longchain 0.4.1

A framework for LLM-powered interactions

longchain

Usage

• Quest - The main object. Ititialize it, then call asyncio.run(quest.run()) to run it.
• Path - A Path is one "step" in a player's experience.
• ActionResolver - This is the real decision-making part of the Path. The two main ones are the SequentialActionResolver, which takes several Actions and just runs them in order, and the LlmActionResolver, which uses an LLM to decide what to do.
  • starts_without_player_action - If this parameter is True, then a ChangePathAction with next_action='path' can immediately call this ActionResolver without waiting for player input.
• AgentAction - This is a single step a path takes. It's something like sending a message or moving to another path. They have a next_action parameter which controls whether the player is prompted for an input or not before the next action happens.
• core vs impl vs plugins - core is the main part of the library, impl is the implementation of the core interfaces, and plugins are more niche implementations. For example, the path following code is in core, the LlmActionResolver is in impl, and the Bag utilities are in plugins.
• Datastore - This is an interface for the quest to keep player state. The only one currently implemented is jsonfile, but another one could be implemented for a database or something.
• Messager - Like Datastore, this is an abstraction layer that allows the quest to send messages to the player. The only one currently implemented is slack.
• LlmActionResolver - There's a lot here so it gets its own bullet point. The LLM automatically talks to the player, and you can add LlmTools in the agent_actions parameter to allow it to do more. The system_prompt parameter can either be a string or a callable taking an LlmContext. This allows the system prompt to change based on what's going on, which is recommended (LLMs are not good at keeping information secret).
  • LlmTool - These describe an action the LLM can take. description should tell the LLM when to call the tool. params is a list of LlmToolParams, which have a name and a schema (which is a JSON Schema object). available and action are callables that take an LlmContext and return a boolean and an AgentAction respectively. available should return True if the tool can be used, and action should return the AgentAction to take if the tool is used. These both get LlmContexts, and action also gets a dict of the params the LLM called the action with. strict can turn on structured output, which forces the LLM's output to conform to the params perfectly but disables some JSON Schema properties. See more in the OpenAI docs.. If strict is off, don't trust the LLM to call your function correctly.

Example

There is an example project in the examples folder.

Development

to build:

.\venv\Scripts\activate
py -m build
python -m twine upload dist/*

to regenerate the bag api:

1. download the update bag.proto
2. pip install grpcio grpcio_tools
3. run cd .\src\longchain\plugins\bag\api\
4. run python -m grpc_tools.protoc -I . --grpc_python_out . --python_out . --pyi_out . bag.proto
5. Go to the generated bag_pb2_grpc.py and change the bag_pb2 import to import longchain.plugins.bag.api.bag_pb2 as bag__pb2

Naming

It's a tool for making long chains of NPC interactions.