feat[agent]: add BaseAgent and standardize planner, memory, tool mana… #385
Conversation
…ger (backward compatible)
| <END_OF_USER_QUERY> | ||
| """ | ||
|
|
||
| OUTPUT_FORMAT = r"""You should provide a JSON object in this structure: |
| class ReActMemory(BaseMemory): | ||
| """ReAct-specific memory implementation.""" | ||
|
|
||
| def retrieve(self, query: str) -> List[Step]: |
There was a problem hiding this comment.
Yup In progress few things still left to add and test. Memory, training.. will let you know once PR is ready for review in some time.
|
can you review and let me know is it good or what until now @liyin2015 pls. |
| @@ -0,0 +1,134 @@ | |||
| """Example demonstrating ReAct agent with vector memory support.""" | |||
There was a problem hiding this comment.
examples should not be in the package, please move it to tutorials
| raise NotImplementedError | ||
|
|
||
|
|
||
| class BaseToolManager(Component): |
There was a problem hiding this comment.
we dont need another base tool manager, the toolmanager is already abstracted enough.
The goal is not to create more than 3 levels of abstraction especially those not doing much work.
The abstraction of the agent should be more focused on the design pattern of agents, such as openai's agent sdk, where it makes it easier to create an agent from user's experience and still remain the control over the prompts if the user wants to modify it.
also we are trying to make it easy to hand over the context (memory) and the step history to all other agents (as a tool).
i have some code and i can share later.
we can do another quick sync call next week.
you can check out openai agent sdk before that.
| raise NotImplementedError | ||
|
|
||
|
|
||
| class BaseAgent(Component): |
There was a problem hiding this comment.
another complexity of the agent is at the auto-optimization, we can make this new agent experimental and it might need multiple iterations to get more mature
Agent Architecture Refactoring
⚠NOTE: TESTING UNDER PROGRESS!!⚠
Overview
Refactored agent architecture to improve modularity and maintainability by introducing a base agent class and restructuring the ReAct agent implementation.
Key Changes
BaseAgentwith standardized interfaces for planning, tool management, and memoryReActAgentto use the new base architectureReActPlanner,ReActToolManager,ReActMemory)Benefits
Breaking Changes
None. Both old (
LegacyReActAgent) and new (NewReActAgent) implementations are available.