Adding support for a new runnable for InlineAgents #340
Conversation
divekarshubham
force-pushed
the
inline_agents_support
branch
2 times, most recently
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.
| response_text = json.dumps(event) | ||
| break | ||
|
|
||
| if "chunk" in event: |
There was a problem hiding this comment.
Be sure to add the files to the output, see https://github.com/jdbaker01/langchain-aws/blob/main/libs/aws/langchain_aws/agents/base.py#L76 and https://github.com/jdbaker01/langchain-aws/blob/main/libs/aws/langchain_aws/agents/base.py#L82. Code Interpreter outputs files as well as the response.
| @@ -609,3 +311,219 @@ def _parse_intermediate_steps( | |||
| return session_id, session_state | |||
|
|
|||
| return None, None | |||
|
|
|||
| class BedrockInlineAgentsRunnable(BaseChatModel): | |||
There was a problem hiding this comment.
What was the motivation behind using the ChatModel as base rather than a Runnable/RunnableSerializable?
ChatModel have specific mechanisms, for example bind_tools to work with tools, that this implementation will bypass, so this might not be the right interface for the agents.
| session_id = response["sessionId"] | ||
| trace_log_elements = [] | ||
| files = [] | ||
| for event in event_stream: |
There was a problem hiding this comment.
Should we have some sort of signal for unknown event types so consumers can more easily identify service model changes?
There was a problem hiding this comment.
Currently this is not a type within the response, will investigate more and update this in a later PR
| ) | ||
| ] | ||
| except Exception as ex: | ||
| raise Exception("Parse exception encountered {}".format(repr(ex))) |
There was a problem hiding this comment.
Is there an exception hierarchy that we can add onto here? Exception is broad.
There was a problem hiding this comment.
Unsure of this one, keeping it broad to encompass all validation + serviceUnavailable exceptions.
divekarshubham
force-pushed
the
inline_agents_support
branch
from
39caa5e to
61ac9a0
Compare
|
@divekarshubham |
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