Agents planning #31702
Agents planning #31702
Conversation
Fix infinite loops in python interpreter
1dce68f to
75d1129
Compare
75d1129 to
8efac17
Compare
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
e00836b to
50cad81
Compare
6dabac6 to
cbba80d
Compare
| **BASE_PYTHON_TOOLS.copy(), | ||
| **self.toolbox.tools, | ||
| }, | ||
| custom_tools=self.custom_tools, |
There was a problem hiding this comment.
This change is made because I needed to separate two sets of tools:
- user-defined tools that are static (at least they do not change within one call to
run) but can be changed in-betweenruncalls - agent-defined tools that will have to change within one run (thus the different handling in python interpreter)
| @@ -365,7 +365,118 @@ def download_prompt(prompt_or_repo_id, agent_name, mode="run"): | |||
| 6. Don't name any new variable with the same name as a tool: for instance don't name a variable 'final_answer'. | |||
| 7. Never create any notional variables in our code, as having these in your logs might derail you from the true variables. | |||
| 8. You can use imports in your code, but only from the following list of modules: <<authorized_imports>> | |||
| 9. Don't give up! You're in charge of solving the task, not providing directions to solve it. | |||
| 9. The state persists between code executions: so if in one step you've created variables or imported modules, these will all persist. | |||
There was a problem hiding this comment.
This precision helps e.g. with matplotlib plots that persist between code snippets.
| @@ -75,8 +85,8 @@ def get_iterable(obj): | |||
| raise InterpreterError("Object is not iterable") | |||
|
|
|||
|
|
|||
| def evaluate_unaryop(expression, state, tools): | |||
| operand = evaluate_ast(expression.operand, state, tools) | |||
| def evaluate_unaryop(expression, state, static_tools, custom_tools): | |||
There was a problem hiding this comment.
We add the distinction between static_tools and custom_tools everywhere: static_tools are the ones that the interpreter is not allowed to overwrite or create new ones (like base Python functions or user-defined tools), whereas custom_tools are the ones it can modify or extend.
There was a problem hiding this comment.
Ok, understood! Can you clarify that in the methods' docstrings?
| ops = [type(op) for op in condition.ops] | ||
|
|
||
| result = True | ||
| current_left = left | ||
|
|
||
| for op, comparator in zip(ops, comparators): | ||
| if op == ast.Eq: | ||
| result = result and (current_left == comparator) |
There was a problem hiding this comment.
This change allows to support pandas "==" statement between a Series and an int for instance
ef0c543 to
7263052
Compare
| f""" | ||
| The code blob you used is invalid: due to the following error: {e} | ||
| This means that the regex pattern {pattern} was not respected: make sure to include code with the correct pattern, for instance: | ||
| Thoughts: Your thoughts |
There was a problem hiding this comment.
We remind the correct pattern to guide the LLM towards correct formatting.
| @@ -655,7 +687,7 @@ def provide_final_answer(self, task) -> str: | |||
| except Exception as e: | |||
| return f"Error in generating final llm output: {e}." | |||
|
|
|||
| def run(self, task: str, stream: bool = False, **kwargs): | |||
| def run(self, task: str, stream: bool = False, reset: bool = True, **kwargs): | |||
There was a problem hiding this comment.
This reset option is provisory, it will be refined later to help multi-turn conversation with an user.
src/transformers/agents/agents.py
Outdated
| @@ -726,6 +764,91 @@ def direct_run(self, task: str, **kwargs): | |||
|
|
|||
| return final_answer | |||
|
|
|||
| def planning_step(self, task, is_first_step=False, iteration: int = None): | |||
There was a problem hiding this comment.
We tried to keep this planning step as simple as possible for now.
There was a problem hiding this comment.
No problem for me to merge this!
As the agents framework grows, it would be nice to complete the docstrings to hve each method be understood better; for example the methods planning_step, provide_final_answer, direct_run, stream_run, etc, are they supposed to be run by the user or by the Agent?
Clarifying this via their docstring would help a lot regarding the accessibility of the framework
| @@ -75,8 +85,8 @@ def get_iterable(obj): | |||
| raise InterpreterError("Object is not iterable") | |||
|
|
|||
|
|
|||
| def evaluate_unaryop(expression, state, tools): | |||
| operand = evaluate_ast(expression.operand, state, tools) | |||
| def evaluate_unaryop(expression, state, static_tools, custom_tools): | |||
There was a problem hiding this comment.
Ok, understood! Can you clarify that in the methods' docstrings?
| @@ -365,7 +365,118 @@ def download_prompt(prompt_or_repo_id, agent_name, mode="run"): | |||
| 6. Don't name any new variable with the same name as a tool: for instance don't name a variable 'final_answer'. | |||
| 7. Never create any notional variables in our code, as having these in your logs might derail you from the true variables. | |||
| 8. You can use imports in your code, but only from the following list of modules: <<authorized_imports>> | |||
| 9. Don't give up! You're in charge of solving the task, not providing directions to solve it. | |||
| 9. The state persists between code executions: so if in one step you've created variables or imported modules, these will all persist. | |||
What does this PR do?
This adds an optional planning steps to agents, to let them insert steps of higher-level planning to their workflow.
It also:
AgentTypebehavior where int and floats or other outputs would be converted toAgentType, which made them less accessible.