Taskgen Module#

class taskgen.Agent(*args, **kwargs)#

Bases: BaseAgent

assign_agent(agent_list)#

Assigns a list of Agents to the main agent, passing in the meta agent as well

assign_agents(agent_list)#

Assigns a list of Agents to the main agent, passing in the meta agent as well

assign_function(function_list)#

Assigns a list of functions to be used in function_map

assign_functions(function_list)#

Assigns a list of functions to be used in function_map

assign_tool(function_list)#

Assigns a list of functions to be used in function_map

assign_tools(function_list)#

Assigns a list of functions to be used in function_map

contribute_agent(author_comments=None)#
Return type:

str

get_next_subtask(task='')#

Based on what the task is and the subtasks completed, we get the next subtask, function and input parameters. Supports user-given task as well if user wants to use this function directly

classmethod load_community_agent(agent_name)#
query(query, output_format, provide_function_list=False, task='')#

Queries the agent with a query and outputs in output_format. If task is provided, we will filter the functions according to the task If you want to provide the agent with the context of functions available to it, set provide_function_list to True (default: False) If task is given, then we will use it to do RAG over functions

reply_user(query='', stateful=True, verbose=True)#

Generate a reply to the user based on the query / agent task and subtasks completed If stateful, also store this interaction into the subtasks_completed If verbose is given, can also override the verbosity of this function

run(task='', overall_task='', num_subtasks=0)#

Attempts to do the task using LLM and available functions Loops through and performs either a function call or LLM call up to num_subtasks number of times If overall_task is filled, then we store it to pass to the inner agents for more context

Return type:

list

select_function(task='')#

Based on the task (without any context), output the next function name and input parameters

select_tool(task='')#

Based on the task (without any context), output the next function name and input parameters

summarise_subtasks_completed(task='')#

Summarise the subtasks_completed list according to task

to_function(meta_agent)#

Converts the agent to a function so that it can be called by another agent The agent will take in an instruction, and output the result after processing

use_agent(agent_name, agent_task)#

Uses an inner agent to do a task for the meta agent. Task outcome goes directly to subtasks_completed of meta agent

use_function(function_name, function_params, subtask='', stateful=True)#

Uses the function. stateful means we store the outcome of the function

use_tool(function_name, function_params, subtask='', stateful=True)#

Uses the function. stateful means we store the outcome of the function

class taskgen.AsyncAgent(*args, **kwargs)#

Bases: BaseAgent

assign_agent(agent_list)#

Assigns a list of Agents to the main agent, passing in the meta agent as well

assign_agents(agent_list)#

Assigns a list of Agents to the main agent, passing in the meta agent as well

assign_function(function_list)#

Assigns a list of functions to be used in function_map

assign_functions(function_list)#

Assigns a list of functions to be used in function_map

assign_tool(function_list)#

Assigns a list of functions to be used in function_map

assign_tools(function_list)#

Assigns a list of functions to be used in function_map

async get_next_subtask(task='')#

Based on what the task is and the subtasks completed, we get the next subtask, function and input parameters. Supports user-given task as well if user wants to use this function directly

async query(query, output_format, provide_function_list=False, task='')#

Queries the agent with a query and outputs in output_format. If task is provided, we will filter the functions according to the task If you want to provide the agent with the context of functions available to it, set provide_function_list to True (default: False) If task is given, then we will use it to do RAG over functions

async reply_user(query='', stateful=True, verbose=True)#

Generate a reply to the user based on the query / agent task and subtasks completed If stateful, also store this interaction into the subtasks_completed If verbose is given, can also override the verbosity of this function

async run(task='', overall_task='', num_subtasks=0)#

Attempts to do the task using LLM and available functions Loops through and performs either a function call or LLM call up to num_subtasks number of times If overall_task is filled, then we store it to pass to the inner agents for more context

Return type:

list

async select_function(task='')#

Based on the task (without any context), output the next function name and input parameters

async select_tool(task='')#

Based on the task (without any context), output the next function name and input parameters

async summarise_subtasks_completed(task='')#

Summarise the subtasks_completed list according to task

to_function(meta_agent)#

Converts the agent to a function so that it can be called by another agent The agent will take in an instruction, and output the result after processing

async use_agent(agent_name, agent_task)#

Uses an inner agent to do a task for the meta agent. Task outcome goes directly to subtasks_completed of meta agent

async use_function(function_name, function_params, subtask='', stateful=True)#

Uses the function. stateful means we store the outcome of the function

async use_tool(function_name, function_params, subtask='', stateful=True)#

Uses the function. stateful means we store the outcome of the function

class taskgen.AsyncChromaDbMemory(*args, **kwargs)#

Bases: BaseChromaDbMemory

async add_file(filepath, text_splitter=None)#
async append(new_memories=[], mapper=None)#

Appends multiple new memories

async append_memory_list(new_memories, metadatas=None)#
async create_embedding(text)#
get_openai_client()#
async get_or_create_collection()#
async remove(memories)#

Removes an existing_memory. existing_memory can be str, or triplet if it is a Knowledge Graph

async remove_by_id(ids=[])#
async reset()#

Clears all memories

async retrieve(task, filter=[])#

Retrieves some memories according to task

class taskgen.AsyncFunction(*args, **kwargs)#

Bases: BaseFunction

async async_init()#

This generates the name for the function using strict_json_async

class taskgen.AsyncMemory(*args, **kwargs)#

Bases: BaseMemory

Retrieves top k memory items based on task - Inputs:

  • memory: List. Default: Empty List. The list containing the memory items

  • top_k: Int. Default: 3. The number of memory list items to retrieve

  • mapper: Function. Maps the memory item to another form for comparison by ranker or LLM. Default: lambda x: x

    • Example mapping: lambda x: x.fn_description (If x is a Class and the string you want to compare for similarity is the fn_description attribute of that class)

  • approach: str. Either retrieve_by_ranker or retrieve_by_llm to retrieve memory items

    • Ranker is faster and cheaper as it compares via embeddings, but are inferior to LLM-based methods for contextual information

  • llm: Function. The llm to use for strict_json llm retriever

  • retrieve_fn: Default: None. Takes in task and outputs top_k similar memories in a list

  • ranker: Ranker. The Ranker which defines a similarity score between a query and a key. Default: OpenAI text-embedding-3-small model.

    • Can be replaced with a function which returns similarity score from 0 to 1 when given a query and key

async retrieve(task)#

Performs retrieval of top_k similar memories according to approach stated

Return type:

list

async retrieve_by_llm(task)#

Performs retrieval via LLMs Returns the key list as well as the value list

Return type:

list

async retrieve_by_ranker(task)#

Performs retrieval of top_k similar memories Returns the memory list items corresponding to top_k matches

Return type:

list

class taskgen.AsyncRanker(model='text-embedding-3-small', ranking_fn=None, database=None)#

Bases: BaseRanker

async get_or_create_embedding_async(text, client)#
class taskgen.ChromaDbMemory(*args, **kwargs)#

Bases: BaseChromaDbMemory

add_file(filepath, text_splitter=None)#
append(new_memories, mapper=None)#

Appends multiple new memories

append_memory_list(new_memories, metadatas=None)#
create_embedding(text)#
get_openai_client()#
get_or_create_collection()#
class taskgen.ConversableAgent(agent, persistent_memory=None, person='User', conversation=None, num_past_conversation=5, verbose=True)#

Bases: object

This class takes an Agent and allows for conversational-based interactions with User / another Agent / Environment. Also updates persistent memory with latest information in conversation

  • Inputs:

    • agent (compulsory): Agent. The agent we want to interact with

    • persistent_memory: dict. What kinds of memory the agent should have that persist over the entire conversation and their descriptions. Uses the same format as output_format of strict_json.

    • person: str. The name of the person you are talking to

    • conversation: List. The current existing conversation. Default: None

    • num_past_conversation: int. The number of past conversations to use for the agent

    • verbose: bool. Default: True. Whether to print the Agent’s inner states

  • ConversableAgent will automatically implement 3 new variables in agent.shared_variables:

    • Persistent Memory: The memory that will be updated as the conversation goes along, defined in persistent_dict

    • Conversation: The entire history of the conversationn

    • Summary of Conversation: A summary of the current conversation

  • ConversableAgent uses chat() which chats with the Agent and the Agent will perform actions and reply the chat message

chat(cur_msg)#

This does one chat with the person, firstly performing actions then replying the person, while updating the important memory

verbose#

Define some external variables for the Agent

class taskgen.ConversationWrapper(agent, persistent_memory=None, person='User', conversation=None, num_past_conversation=5, verbose=True)#

Bases: Agent

This class takes an Agent and allows for conversational-based interactions with User / another Agent / Environment. Also updates persistent memory with latest information in conversation

  • Inputs:

    • agent (compulsory): Agent. The agent we want to interact with

    • persistent_memory: dict. What kinds of memory the agent should have that persist over the entire conversation and their descriptions. Uses the same format as output_format of strict_json.

    • person: str. The name of the person you are talking to

    • conversation: List. The current existing conversation. Default: None

    • num_past_conversation: int. The number of past conversations to use for the agent

    • verbose: bool. Default: True. Whether to print the Agent’s inner states

  • ConversationWrapper will automatically implement 3 new variables in agent.shared_variables:

    • Persistent Memory: The memory that will be updated as the conversation goes along, defined in persistent_dict

    • Conversation: The entire history of the conversation

    • Summary of Conversation: A summary of the current conversation

  • ConversationWrapper uses chat() which chats with the Agent and the Agent will perform actions and reply the chat message

chat(cur_msg)#

This does one chat with the person, firstly performing actions then replying the person, while updating the important memory

verbose#

Define some external variables for the Agent

class taskgen.Function(*args, **kwargs)#

Bases: BaseFunction

class taskgen.Memory(*args, **kwargs)#

Bases: BaseMemory

Retrieves top k memory items based on task - Inputs:

  • memory: List. Default: Empty List. The list containing the memory items

  • top_k: Int. Default: 3. The number of memory list items to retrieve

  • mapper: Function. Maps the memory item to another form for comparison by ranker or LLM. Default: lambda x: x

    • Example mapping: lambda x: x.fn_description (If x is a Class and the string you want to compare for similarity is the fn_description attribute of that class)

  • approach: str. Either retrieve_by_ranker or retrieve_by_llm to retrieve memory items

    • Ranker is faster and cheaper as it compares via embeddings, but are inferior to LLM-based methods for contextual information

  • llm: Function. The llm to use for strict_json llm retriever

  • retrieve_fn: Default: None. Takes in task and outputs top_k similar memories in a list

  • ranker: Ranker. The Ranker which defines a similarity score between a query and a key. Default: OpenAI text-embedding-3-small model.

    • Can be replaced with a function which returns similarity score from 0 to 1 when given a query and key

retrieve(task)#

Performs retrieval of top_k similar memories according to approach stated

Return type:

list

retrieve_by_llm(task)#

Performs retrieval via LLMs Returns the key list as well as the value list

Return type:

list

retrieve_by_ranker(task)#

Performs retrieval of top_k similar memories Returns the memory list items corresponding to top_k matches

Return type:

list

class taskgen.MemoryTemplate#

Bases: ABC

A generic template provided for all memories

abstract append(memory_list, mapper=None)#

Appends multiple new memories

read_docx(filepath)#
read_file(filepath, text_splitter=None)#
read_pdf(filepath)#
abstract remove(existing_memory)#

Removes an existing_memory. existing_memory can be str, or triplet if it is a Knowledge Graph

abstract reset()#

Clears all memories

abstract retrieve(task)#

Retrieves some memories according to task

class taskgen.Ranker(model='text-embedding-3-small', ranking_fn=None, database=None)#

Bases: BaseRanker

get_or_create_embedding(text, client)#
taskgen.chat(system_prompt, user_prompt, model='gpt-4o-mini', temperature=0, verbose=False, host='openai', llm=None, **kwargs)#

Performs a chat with the host’s LLM model with system prompt, user prompt, model, verbose and kwargs Returns the output string res - system_prompt: String. Write in whatever you want the LLM to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - model: String. The LLM model to use for json generation - verbose: Boolean (default: False). Whether or not to print out the system prompt, user prompt, GPT response - host: String. The provider of the LLM - llm: User-made llm function.

  • Inputs:

    • system_prompt: String. Write in whatever you want the LLM to become. e.g. “You are a <purpose in life>”

    • user_prompt: String. The user input. Later, when we use it as a function, this is the function input

  • Output:

    • res: String. The response of the LLM call

  • **kwargs: Dict. Additional arguments for LLM chat

async taskgen.chat_async(system_prompt, user_prompt, model='gpt-4o-mini', temperature=0, verbose=False, host='openai', llm=None, **kwargs)#

Performs a chat with the host’s LLM model with system prompt, user prompt, model, verbose and kwargs Returns the output string res - system_prompt: String. Write in whatever you want the LLM to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - model: String. The LLM model to use for json generation - verbose: Boolean (default: False). Whether or not to print out the system prompt, user prompt, GPT response - host: String. The provider of the LLM - llm: User-made llm function.

  • Inputs:

    • system_prompt: String. Write in whatever you want the LLM to become. e.g. “You are a <purpose in life>”

    • user_prompt: String. The user input. Later, when we use it as a function, this is the function input

  • Output:

    • res: String. The response of the LLM call

  • **kwargs: Dict. Additional arguments for LLM chat

taskgen.strict_function#

alias of Function

taskgen.strict_json(system_prompt, user_prompt, output_format, return_as_json=False, custom_checks=None, check_data=None, delimiter='###', num_tries=3, openai_json_mode=False, **kwargs)#

Ensures that OpenAI will always adhere to the desired output JSON format defined in output_format. Uses rule-based iterative feedback to ask GPT to self-correct. Keeps trying up to num_tries it it does not. Returns empty JSON if unable to after num_tries iterations.

Inputs (compulsory): - system_prompt: String. Write in whatever you want GPT to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - output_format: Dict. JSON format with the key as the output key, and the value as the output description

Inputs (optional): - return_as_json: Bool. Default: False. Whether to return the output as a json. If False, returns as Python dict. If True, returns as json string - custom_checks: Dict. Key is output key, value is function which does checking of content for output field - check_data: Any data type. The additional data for custom_checks to use if required - delimiter: String (Default: ‘###’). This is the delimiter to surround the keys. With delimiter ###, key becomes ###key### - num_tries: Integer (default: 3). The number of tries to iteratively prompt GPT to generate correct json format - openai_json_mode: Boolean (default: False). Whether or not to use OpenAI JSON Mode - **kwargs: Dict. Additional arguments for LLM chat

Output: - res: Dict. The JSON output of the model. Returns {} if JSON parsing failed.

async taskgen.strict_json_async(system_prompt, user_prompt, output_format, return_as_json=False, custom_checks=None, check_data=None, delimiter='###', num_tries=3, openai_json_mode=False, **kwargs)#

Ensures that OpenAI will always adhere to the desired output JSON format defined in output_format. Uses rule-based iterative feedback to ask GPT to self-correct. Keeps trying up to num_tries it it does not. Returns empty JSON if unable to after num_tries iterations.

Inputs (compulsory): - system_prompt: String. Write in whatever you want GPT to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - output_format: Dict. JSON format with the key as the output key, and the value as the output description

Inputs (optional): - return_as_json: Bool. Default: False. Whether to return the output as a json. If False, returns as Python dict. If True, returns as json string - custom_checks: Dict. Key is output key, value is function which does checking of content for output field - check_data: Any data type. The additional data for custom_checks to use if required - delimiter: String (Default: ‘###’). This is the delimiter to surround the keys. With delimiter ###, key becomes ###key### - num_tries: Integer (default: 3). The number of tries to iteratively prompt GPT to generate correct json format - openai_json_mode: Boolean (default: False). Whether or not to use OpenAI JSON Mode - **kwargs: Dict. Additional arguments for LLM chat

Output: - res: Dict. The JSON output of the model. Returns {} if JSON parsing failed.

taskgen.strict_output(system_prompt, user_prompt, output_format, return_as_json=False, custom_checks=None, check_data=None, delimiter='###', num_tries=3, openai_json_mode=False, **kwargs)#

Ensures that OpenAI will always adhere to the desired output JSON format defined in output_format. Uses rule-based iterative feedback to ask GPT to self-correct. Keeps trying up to num_tries it it does not. Returns empty JSON if unable to after num_tries iterations.

Inputs (compulsory): - system_prompt: String. Write in whatever you want GPT to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - output_format: Dict. JSON format with the key as the output key, and the value as the output description

Inputs (optional): - return_as_json: Bool. Default: False. Whether to return the output as a json. If False, returns as Python dict. If True, returns as json string - custom_checks: Dict. Key is output key, value is function which does checking of content for output field - check_data: Any data type. The additional data for custom_checks to use if required - delimiter: String (Default: ‘###’). This is the delimiter to surround the keys. With delimiter ###, key becomes ###key### - num_tries: Integer (default: 3). The number of tries to iteratively prompt GPT to generate correct json format - openai_json_mode: Boolean (default: False). Whether or not to use OpenAI JSON Mode - **kwargs: Dict. Additional arguments for LLM chat

Output: - res: Dict. The JSON output of the model. Returns {} if JSON parsing failed.

taskgen.strict_text(system_prompt, user_prompt, output_format, return_as_json=False, custom_checks=None, check_data=None, delimiter='###', num_tries=3, openai_json_mode=False, **kwargs)#

Ensures that OpenAI will always adhere to the desired output JSON format defined in output_format. Uses rule-based iterative feedback to ask GPT to self-correct. Keeps trying up to num_tries it it does not. Returns empty JSON if unable to after num_tries iterations.

Inputs (compulsory): - system_prompt: String. Write in whatever you want GPT to become. e.g. “You are a <purpose in life>” - user_prompt: String. The user input. Later, when we use it as a function, this is the function input - output_format: Dict. JSON format with the key as the output key, and the value as the output description

Inputs (optional): - return_as_json: Bool. Default: False. Whether to return the output as a json. If False, returns as Python dict. If True, returns as json string - custom_checks: Dict. Key is output key, value is function which does checking of content for output field - check_data: Any data type. The additional data for custom_checks to use if required - delimiter: String (Default: ‘###’). This is the delimiter to surround the keys. With delimiter ###, key becomes ###key### - num_tries: Integer (default: 3). The number of tries to iteratively prompt GPT to generate correct json format - openai_json_mode: Boolean (default: False). Whether or not to use OpenAI JSON Mode - **kwargs: Dict. Additional arguments for LLM chat

Output: - res: Dict. The JSON output of the model. Returns {} if JSON parsing failed.

wrapper

base

On this page
Show Source