Table of Contents
- Structure of Coder
- CtrlExMem_CoderFlow
- run_coder
- Planner_CoderFlow
- Controller_CoderFlow
- UpdatePlanAtomicFlow
- CoderFlow
- __init__
Structure of Coder
goal, memory_files (dict)
|
v
+-------------------+
| MemoryReading | Reads in the content of the memory files
| Flow |
+-------------------+
|
| (memory_content)
|
v
+-------------------+
| PlanWriter | Writes a step-by-step plan to achieve the goal
+-------------------+
|
| (plan)
|
v
+-------------------+
| CtrlExMemFlow | Illustrated below. Carries out the plan in an controller-executor fashion,
| | with memory management mechanisms.
+-------------------+
|
(summary, result)
Here is the structure of the CtrlExMemFlow
:
plan, memory_files, memory_content, goal
|
v
+---------------+
| Controller | --------<<<<-----------+
+---------------+ |
| |
| (command, command args) |
| |
v |
+------------------+ |
| Executor | Each branch is an |
| (Tree Structure) | executor |
+------------------+ |
| ^
| (execution results) ^
| ^
v ^
+---------------+ ^
| MemWriteFlow | Updates memory files ^
+---------------+ ^
| ^
| (summary) |
| |
v |
+---------------+ |
| MemReadFlow | Reads updated memory |
+---------------+ |
| |
| (updated memory content) |
| |
+-> goes back to the Controller>-+
Structure of the Executors:
+-------------------+
| Branching |
| Executor |
+-------------------+
/ | | | \
/ | | | \
/ | | | \
/ | | | \
Extend_library ask_user re_plan update_plan run_code
Memory files of Coder:
- plan_coder.txt
- logs_coder.txt
- library.py
About the branches:
- ExtendLibrary: Writes and tests code functions in an interactive fashion, finally saves the written function to the code library.
- ask_user: Ask user for info / confirmation, etc.
- re_plan: One branch of the executors, when something goes wrong, re-draft the plan.
- update_plan: One branch of the executors, when the controller realizes that one (or some, depending on the LLM's response) step of the plan is (are) done, it generates a new plan that marks the step(s) as done.
- run_code: Runs the code written by the Controller, will first open up a temp file with the code for user confirmation and editing, then the code is passed to the InterpreterFlow.
CtrlExMem_CoderFlow
CtrlExMem_CoderFlow Objects
class CtrlExMem_CoderFlow(CtrlExMemFlow)
This class inherits from the CtrlExMemFlow class from AbstractBossFlowModule. See: https://huggingface.co/aiflows/AbstractBossFlowModule/blob/main/CtrlExMemFlow.py
Input Interface:
plan
logs
code_library
: the signatures and docstring of the functions in the code library.memory_files
goal
Output Interface
result
(str): The result of the flow, the result will be returned to the caller.summary
(str): The summary of the flow, the summary will be logged into the logs of the caller flow.
Configuration Parameters:
- See the configuration parameters of the CtrlExMemFlow class from AbstractBoss (https://huggingface.co/aiflows/AbstractBossFlowModule/blob/main/CtrlExMemFlow.py).
input_interface
: the input interface of the flowoutput_interface
: the output interface of the flowsubflows_config
: the configuration of the subflows
detect_finish_or_continue
@CircularFlow.output_msg_payload_processor
def detect_finish_or_continue(output_payload: Dict[str, Any],
src_flow) -> Dict[str, Any]
This method is called when the flow receives a message from a subflow.
This method will check the message and decide whether the flow should continue or finish.
Arguments:
output_payload
: the output payload of the subflowsrc_flow
: the source flow of the message
Returns:
Dict[str, Any]
: the output payload of the flow
run_coder
Planner_CoderFlow
Planner_CoderFlow Objects
class Planner_CoderFlow(PlanWriterFlow)
Planner of the coder flow, it inherits from PlanWriterFlow, see: https://huggingface.co/aiflows/PlanWriterFlowModule This flow is responsible for generating the plan for the coder flow.
Input Interface
goal
memory_files
code_library
Output Interface
plan
summary
status
Configuration Parameters:
- Look at the configuration parameters of PlanWriterFlowModule for detailed info. (https://huggingface.co/aiflows/PlanWriterFlowModule)
input_interface
: the input interface of the flow, default:["goal", "memory_files", "code_library"]
output_interface
: the output interface of the flow, default:["plan", "summary", "status"]
subflows_config
: configuration of the subflows.topology
: topology of the subflows.
Controller_CoderFlow
Controller_CoderFlow Objects
class Controller_CoderFlow(ChatAtomicFlow)
Refer to: https://huggingface.co/aiflows/JarvisFlowModule/blob/main/Controller_JarvisFlow.py This flow is used to control the coder flow.
Input Interface Non Initialized:
goal
plan
code_library
logs
memory_files
Input Interface Initialized:
goal
plan
code_library
logs
memory_files
result
Output Interface:
command
command_args
Configuration Parameters:
Input Interface Non Initialized
: Input interface before the conversation is initialized.Input Interface Initialized
: Input interface after the conversation is initialized.Output Interface
: Output interface.backend
: The backend of the LLM.command
: A list of available commands for the controller to call.system_message_prompt_template
: The template of the system message prompt.init_human_message_prompt_template
: The template of the initial human message prompt.human_message_prompt_template
: The template of the human message prompt.previous_messages
: The sliding window of previous messages.
__init__
def __init__(commands: List[Command], **kwargs)
Initialize the flow.
Arguments:
commands
(List[Command]
): A list of available commands for the controller to call.kwargs
(Dict[str, Any]
): Refer to the configuration parameters.
instantiate_from_config
@classmethod
def instantiate_from_config(cls, config)
Instantiate the flow from the configuration.
Arguments:
config
(Dict[str, Any]
): The configuration.
Returns:
Controller_CoderFlow
: The instantiated flow.
run
def run(input_data: Dict[str, Any]) -> Dict[str, Any]
Run the flow.
Arguments:
input_data
(Dict[str, Any]
): The input data.
Returns:
Dict[str, Any]
: The output data.
UpdatePlanAtomicFlow
UpdatePlanAtomicFlow Objects
class UpdatePlanAtomicFlow(AtomicFlow)
This flow updates the plan file with the updated plan. The controller should pass the updated plan to this flow. This design (controller reflect on the existing plan--update plan) is intended to let the controller more aware of the plan it has. However one setback is that this process in then not deterministic.
Input Interface
updated_plan
: the updated plan in string format
Output Interface
result
: the result of the update plan operation
Configuration Parameters:
input_interface
: the input interface of the flow, default:["updated_plan"]
output_interface
: the output interface of the flow, default:["result"]
run
def run(input_data: Dict[str, Any])
The run function of the flow.
Arguments:
input_data
(Dict[str, Any]
): the input data to the flow
Returns:
Dict[str, Any]
: the output data of the flow
CoderFlow
CoderFlow Objects
class CoderFlow(AbstractBossFlow)
Coder flow is one executor branch of the Jarvis flow. At a higher level, it is a flow that writes and runs code given a goal. In the Jarvis flow, the Coder flow in invoked by the controller, The Coder flow receives the goal generated by the controller, writes and runs code in an interactive fashion.
The Coder flow has the similar structure as the Jarvis flow (inherited from AbstractBossFlow).
Input Interface (expected input)
goal
(str): The goal from the caller (source flow, i.e. JarvisFlow)
Output Interface (expected output)
result
(str): The result of the flow, the result will be returned to the caller (i.e. JarvisFlow).summary
(str): The summary of the flow, the summary will be logged into the logs of the caller flow (i.e. JarvisFlow).
Configuration Parameters: (Also see super class: AbstractBossFlow)
memory_files
(dict): The memory files of the flow. The memory files are the files that the flow reads and writes. Typically it should contain plan, logs, and code_library.
Typical workflow of Coder: 0. JarvisFlow calls Coder with a goal.
- MemoryReading reads plans, logs and code library.
- Planner makes plan based on goal.
- Extend library with the goal given by the controller.
- Run code with code (possibly calls the newly written function) given by the controller.
- Finish and give an answer.
run
def run(input_data: Dict[str, Any]) -> Dict[str, Any]
The run function of the Coder flow.
Arguments:
input_data
(Dict[str, Any]
): The input data of the flow.
Returns:
Dict[str, Any]
: The output data of the flow.