This is the bare bones of a higher level abstraction that can support various higher level functionality with an AI Model.
The main abstractions are a ChatAgent that takes uses multiple PromptContextTransformers as collaborating objects. A PromptContextTransformer is a function that takes a PromptContext as input and output.
The implementation DefaultChatAgent can handle many use cases. It allows one to build an agent that can invoke functions (at the moment contained in the prompt) and execute various RAG algorithms. Augmentation can come from Node implementations, which are Documents and Messages. Augmenting Messages would be done for the case of including chat history into the prompt.
A fluent builder API for DefaultChatAgent puts PromptContextTransformers into different collections so that they can be executed at the right time in the default workflow inside the agent's call method. The DefaultChatAgent also contains a list of listeners that are invoked at the end of the ChatAgent's work so that state in external systems (such as a MessageHistory or MessageStore) can be updated.
There is a new Evaluator interface and one implementation RelevancyEvaluator
The test class OpenAiDefaultChatAgentIT minimally exercises this new functionality.
Open issues to consider:
- Adding a list of functions availalbe for the Agent to call can be achieved by transforming the prompt and adding function, but mayb
- PromptContext is currently mutable, this made implementation easy but we may want to consider returning a copy of a PromptContext from the transformer.
- MessageHistory has not been implemented.
Comment From: tzolov
Replaced by #634