-
🥞 Layered agents!
- Agents for your agents!
- Tool-use, structured output ("extraction"), and layering FTW 💪
- Ever wanted a recursive agent? Now you can have one! 🤯
- Parallel tool-calling by default.
- Fully asyncio.
- 100% Python type hints.
- Functional-style 😎
- (optional) Easy & pluggable caching! 🏦
-
🚣 Streamable!
- Event streams for everything.
- Asyncio generators are awesome.
-
🗃️ Easy database integration!
- Don't rage when trying to store raw messages and token counts. 😡 🤬
- Yes, you can have both streaming and easy database storage.
-
↔️ Provider/model agnostic and interoperable!- Native support for OpenAI, Anthropic, NVIDIA NIM/NGC (+ more to come).
- Message representations are canonized. 😇
- Supports vision!
- Easily build committees!
- Swap providers or models mid-conversation.
- Delegate tasks among model providers or model sizes.
- Parallelize all the things.
pip install -U lasagna-ai[openai,anthropic]If you want to easily run all the ./examples, then you can install the extra dependencies used by those examples:
pip install -U lasagna-ai[openai,anthropic,example-deps]Lasagna is used in production by:
Here is the most simple agent (it doesn't add anything to the underlying model). More complex agents would add tools and/or use layers of agents, but not this one! Anyway, run it in your terminal and you can chat interactively with the model. 🤩
(taken from ./examples/quickstart.py)
from lasagna import (
known_models,
build_simple_agent,
)
from lasagna.tui import (
tui_input_loop,
)
from typing import List, Callable
import asyncio
from dotenv import load_dotenv; load_dotenv()
MODEL_BINDER = known_models.BIND_OPENAI_gpt_4o_mini()
async def main() -> None:
system_prompt = "You are grumpy."
tools: List[Callable] = []
my_agent = build_simple_agent(name = 'agent', tools = tools)
my_bound_agent = MODEL_BINDER(my_agent)
await tui_input_loop(my_bound_agent, system_prompt)
if __name__ == '__main__':
asyncio.run(main())Want to add your first tool? LLMs can't natively do arithmetic (beyond simple arithmetic with small numbers), so let's give our model a tool for doing arithmetic! 😎
(full example at ./examples/quickstart_with_math_tool.py)
import sympy as sp
...
def evaluate_math_expression(expression: str) -> float:
"""
This tool evaluates a math expression and returns the result.
Pass math expression as a string, for example:
- "3 * 6 + 1"
- "cos(2 * pi / 3) + log(8)"
- "(4.5/2) + (6.3/1.2)"
- ... etc
:param: expression: str: the math expression to evaluate
"""
expr = sp.sympify(expression)
result = float(expr.evalf())
return result
...
...
tools: List[Callable] = [
evaluate_math_expression,
]
my_agent = build_simple_agent(name = 'agent', tools = tools)
...
...Simple RAG: Everyone's favorite tool: Retrieval Augmented Generation (RAG). Let's GO! 📚💨
See: ./examples/demo_rag.py
This library logs using Python's builtin logging module. It logs mostly to INFO, so here's a snippet of code you can put in your app to see those traces:
import logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
)
# ... now use Lasagna as you normally would, but you'll see extra log traces!Special thanks to those who inspired this library:
- Numa Dhamani (buy her book: Introduction to Generative AI)
- Dave DeCaprio's voice-stream library
lasagna-ai is distributed under the terms of the MIT license.
Layered Agents with toolS And aGeNts and Ai