github/langchain
1
0
Fork 0
mirror of https://github.com/hwchase17/langchain.git synced 2025-06-24 07:35:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

agent refactors (#997)

Browse Source
This commit is contained in:
Harrison Chase 2023-02-12 23:02:13 -08:00 committed by GitHub
parent 7fb33fca47
commit 0f0e69adce
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 59 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
langchain/agents/agent.py
View File

@ -375,35 +375,22 @@ class AgentExecutor(Chain, BaseModel):
final_output["intermediate_steps"] = intermediate_steps final_output["intermediate_steps"] = intermediate_steps
return final_output return final_output
def _call(self, inputs: Dict[str, str]) -> Dict[str, Any]: def _take_next_step(
"""Run text through and get agent response.""" self,
# Make sure that every tool is synchronous (not a coroutine) name_to_tool_map: Dict[str, Tool],
for tool in self.tools: color_mapping: Dict[str, str],
if asyncio.iscoroutinefunction(tool.func): inputs: Dict[str, str],
raise ValueError( intermediate_steps: List[Tuple[AgentAction, str]],
"Tools cannot be asynchronous for `run` method. " ) -> Union[AgentFinish, Tuple[AgentAction, str]]:
"Please use `arun` instead." """Take a single step in the thought-action-observation loop.
)
# Do any preparation necessary when receiving a new input. Override this to take control of how the agent makes and acts on choices.
self.agent.prepare_for_new_call() """
# Construct a mapping of tool name to tool for easy lookup
name_to_tool_map = {tool.name: tool for tool in self.tools}
# We construct a mapping from each tool to a color, used for logging.
color_mapping = get_color_mapping(
[tool.name for tool in self.tools], excluded_colors=["green"]
)
intermediate_steps: List[Tuple[AgentAction, str]] = []
# Let's start tracking the iterations the agent has gone through
iterations = 0
# We now enter the agent loop (until it returns something).
while self._should_continue(iterations):
# Call the LLM to see what to do. # Call the LLM to see what to do.
output = self.agent.plan(intermediate_steps, **inputs) output = self.agent.plan(intermediate_steps, **inputs)
# If the tool chosen is the finishing tool, then we end and return. # If the tool chosen is the finishing tool, then we end and return.
if isinstance(output, AgentFinish): if isinstance(output, AgentFinish):
return self._return(output, intermediate_steps) return output
# Otherwise we lookup the tool # Otherwise we lookup the tool
if output.tool in name_to_tool_map: if output.tool in name_to_tool_map:
tool = name_to_tool_map[output.tool] tool = name_to_tool_map[output.tool]
@ -436,11 +423,41 @@ class AgentExecutor(Chain, BaseModel):
llm_prefix=llm_prefix, llm_prefix=llm_prefix,
verbose=self.verbose, verbose=self.verbose,
) )
intermediate_steps.append((output, observation))
if return_direct: if return_direct:
# Set the log to "" because we do not want to log it. # Set the log to "" because we do not want to log it.
output = AgentFinish({self.agent.return_values[0]: observation}, "") return AgentFinish({self.agent.return_values[0]: observation}, "")
return self._return(output, intermediate_steps) return output, observation
def _call(self, inputs: Dict[str, str]) -> Dict[str, Any]:
"""Run text through and get agent response."""
# Make sure that every tool is synchronous (not a coroutine)
for tool in self.tools:
if asyncio.iscoroutinefunction(tool.func):
raise ValueError(
"Tools cannot be asynchronous for `run` method. "
"Please use `arun` instead."
)
# Do any preparation necessary when receiving a new input.
self.agent.prepare_for_new_call()
# Construct a mapping of tool name to tool for easy lookup
name_to_tool_map = {tool.name: tool for tool in self.tools}
# We construct a mapping from each tool to a color, used for logging.
color_mapping = get_color_mapping(
[tool.name for tool in self.tools], excluded_colors=["green"]
)
intermediate_steps: List[Tuple[AgentAction, str]] = []
# Let's start tracking the iterations the agent has gone through
iterations = 0
# We now enter the agent loop (until it returns something).
while self._should_continue(iterations):
next_step_output = self._take_next_step(
name_to_tool_map, color_mapping, inputs, intermediate_steps
)
if isinstance(next_step_output, AgentFinish):
return self._return(next_step_output, intermediate_steps)
intermediate_steps.append(next_step_output)
iterations += 1 iterations += 1
output = self.agent.return_stopped_response( output = self.agent.return_stopped_response(
self.early_stopping_method, intermediate_steps, **inputs self.early_stopping_method, intermediate_steps, **inputs