chain.md

Chain

If you have worked on LangChain before, you already knows 80% of what chain does. Chain is the stateful orchestrator for agent, which controls when to involve agent in which way. It offers a framework to interact with agents by controlling the information flow. Chain leverages a memory component that memorizes past conversations and any other intermediate steps, such as tool outputs.

For a typical chain, it contains an agent to interact with, list of tools that agent might use and memory that stored stateful information.

Flow diagram describes the high level picture of the default chain interaction with an agent.

Differences with LangChain

To remove abstraction and internal concepts, we expose a more flatten and simplified interface for chain and explained in details below. The main difference is we simply the flow and removed as many internal concepts as possible.

BaseChain and Chain

This is the most generic interface for implementing any chain in AutoChain. It contains a few features user could override. BaseChain is the generic interface where Chain is the default chain implementation by implementing the only abstract method take_next_step

run

this is the entry point to interact with chain with user_query. Because there are often different ways to interact with the chain, any chain could override the _run function that handles the business logics for agent interaction, while still be benefited from input and output standardization provided by the BaseChain.

_run

This provide the standard way to manage memories and determines when the agent should stop answering user query. Most of the time, user could reuse how we manage memory and just need to change what is the next step agent should do given inputs including user_query and memories. In that case, user would need to implement the take_next_step function.

take_next_step

This is an abstract method implements the way you would like to interact with agent and asking agent to come up with the next step. The default implementation is in Chain, where it asks the agent to plan for next step and execute tools selected by agent

should_answer

It is often unclear when agent should stop responding to user query. Sometimes user would just say "Thank you" in the end but agent might not understand this as end of the conversation, so it could still try to respond with more contents, even clarifying questions in some cases. By default, agent will always respond to user until user stops. In the case that is not desired, we introduce the should_answer step in BaseChain to stop agent from further interaction.