llm-limiters 0.1.4

A thread-safe and async rate limiter for Gemini and OpenAI models.

llm-limiters

A lightweight, provider-aware rate-limiting library for LLM text, image, and live (streaming) models, supporting Gemini and OpenAI with tier-based quotas, fallbacks, and custom overrides.

Designed for:

• API gateways
• Backend services
• Agent systems
• Multi-model routing with hard quota enforcement

---

Features

• Tier-based rate limiting (free, tier1, tier2, …)
• Gemini and OpenAI support
• Text, Image, and Live (WebSocket) limiters
• RPM / TPM / RPD enforcement
• Automatic cooldown on quota and 429 errors
• Model-priority fallback
• Custom per-model overrides
• Thread-safe (sync and async)

---

Installation

pip install llm-limiters

Quick Start

Import

from llm_limiters import RateLimiter
from llm_limiters.live import LiveRateLimiter
from llm_limiters.image import ImageRateLimiter

Text Model Rate Limiting (Gemini + OpenAI)

Create a limiter

limiter = RateLimiter(tier="free")

Use with model fallback

@limiter.limit([
    "gemini-2.5-flash",
    "gpt-4o-mini"
])
def generate_text(prompt: str, model_name: str):
    return client.generate(
        model=model_name,
        prompt=prompt
    )

• Enforces RPM, TPM, and RPD
• Falls back to the next model when limits are reached
• Applies cooldowns automatically on rate-limit errors

Image Generation Rate Limiting

Supports

• Gemini Image / Imagen models
• OpenAI gpt-image-1, gpt-image-1-mini

Image generation requires billing (no free tier).

Create an image limiter

image_limiter = ImageRateLimiter(tier="tier1")

Use with priority routing

@image_limiter.limit([
    "gpt-image-1-mini",
    "gpt-image-1",
    "gemini-2.5-flash-preview-image",
    "imagen-4.0-fast-generate"
])
def generate_image(prompt: str, model_name: str):
    if model_name.startswith("gpt-image"):
        return openai.images.generate(
            model=model_name,
            prompt=prompt
        )
    else:
        return gemini.images.generate(
            model=model_name,
            prompt=prompt
        )

• RPM enforced for all image models
• TPM enforced for OpenAI image models
• RPD enforced for Gemini / Imagen models
• Automatic cooldown on quota errors

Live / Streaming Rate Limiter (Async)

Designed for:

• WebSockets
• Audio and real-time Gemini Live models

Create a live limiter

live_limiter = LiveRateLimiter(tier="free")

Check availability before starting a session

allowed, reason = await live_limiter.check_availability(
    "gemini-2.0-flash-live-001"
)

if not allowed:
    raise Exception(reason)

Track usage

await live_limiter.acquire_session(model)
await live_limiter.record_token_usage(model, token_count)

Trigger cooldown on provider rate-limit

await live_limiter.trigger_cooldown(model, wait_seconds=60)

Custom Limits and Overrides

Custom limits can override tier defaults or add new models.

Override a single model

limiter = ImageRateLimiter(
    tier="tier2",
    custom_limits={
        "gpt-image-1": {
            "rpm": 100,
            "tpm": 500_000,
            "rpd": 0
        }
    }
)

Add a custom or internal model

limiter = ImageRateLimiter(
    tier="tier1",
    custom_limits={
        "internal-image-model": {
            "rpm": 20,
            "tpm": 0,
            "rpd": 100
        }
    }
)

Custom limits always take precedence over built-in tiers.

Error Handling

from llm_limiters import RateLimitExceededError, ModelNotFoundError

• RateLimitExceededError — all models in the priority list are exhausted
• ModelNotFoundError — requested model is not configured

Built-In Limits

Built-in limits are defined in constants.py and include:

• Gemini text models
• OpenAI text models
• Gemini Live models
• Gemini Image / Imagen models
• OpenAI image models

Tier-2 and higher image limits are intentionally empty until discovered and overridden.

License

MIT License Copyright (c) 2025 Nagomi Jayamani