change return of LLMChain.run/2 - breaking change

[brainlid]

• Remove the "last_message" from the return tuple of LLMChain.run/2
• Data is still available in %LLMChain{last_message: last_message}
• Adds %LLMChain{exchanged_messages: exchanged_messages}.

This is a breaking change. Let's cover why this is being done and how to adapt to it.

Why the change

Before this change, an LLMChain's run function returned {:ok, updated_chain, last_message}.

When an assistant (ie LLM) issues a ToolCall and when run is in the mode :until_success or :while_need_response, the LLMChain will automatically execute the function and return the result as a new Message back to the LLM. This works great!

The problem comes when an application needs to keep track of all the messages being exchanged during a run operation. That can be done by using callbacks and sending and receiving messages, but that's far from ideal. It makes more sense to have access to that information directly after the run operation completes.

What this change does

This PR changes the returned type to {:ok, updated_chain}.

The last_message is available in updated_chain.last_message. This cleans up the return API.

This change also adds %LLMChain{exchanged_messages: exchanged_messages},or updated_chain.exchanged_messages which is a list of all the messages exchanged between the application and the LLM during the execution of the run function.

This breaks the return contract for the run function.

How to adapt to this change

To adapt to this, if the application isn't using the last_message in {:ok, updated_chain, _last_message}, then delete the third position in the tuple. Ex: {:ok, updated_chain}.

Access to the last_message is available on the updated_chain.

{:ok, updated_chain} =
  %{llm: model}
  |> LLMChain.new!()
  |> LLMChain.run()

last_message = updated_chain.last_message

NOTE: that the updated_chain now includes updated_chain.exchanged_messages which can also be used.