Documentation | Discord | PyPI | Roadmap
Harness the power of next-generation AI while maintaining complete control and reliability.
The AI revolution is transforming how we work - but with great power comes great responsibility. LION provides the control mechanisms and reliability features needed to safely integrate advanced AI capabilities into enterprise workflows.
LION is designed to be:
- 🔒 Controlled: Built-in safety mechanisms and verification
- 🎯 Precise: Exact control over AI behaviors
- 🔧 Flexible: Build any workflow you need
- 🚀 Efficient: Minimal dependencies, maximum performance
LION maintains minimal dependencies for maximum reliability:
uv pip install lionagiDependencies:
- litellm
- jinja2
- pandas
- pillow
- python-dotenv
from lionagi import iModel, Branch
# Initialize model
gpt4o = iModel(provider="openai", task="chat", model="gpt-4o")
hunter = Branch(
system="you are a hilarious dragon hunter who responds in 10 words rhymes",
imodel=gpt4o,
)
# Chat asynchronously
print(await hunter.communicate("I am a dragon"))You claim to be a dragon, oh what a braggin'!
LION provides a unified interface for interacting with any AI model, regardless of the underlying architecture. This allows you to easily switch between models without changing your code.
from pydantic import BaseModel
class Joke(BaseModel):
joke: str
sonnet = iModel(
provider="anthropic",
model="claude-3-5-sonnet-20241022",
max_tokens=100, # max_tokens is required for anthropic models
)
response = await hunter.communicate(
instruction="I am a dragon",
response_format=Joke, # structured output in given pydantic model
clear_messages=True, # refresh the conversation
imodel=sonnet, # use sonnet model, which doesn't support structured output
)
print(type(response))
print(response.joke)<class '__main__.Joke'>
Joke(joke='With fiery claws, dragons hide their laughter flaws!')
# using perplexity model
pplx_small = iModel(
provider="perplexity",
task="chat/completions",
model="llama-3.1-sonar-small-128k-online",
max_tokens=1000,
)
b = await hunter.communicate(
instruction="What makes a well-behaved dragon?",
clear_messages=True, # refresh the conversation
imodel=pplx_small, # use perplexity model
)
print(b)A well-behaved dragon is one that's calm and bright,
No stress or fear, just a peaceful night.
It's active, not lethargic, with a happy face,
And behaviors like digging, not a frantic pace.
It's social, friendly, and never a fright,
Just a gentle soul, shining with delight
hunter.msgs.last_response.model_response{'id': '1be10f4c-0936-4050-ab48-91bd86ab11a5',
'model': 'llama-3.1-sonar-small-128k-online',
'object': 'chat.completion',
'created': 1734369700,
'choices': [{'index': 0,
'message': {'role': 'assistant',
'content': "A well-behaved dragon is one that's calm and bright,\nNo stress or fear, just a peaceful night.\nIt's active, not lethargic, with a happy face,\nAnd behaviors like digging, not a frantic pace.\nIt's social, friendly, and never a fright,\nJust a gentle soul, shining with delight"},
'finish_reason': 'stop',
'delta': {'role': 'assistant', 'content': ''}}],
'usage': {'prompt_tokens': 40, 'completion_tokens': 69, 'total_tokens': 109},
'citations': [{'url': 'https://dragonsdiet.com/blogs/dragon-care/15-bearded-dragon-behaviors-and-what-they-could-mean'},
{'url': 'https://masterbraeokk.tripod.com/dragons/behavior.html'},
{'url': 'https://files.eric.ed.gov/fulltext/ED247607.pdf'},
{'url': 'https://www.travelchinaguide.com/intro/social_customs/zodiac/dragon/five-elements.htm'},
{'url': 'https://www.travelchinaguide.com/intro/social_customs/zodiac/dragon/'}]}
# chain of thoughts
from pydantic import Field
class Reason(BaseModel):
reason: str
confidence_score: float
class Thought(BaseModel):
thought: str
class Analysis(BaseModel):
thought: list[Thought] = Field(
default_factory=list,
description="concise Chain of thoughts from you, 3 step, each in 8 words"
)
analysis: str = Field(
...,
description="Final analysis of the dragon's psyche in 20 words",
)
reason: list[Reason] = Field(
default_factory=list,
description="Concise Reasoning behind the analysis, 3 support, each in 8 words"
)
context1 = "I am a dragon, I think therefore I am, I suffer from shiny objects syndrome"
context2 = "I like food and poetry, I use uv sometimes, it's cool but I am not familiar with pip"
async def analyze(context) -> Analysis:
psychologist = Branch(
system="you are a renowned dragon psychologist",
imodel=gpt4o,
)
return await psychologist.communicate(
instruction="analyze the dragon's psyche using chain of thoughts",
guidance="think step by step, reason with logic",
context=context,
response_format=Analysis,
)result1 = await analyze(context1)
print("\nThoughts:")
for i in result1.thought:
print(i.thought)
print("\nAnalysis:")
print(result1.analysis)
print("\nReasoning:")
for i in result1.reason:
print(i.reason)
Thoughts:
Dragons are attracted to shiny objects naturally.
This suggests a strong affinity for hoarding.
Reflects the dragon's inherent desire for possession.
Analysis:
The dragon demonstrates a compulsive hoarding behavior linked to attraction for shiny objects.
Reasoning:
Shiny objects trigger instinctual hoarding behavior.
Possession indicates a symbol of power and security.
Hoarding is reinforced by evolutionary survival mechanisms.
result2 = await analyze(context2)
print("\nThoughts:")
for i in result2.thought:
print(i.thought)
print("\nAnalysis:")
print(result2.analysis)
print("\nReasoning:")
for i in result2.reason:
print(i.reason)Thoughts:
Dragon enjoys both food and poetry regularly.
Dragon uses uv light with frequent interest.
Dragon is unfamiliar and not comfortable with pip.
Analysis:
The dragon is curious and exploratory, yet selectively cautious about unfamiliar methodologies.
Reasoning:
Preference for food and poetry suggests curiosity.
Frequent uv light use indicates exploratory nature.
Discomfort with pip usage shows selective caution.
Below is an example of what you can build with LION. Note that these are sample implementations - LION provides the building blocks, you create the workflows that fit your needs.
sequenceDiagram
autonumber
participant Client
participant Orchestrator
participant ResearchAgent
participant AnalysisAgent
participant ValidationAgent
participant Tools
Client->>+Orchestrator: Submit Complex Task
Note over Orchestrator: Task Analysis & Planning
%% Research Phase
Orchestrator->>+ResearchAgent: Delegate Research
activate ResearchAgent
ResearchAgent->>Tools: Access Data Sources
Tools-->>ResearchAgent: Raw Data
ResearchAgent-->>-Orchestrator: Research Results
deactivate ResearchAgent
%% Analysis Phase
Orchestrator->>+AnalysisAgent: Process Data
activate AnalysisAgent
AnalysisAgent->>Tools: Apply Models
Tools-->>AnalysisAgent: Analysis Results
AnalysisAgent-->>-Orchestrator: Processed Insights
deactivate AnalysisAgent
%% Validation Phase
Orchestrator->>+ValidationAgent: Verify Results
activate ValidationAgent
ValidationAgent->>Tools: Apply Safety Checks
Tools-->>ValidationAgent: Validation Status
ValidationAgent-->>-Orchestrator: Verified Results
deactivate ValidationAgent
Orchestrator-->>-Client: Return Validated Output
Join our Discord community to:
- Share ideas
- Report issues
- Contribute code
- Learn from others
@software{Li_LionAGI_2023,
author = {Haiyang Li},
month = {12},
year = {2023},
title = {LionAGI: Towards Automated General Intelligence},
url = {https://github.com/lion-agi/lionagi},
}