planning-library 0.1.4

Description:

planninglibrary 0.1.4

🤖✨ Planning Library
Library with planning algorithms for AI Agents built with LangChain and LangGraph.
Installation
As a package


For Poetry:
poetry add planning-library



For pip:
pip install planning-library



For development
Step 0: Install prerequisites
In general, the only prerequisite is :snake: Python. However, note the TextWorld requirements if you run into any issues.

You can use pyenv to set the specific Python version.

Step 2: Clone repository
git clone [email protected]:JetBrains-Research/planning-library.git

Step 3: Install Python dependencies

For Poetry: run poetry install.

Note. If you do not need to run code quality checks or to run examples, you can exclude the corresponding dependencies groups: poetry install --without dev,examples



Quick Tour
Currently, we have two types of strategies: custom strategies and
LangGraph strategies.
Custom strategies
Custom strategies follow the interface provided
by BaseCustomStrategy.
Example: Tree of Thoughts + DFS
Initializing strategy
Each custom strategy can be created by invoking a static method create with at least agent and tools.
from planning_library.strategies import TreeOfThoughtsDFSStrategy

agent = ... # any runnable that follows either RunnableAgent or RunnableMultiActionAgent
tools = [...] # any sequence of tools
strategy_executor = TreeOfThoughtsDFSStrategy.create(
agent=agent,
tools=tools,
)

Some strategies contain other meaningful components (e.g., an evaluator, which is responsible for evaluating
intermediate steps). :construction: We will provide some default implementations for such components, but they can also
be redefined with custom runnables tailored for specific tasks.
Using strategy
Each custom strategy is an instance of Chain and mostly can be
used the same
way as the default AgentExecutor from
LangChain.
strategy_executor.invoke({"inputs": "Hello World"})

LangGraph strategies
Strategies powered by LangGraph library follow the interface provided
by BaseLangGraphStrategy.
Example: Reflexion
Initializing strategy
Each LangGraph strategy can be created by invoking a static method create with (at least) agent and tools.
from planning_library.strategies import ReflexionStrategy

agent = ... # any runnable that follows either RunnableAgent or RunnableMultiActionAgent
tools = [...] # any sequence of tools
strategy_graph = ReflexionStrategy.create(agent=agent, tools=tools)

Some strategies contain other meaningful components (e.g., an evaluator, which is responsible for evaluating
intermediate steps). :construction: We will provide some default implementations for such components, but they can also
be redefined with custom runnables tailored for specific tasks.
Using strategy
BaseLangGraphStrategy.create returns a
compiled StateGraph that exposes the same
interface as any LangChain runnable.
strategy_graph.invoke({"inputs": "Hello World"})

Available Strategies



Name
Implementation
Type
Paper




Tree of Thoughts + DFS / DFSDT
TreeOfThoughtsDFSStrategy
Custom
:scroll: ToT, :scroll: DFSDT


Reflexion
ReflexionStrategy
LangGraph
:scroll:


ADaPT
ADaPTStrategy
Custom
:scroll:


Simple/ReAct
SimpleStrategy
Custom
:scroll:



Available Environments
:two::four: Game of 24

Game of 24 is a mathematical reasoning task. The goal is to reach the number 24 by applying arithmetical operations
to four given numbers. See :scroll: Tree of Thoughts paper for more details.

Our implementation of Game of 24 is available under environments/game_of_24 folder. It
includes a set of prompts, a set of tools and examples of running available strategies on Game of 24.

Common:

Gymnasium env for Game of
24: environments/game_of_24/common/environment.py
Tools for Game of 24: environments/game_of_24/common/tools.py



:snowflake: FrozenLake

FrozenLake is a simple environment that requires crossing a frozen lake from start to goal without falling into any
holes.
See Gymnasium docs for more details.

Our implementation of FrozenLake is available under environments/frozen_lake folder.

Common:

Env wrapper for
FrozenLake: environments/frozen_lake/common/environment.py
Tools for FrozenLake: environments/frozen_lake/common/tools.py



:house: ALFWorld

ALFWorld contains interactive TextWorld environments for household navigation. See :scroll:
ALFWorld paper or project website for more
information.

Our implementation of ALFWorld is available under environments/alfword folder.

Common:

Env wrapper for
ALFWorld: environments/alfworld/common/environment.py
Tools for ALFWorld: environments/alfworld/common/tools.py



Strategies usage examples
Examples are available under examples folder.