingAsync(messages))
        {
            if (chunk.ContentUpdate.Count > 0)
            {
                var text = chunk.ContentUpdate[0].Text;
                result.Append(text);
                Console.Write(text);
            }
        }

        return result.ToString();
    }
}

4. Using with Azure OpenAI (Recommended for Azure Users)

Note: Logging is optional. The WithRateLimit() extension method accepts an optional ILoggerFactory parameter for diagnostics. If not provided, logging is disabled.

using Azure;
using Azure.AI.OpenAI;
using TokenRateGate.Azure;

public class AzureChatService
{
    private readonly ITokenRateGate _rateGate;

    public AzureChatService(ITokenRateGate rateGate)
    {
        _rateGate = rateGate;
    }

    public async Task<string> AskQuestionAsync(string question)
    {
        var azureClient = new AzureOpenAIClient(
            new Uri("https://your-resource.openai.azure.com/"),
            new AzureKeyCredential("your-api-key"));

        // Wrap with rate limiting (deployment name + model name for token counting)
        var rateLimitedClient = azureClient.WithRateLimit(
            _rateGate,
            deploymentName: "my-gpt4-deployment",
            modelName: "gpt-4");

        var messages = new[] { new UserChatMessage(question) };
        var response = await rateLimitedClient.CompleteChatAsync(messages);

        return response.Content[0].Text;
    }
}

Multi-Tenant Configuration

Support different rate limits for different users, models, or tenants:

// Registration in Program.cs
builder.Services.AddTokenRateGateFactory();

builder.Services.AddNamedTokenRateGate("basic-tier", options =>
{
    options.TokenLimit = 100000;  // 100K tokens/min for basic users
    options.WindowSeconds = 60;
});

builder.Services.AddNamedTokenRateGate("premium-tier", options =>
{
    options.TokenLimit = 1000000; // 1M tokens/min for premium users
    options.WindowSeconds = 60;
});

// Usage in your service
public class MultiTenantChatService
{
    private readonly ITokenRateGateFactory _factory;

    public MultiTenantChatService(ITokenRateGateFactory factory)
    {
        _factory = factory;
    }

    public async Task<string> AskQuestionAsync(string question, string tier)
    {
        // Get rate gate for the tenant's tier
        var rateGate = _factory.GetOrCreate(tier);

        var client = new ChatClient("gpt-4", "your-api-key");
        var rateLimitedClient = client.WithRateLimit(rateGate, "gpt-4");

        var messages = new[] { new UserChatMessage(question) };
        var response = await rateLimitedClient.CompleteChatAsync(messages);

        return response.Content[0].Text;
    }
}

Standalone Usage (Without DI)

You can also use TokenRateGate without dependency injection:

using TokenRateGate.Core;
using TokenRateGate.Core.Options;
using TokenRateGate.OpenAI;
using Microsoft.Extensions.Logging;

// Create rate gate manually
var options = new TokenRateGateOptions
{
    TokenLimit = 500000,
    WindowSeconds = 60,
    MaxConcurrentRequests = 10
};

using var loggerFactory = LoggerFactory.Create(builder =>
{
    builder.AddConsole();
});

var rateGate = new TokenRateGate(options, loggerFactory);

// Use with OpenAI
var client = new ChatClient("gpt-4", "your-api-key");
var rateLimitedClient = client.WithRateLimit(rateGate, "gpt-4", loggerFactory);

var messages = new[] { new UserChatMessage("Hello!") };
var response = await rateLimitedClient.CompleteChatAsync(messages);
Console.WriteLine(response.Content[0].Text);

Monitoring Usage

public class MonitoringService
{
    private readonly ITokenRateGate _rateGate;

    public MonitoringService(ITokenRateGate rateGate)
    {
        _rateGate = rateGate;
    }

    public void LogCurrentUsage()
    {
        var stats = _rateGate.GetUsageStats();

        Console.WriteLine($"Current Usage: {stats.CurrentUsage}/{stats.EffectiveCapacity} tokens");
        Console.WriteLine($"Reserved: {stats.ReservedTokens} tokens");
        Console.WriteLine($"Available: {stats.AvailableTokens} tokens");
        Console.WriteLine($"Usage: {stats.UsagePercentage:F1}%");
        Console.WriteLine($"Near Capacity: {stats.IsNearCapacity}");
    }
}

Configuration Options

Output Estimation Strategies

• FixedMultiplier: Multiply input tokens by OutputMultiplier (default 0.5)
• FixedAmount: Add a fixed DefaultOutputTokens (default 1000)
• Conservative: Assume output = input (reserve 2x input tokens)

How It Works

TokenRateGate uses a dual-component capacity system:

Current Capacity = Historical Usage + Active Reservations

1. Token Estimation: Before making an API call, estimate input and output tokens
2. Capacity Check: Checks both:
   • TPM Check: (Historical Usage + Active Reservations + Requested Tokens) <= (TokenLimit - SafetyBuffer)
   • RPM Check: Current Request Count < MaxRequestsPerMinute
   • Both must pass - whichever is more restrictive applies
3. Reservation: If capacity available, reserves tokens immediately. Otherwise, queues the request with timeout.
4. API Call: Make your LLM API call
5. Record Actual Usage (Optional but recommended): Call RecordActualUsage() with actual tokens from response
   • If recorded: Actual usage tracked in sliding window for WindowSeconds
   • If not recorded: Reserved capacity freed immediately on disposal
6. Disposal: When using block ends, reservation is released and queued requests are processed
7. Sliding Window Cleanup:
   • Token timeline cleaned up every WindowSeconds
   • Request timeline cleaned up every RequestWindowSeconds (separate window for RPM)
   • Stale reservations removed after 10x WindowSeconds

Advanced Topics

Health Checks

builder.Services.AddHealthChecks()
    .AddTokenRateGate(name: "tokenrategate", tags: ["rate-limiting"]);

Custom Token Estimation

// Configure estimation strategy
builder.Services.AddTokenRateGate(options =>
{
    options.OutputEstimationStrategy = OutputEstimationStrategy.Conservative;
    // Now reserves 2x input tokens (assumes output = input)
});

Logging

TokenRateGate provides detailed structured logging:

builder.Services.AddLogging(logging =>
{
    logging.AddConsole();
    logging.SetMinimumLevel(LogLevel.Debug);  // See detailed token tracking
});

Samples

Check the samples/ directory for complete examples:

• OpenAIIntegration: Basic OpenAI usage, streaming, monitoring
• AzureOpenAI.BasicUsage: Azure OpenAI integration
• More samples available in the repository

Performance

• Minimal Overhead: Token estimation uses efficient tiktoken library
• Optimized Queuing: Fast capacity checks with double-check locking
• High Throughput: Achieves >95% capacity utilization under load
• Concurrent Requests: Supports high concurrency with proper synchronization

See tests/TokenRateGate.PerformanceTests for benchmarks.

Testing

# Run all tests
dotnet test

# Run specific test categories
dotnet test --filter "Category=Integration"
dotnet test --filter "Category=Performance"

Requirements

• .NET 6.0, 8.0, or 9.0
• OpenAI SDK (for OpenAI integration): OpenAI NuGet package
• Azure OpenAI SDK (for Azure integration): Azure.AI.OpenAI NuGet package

Packages

User-Facing Packages

• TokenRateGate ⭐: Complete solution - includes Base + OpenAI + Azure (recommended for most users)
• TokenRateGate.Base: Core engine + DI + Extensions (for custom LLM providers)
• TokenRateGate.OpenAI: OpenAI SDK integration (can be used with Base)
• TokenRateGate.Azure: Azure OpenAI SDK integration (can be used with Base)

Internal Packages (Included in Base)

You don't need to install these individually - they're included in TokenRateGate.Base:

• TokenRateGate.Core: Core rate limiting engine
• TokenRateGate.Abstractions: Interfaces and abstractions
• TokenRateGate.Extensions: Base implementations
• TokenRateGate.Extensions.DependencyInjection: DI support

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Make your changes with tests
4. Submit a pull request

License

MIT License - Copyright © 2025 Marko Mrdja

See LICENSE for details.

Acknowledgments

• Uses tiktoken for accurate token counting
• Built for the OpenAI SDK for .NET
• Supports Azure OpenAI SDK