tinyloop 0.1.35

A super lightweight library for LLM-based applications

A lightweight Python library for building AI-powered applications with clean function calling, vision support, and MLflow integration.

TinyLoop is fully built on top of LiteLLM, providing 100% compatibility with the LiteLLM API while adding powerful abstractions and utilities. This means you can use any model, provider, or feature that LiteLLM supports, including:

• All LLM Providers: OpenAI, Anthropic, Google, Azure, Cohere, and 100+ more
• All Model Types: Chat, completion, embedding, and vision models
• Advanced Features: Streaming, function calling, structured outputs, and more
• Ops Features: Retries, fallbacks, caching, and cost tracking

TinyLoop provides a clean, intuitive interface for working with Large Language Models (LLMs), featuring:

• 🎯 Clean Function Calling: Convert Python functions to JSON tool definitions automatically
• 🔍 MLflow Integration: Built-in tracing and monitoring with customizable span names
• 👁️ Vision Support: Handle images and vision models seamlessly
• 📊 Structured Output: Generate structured data from LLM responses using Pydantic
• 🔄 Tool Loops: Execute multi-step tool calling workflows
• ⚡ Async Support: Full async/await support for all operations

📦 Installation

pip install tinyloop

🚀 Quick Start

Basic LLM Usage

Synchronous Calls

from tinyloop.inference.litellm import LLM

# Initialize the LLM
llm = LLM(model="openai/gpt-3.5-turbo", temperature=0.1)

# Simple text generation
response = llm(prompt="Hello, how are you?")
print(response)

# Get conversation history
history = llm.get_history()

# Access comprehensive response information
print(f"Response: {response}")
print(f"Cost: ${response.cost:.6f}")
print(f"Tool calls: {response.tool_calls}")
print(f"Raw response: {response.raw_response}")
print(f"Message history: {len(response.message_history)} messages")

Asynchronous Calls

from tinyloop.inference.litellm import LLM

llm = LLM(model="openai/gpt-3.5-turbo", temperature=0.1)

# Async text generation
response = await llm.acall(prompt="Hello, how are you?")
print(response)

🔄 Tool Loops

Execute multi-step tool calling workflows:

from tinyloop.modules.tool_loop import ToolLoop
from tinyloop.features.function_calling import Tool
from pydantic import BaseModel
import random

def roll_dice():
    """Roll a dice and return the result"""
    return random.randint(1, 6)

class FinalAnswer(BaseModel):
    last_roll: int
    reached_goal: bool

# Create tool loop
loop = ToolLoop(
    model="openai/gpt-4.1",
    system_prompt="""
    You are a dice rolling assistant.
    Roll a dice until you get the number indicated in the prompt.
    Use the roll_dice function to roll the dice.
    Return the last roll and whether you reached the goal.
    """,
    temperature=0.1,
    output_format=FinalAnswer,
    tools=[Tool(roll_dice)]
)

# Execute the loop
response = loop(
    prompt="Roll a dice until you get a 6",
    parallel_tool_calls=False,
)

print(f"Last roll: {response.last_roll}")
print(f"Reached goal: {response.reached_goal}")

Supported Features

🎯 Structured Output Generation

Generate structured data using Pydantic models:

from tinyloop.inference.litellm import LLM
from pydantic import BaseModel
from typing import List

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: List[str]

class EventsList(BaseModel):
    events: List[CalendarEvent]

# Initialize LLM with structured output
llm = LLM(
    model="openai/gpt-4.1-nano",
    temperature=0.1,
)

# Generate structured data
response = llm(
    prompt="List 5 important events in the XIX century",
    response_format=EventsList
)

# Access structured data
for event in response.events:
    print(f"{event.name} - {event.date}")
    print(f"Participants: {', '.join(event.participants)}")

👁️ Vision

Work with images using various input methods:

from tinyloop.inference.litellm import LLM
from tinyloop.features.vision import Image
from PIL import Image as PILImage

llm = LLM(model="openai/gpt-4.1-nano", temperature=0.1)

# From PIL Image
pil_image = PILImage.open("image.jpg")
image = Image.from_PIL(pil_image)

# From file path
image = Image.from_file("image.jpg")

# From URL
image = Image.from_url("https://example.com/image.jpg")

# Analyze image
response = llm(prompt="Describe this image", images=[image])
print(response)

🔧 Function Calling

Convert Python functions to LLM tools with automatic schema generation:

from tinyloop.inference.litellm import LLM
from tinyloop.features.function_calling import Tool
import json

def get_current_weather(location: str, unit: str):
    """Get the current weather in a given location

    Args:
        location: The city and state, e.g. San Francisco, CA
        unit: Temperature unit {'celsius', 'fahrenheit'}

    Returns:
        A sentence indicating the weather
    """
    if location == "Boston, MA":
        return "The weather is 12°F"
    return f"Weather in {location} is sunny"

# Create LLM instance
llm = LLM(model="openai/gpt-4.1-nano", temperature=0.1)

# Create tool from function
weather_tool = Tool(get_current_weather)

# Use function calling
inference = llm(
    prompt="What is the weather in Boston, MA?",
    tools=[weather_tool],
)

# Process tool calls
for tool_call in inference.raw_response.choices[0].message.tool_calls:
    tool_name = tool_call.function.name
    tool_args = json.loads(tool_call.function.arguments)
    print(f"Tool: {tool_name}")
    print(f"Args: {tool_args}")
    print(weather_tool(**tool_args))

# Access comprehensive response information
print(f"Total cost: ${inference.cost:.6f}")
print(f"Tool calls made: {len(inference.tool_calls) if inference.tool_calls else 0}")
print(f"Conversation length: {len(inference.message_history)} messages")

📝 Generate Module

Simple text generation with a clean interface:

from tinyloop.modules.generate import Generate

# Synchronous generation
response = Generate.run(
    prompt="Write a haiku about programming",
    model="openai/gpt-3.5-turbo",
    temperature=0.7
)
print(response.response)

# Async generation
response = await Generate.arun(
    prompt="Explain quantum computing",
    model="openai/gpt-4",
    temperature=0.3
)
print(response.response)

# Using the class for multiple calls
generator = Generate(
    model="openai/gpt-3.5-turbo",
    temperature=0.5,
    system_prompt="You are a helpful coding assistant."
)

response1 = generator.call("How do I implement a binary search?")
response2 = generator.call("What's the time complexity?")

🎨 Prompt Rendering

Manage prompts with YAML templates and Jinja2:

from tinyloop.utils.prompt_renderer import PromptRenderer, render_base_prompts

# Using PromptRenderer class
renderer = PromptRenderer("prompts/chat.yaml")
system_prompt = renderer.render("system", user_name="Alice", context="coding")
user_prompt = renderer.render("user", question="How do I debug Python?")

Example YAML prompt file (prompts/chat.yaml):

system: |
  You are {{ user_name }}, a helpful AI assistant specializing in {{ context }}.
  Always provide clear, actionable advice.

user: |
  {{ user_name }}, I have a question: {{ question }}

  Please provide a detailed response with examples if relevant.

🌊 Streaming Responses

Get real-time responses as they're generated:

from tinyloop.inference.litellm import LLM

llm = LLM(model="openai/gpt-3.5-turbo", temperature=0.1)

# Stream responses
for chunk in llm.stream(prompt="Write a story about a robot"):
    print(chunk.response, end="", flush=True)

🔄 Async Tool Loops

Execute tool loops asynchronously for better performance:

import asyncio
from tinyloop.modules.tool_loop import ToolLoop
from tinyloop.features.function_calling import Tool
from pydantic import BaseModel

def fetch_data(source: str):
    """Fetch data from a source"""
    return f"Data from {source}: [1, 2, 3, 4, 5]"

def process_data(data: str):
    """Process the fetched data"""
    return f"Processed: {data}"

class AnalysisResult(BaseModel):
    final_result: str
    steps_completed: int

async def main():
    loop = ToolLoop(
        model="openai/gpt-4",
        system_prompt="You are a data analyst. Fetch and process data step by step.",
        temperature=0.1,
        output_format=AnalysisResult,
        tools=[Tool(fetch_data), Tool(process_data)]
    )

    result = await loop.acall(
        prompt="Fetch data from 'api' and process it"
    )
    print(f"Final result: {result.final_result}")
    print(f"Steps completed: {result.steps_completed}")

# Run the async function
asyncio.run(main())

🔍 Advanced Observability: MLflow Integration

Custom Span Names

Create custom MLflow spans with meaningful names:

from tinyloop.utils.mlflow import mlflow_trace
from tinyloop.features.function_calling import Tool
import mlflow

def get_weather(location: str, unit: str = "celsius"):
    """Get weather for a location"""
    return f"Weather in {location}: 20°{unit}"

def get_stock_price(symbol: str, currency: str = "USD"):
    """Get stock price for a symbol"""
    return f"Stock price for {symbol}: $150.00 {currency}"

# Create tools with custom names for better tracing
weather_tool = Tool(get_weather, name="weather_service")
stock_tool = Tool(get_stock_price, name="stock_service")

# Start MLflow run
with mlflow.start_run():
    # Call tools - these will create spans with custom names
    weather_result = weather_tool("London", "fahrenheit")
    stock_result = stock_tool("AAPL", "USD")

    # The MLflow spans will be named:
    # - "weather_service.__call__" for the weather tool
    # - "stock_service.__call__" for the stock tool

Custom Agent Tracing

from tinyloop.utils.mlflow import mlflow_trace

class ResearchAgent:
    def __init__(self):
        self.llm = LLM(model="openai/gpt-4", temperature=0.1)

    @mlflow_trace(mlflow.entities.SpanType.AGENT)
    def research_topic(self, topic: str):
        """Research a topic comprehensively"""
        response = self.llm(
            prompt=f"Research the topic: {topic}. Provide key insights and sources."
        )
        return response

    @mlflow_trace(mlflow.entities.SpanType.AGENT)
    def analyze_findings(self, findings: str):
        """Analyze research findings"""
        response = self.llm(
            prompt=f"Analyze these findings: {findings}. What are the implications?"
        )
        return response

# Usage with automatic tracing
agent = ResearchAgent()
research_result = agent.research_topic("artificial intelligence")
analysis_result = agent.analyze_findings(research_result.response)

🛡️ Error Handling and Retries

Handle errors gracefully with retry patterns:

from tinyloop.inference.litellm import LLM
import time
import random

def robust_llm_call(llm, prompt, max_retries=3, delay=1):
    """Make LLM calls with retry logic"""
    for attempt in range(max_retries):
        try:
            response = llm(prompt=prompt)
            return response
        except Exception as e:
            if attempt == max_retries - 1:
                raise e
            print(f"Attempt {attempt + 1} failed: {e}")
            time.sleep(delay * (2 ** attempt) + random.uniform(0, 1))

    return None

# Usage
llm = LLM(model="openai/gpt-3.5-turbo", temperature=0.1)
response = robust_llm_call(
    llm,
    "Explain the concept of machine learning",
    max_retries=3
)
print(response.response)

🔭 Observability (Opt-in)

TinyLoop includes optional tracing integrations. By default, tracing/export is disabled (no-op) to avoid surprise side effects such as:

• Creating local mlruns/ directories
• Noisy “Exception while exporting Span” errors when no collector/server is running

You can enable each integration explicitly with environment variables (recommended), or via module parameters where available.

Langfuse / OpenTelemetry (no-op by default)

TinyLoop uses a safe wrapper for Langfuse’s observe decorator. Unless enabled, all @observe(...) decorators are no-ops.

• Enable:

export TINYLOOP_ENABLE_LANGFUSE=1

• Disable (default):

export TINYLOOP_ENABLE_LANGFUSE=0

If you enable Langfuse, make sure your OTEL/Langfuse endpoint is running and configured in your environment.

MLflow (autolog is opt-in)

TinyLoop provides MLflow tracing helpers (e.g. mlflow_trace) and can optionally enable MLflow + LiteLLM autologging.

• Enable MLflow LiteLLM autologging (for ToolLoop):

export TINYLOOP_ENABLE_MLFLOW=1

• Disable (default):

export TINYLOOP_ENABLE_MLFLOW=0

You can also enable/disable it per ToolLoop instance:

from tinyloop.modules.tool_loop import ToolLoop

loop = ToolLoop(
    model="openai/gpt-4.1",
    tools=[],
    output_format=dict,  # example only
    enable_mlflow=False,  # default is None (use env var)
)

🔍 Observability: MLflow Integration

Automatic Tracing

TinyLoop supports MLflow tracing utilities, and (optionally) MLflow + LiteLLM autologging.

• By default, TinyLoop does not enable MLflow autologging at import time (to avoid creating local mlruns/ unexpectedly).
• To enable MLflow LiteLLM autologging for ToolLoop, set TINYLOOP_ENABLE_MLFLOW=1 or pass enable_mlflow=True to ToolLoop.

from tinyloop.utils.mlflow import mlflow_trace

class Agent:
    @mlflow_trace(mlflow.entities.SpanType.AGENT)
    def __call__(self, prompt: str, **kwargs):
        self.llm.add_message(self.llm._prepare_user_message(prompt))
        for _ in range(self.max_iterations):
            response = self.llm(
                messages=self.llm.get_history(), tools=self.tools, **kwargs
            )
            if response.tool_calls:
                should_finish = False
                for tool_call in response.tool_calls:
                    tool_response = self.tools_map[tool_call.function_name](
                        **tool_call.args
                    )

                    self.llm.add_message(
                        self._format_tool_response(tool_call, str(tool_response))
                    )

                    if tool_call.function_name == "finish":
                        should_finish = True
                        break

                if should_finish:
                    break

        return self.llm(
            messages=self.llm.get_history(),
            response_format=self.output_format,
    )

🏗️ Project Structure

tinyloop/
├── features/
│   ├── function_calling.py  # Function calling utilities
│   └── vision.py           # Vision model support
├── inference/
│   ├── base.py             # Base inference classes
│   └── litellm.py          # LiteLLM integration
├── modules/
│   ├── base_loop.py        # Base loop implementation
│   ├── generate.py         # Generation modules
│   └── tool_loop.py        # Tool execution loop
└── utils/
    └── mlflow.py           # MLflow utilities

🧪 Development

Running Tests

# Run all tests
pytest tests/

# Run specific test file
pytest tests/test_function_calling.py -v

# Run with coverage
pytest tests/ --cov=tinyloop

Examples

Check out the Jupyter notebooks for more detailed examples:

• basic_usage.ipynb - Basic usage examples
• modules.ipynb - Advanced module usage

🤝 Contributing

We welcome contributions! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Made with ❤️ for the AI community