Adding support for a new runnable for InlineAgents #340
Conversation
divekarshubham
force-pushed
the
inline_agents_support
branch
from
1d963b7 to
10ad334
Compare
divekarshubham
force-pushed
the
inline_agents_support
branch
from
10ad334 to
5f9ad6c
Compare
| prompt_override_configuration: Optional[Dict] | ||
| inline_session_state: Optional[Dict] | ||
|
|
||
| class BedrockInlineAgentsRunnable(RunnableSerializable[Dict, OutputType]): |
There was a problem hiding this comment.
The contract seems to deviate from how other Runnables are implemented. Notice here in the Anthropic Chat Model that they use message as input and output: https://github.com/langchain-ai/langchain/blob/master/libs/core/langchain_core/language_models/chat_models.py#L127
If we support message in and message out then this will easily plugin to LCEL, LG and other constructs they have.
| get_boto_session, | ||
| ) | ||
|
|
||
| class BedrockInlineAgentsChatModel(BaseChatModel): |
There was a problem hiding this comment.
Let's not have both BedrockInlineAgentRunnable and BedrockInlineAgentsChatModel. Remove the state based BedrockInlineAgentRunnable implementation and rename BedrockInlineAgentsChatModel to BedrockInlineAgentRunnable. I like how the new implementation supports the standardized message interface and will integrate with LangGraph better.
Please update the inline_agent_langgraph notebook to use this implementation instead of the current state based approach. I would like use to completely remove "should_continue" just like here in this sample they have no should_continue: https://langchain-ai.github.io/langgraph/how-tos/create-react-agent/#usage
There was a problem hiding this comment.
Thanks, merged the two interfaces and added a langgraph integration notebook.
We'd still need the custom parsers for running the graph because of the way RoC inputs are passed to the agent. We can investigate more on how to handle this natively within the runnable itself in next iterations.
Description
This PR introduces a new Bedrock Inline-Agents Runnable that allows using InvokeInlineAgent API with return of control functions as tools and configurable agentic properties.
Usage
See Jupyter notebook for usage with
AgentExecutorCloses #309