Refactor agents and tools (#91)
* Move tools to agents * Move agents to dedicate place * Remove subclassing BaseAgent from BaseTool
This commit is contained in:
Nguyen Trung Duc (john)
committed by
GitHub
GitHub
parent
4256030b4f
commit
8e3a1d193f
3
knowledgehub/agents/rewoo/__init__.py
Normal file
3
knowledgehub/agents/rewoo/__init__.py
Normal file
@@ -0,0 +1,3 @@
|
||||
from .agent import RewooAgent
|
||||
|
||||
__all__ = ["RewooAgent"]
|
||||
279
knowledgehub/agents/rewoo/agent.py
Normal file
279
knowledgehub/agents/rewoo/agent.py
Normal file
@@ -0,0 +1,279 @@
|
||||
import logging
|
||||
import re
|
||||
from concurrent.futures import ThreadPoolExecutor
|
||||
from typing import Any
|
||||
|
||||
from theflow import Param
|
||||
|
||||
from kotaemon.base.schema import Document
|
||||
from kotaemon.llms import LLM, ChatLLM, PromptTemplate
|
||||
from kotaemon.pipelines.citation import CitationPipeline
|
||||
|
||||
from ..base import AgentType, BaseAgent, BaseLLM, BaseTool
|
||||
from ..output.base import BaseScratchPad
|
||||
from ..utils import get_plugin_response_content
|
||||
from .planner import Planner
|
||||
from .solver import Solver
|
||||
|
||||
|
||||
class RewooAgent(BaseAgent):
|
||||
"""Distributive RewooAgent class inherited from BaseAgent.
|
||||
Implementing ReWOO paradigm https://arxiv.org/pdf/2305.18323.pdf"""
|
||||
|
||||
name: str = "RewooAgent"
|
||||
agent_type: AgentType = AgentType.rewoo
|
||||
description: str = "RewooAgent for answering multi-step reasoning questions"
|
||||
llm: BaseLLM | dict[str, BaseLLM] # {"Planner": xxx, "Solver": xxx}
|
||||
prompt_template: dict[str, PromptTemplate] = Param(
|
||||
default_callback=lambda _: {},
|
||||
help="A dict to supply different prompt to the agent.",
|
||||
)
|
||||
plugins: list[BaseTool] = Param(
|
||||
default_callback=lambda _: [], help="A list of plugins to be used in the model."
|
||||
)
|
||||
examples: dict[str, str | list[str]] = Param(
|
||||
default_callback=lambda _: {}, help="Examples to be used in the agent."
|
||||
)
|
||||
|
||||
def _get_llms(self):
|
||||
if isinstance(self.llm, ChatLLM) or isinstance(self.llm, LLM):
|
||||
return {"Planner": self.llm, "Solver": self.llm}
|
||||
elif (
|
||||
isinstance(self.llm, dict)
|
||||
and "Planner" in self.llm
|
||||
and "Solver" in self.llm
|
||||
):
|
||||
return {"Planner": self.llm["Planner"], "Solver": self.llm["Solver"]}
|
||||
else:
|
||||
raise ValueError("llm must be a BaseLLM or a dict with Planner and Solver.")
|
||||
|
||||
def _parse_plan_map(
|
||||
self, planner_response: str
|
||||
) -> tuple[dict[str, list[str]], dict[str, str]]:
|
||||
"""
|
||||
Parse planner output. It should be an n-to-n mapping from Plans to #Es.
|
||||
This is because sometimes LLM cannot follow the strict output format.
|
||||
Example:
|
||||
#Plan1
|
||||
#E1
|
||||
#E2
|
||||
should result in: {"#Plan1": ["#E1", "#E2"]}
|
||||
Or:
|
||||
#Plan1
|
||||
#Plan2
|
||||
#E1
|
||||
should result in: {"#Plan1": [], "#Plan2": ["#E1"]}
|
||||
This function should also return a plan map.
|
||||
|
||||
Returns:
|
||||
tuple[Dict[str, List[str]], Dict[str, str]]: A list of plan map
|
||||
"""
|
||||
valid_chunk = [
|
||||
line
|
||||
for line in planner_response.splitlines()
|
||||
if line.startswith("#Plan") or line.startswith("#E")
|
||||
]
|
||||
|
||||
plan_to_es: dict[str, list[str]] = dict()
|
||||
plans: dict[str, str] = dict()
|
||||
for line in valid_chunk:
|
||||
if line.startswith("#Plan"):
|
||||
plan = line.split(":", 1)[0].strip()
|
||||
plans[plan] = line.split(":", 1)[1].strip()
|
||||
plan_to_es[plan] = []
|
||||
elif line.startswith("#E"):
|
||||
plan_to_es[plan].append(line.split(":", 1)[0].strip())
|
||||
|
||||
return plan_to_es, plans
|
||||
|
||||
def _parse_planner_evidences(
|
||||
self, planner_response: str
|
||||
) -> tuple[dict[str, str], list[list[str]]]:
|
||||
"""
|
||||
Parse planner output. This should return a mapping from #E to tool call.
|
||||
It should also identify the level of each #E in dependency map.
|
||||
Example:
|
||||
{
|
||||
"#E1": "Tool1", "#E2": "Tool2",
|
||||
"#E3": "Tool3", "#E4": "Tool4"
|
||||
}, [[#E1, #E2], [#E3, #E4]]
|
||||
|
||||
Returns:
|
||||
tuple[dict[str, str], List[List[str]]]:
|
||||
A mapping from #E to tool call and a list of levels.
|
||||
"""
|
||||
evidences: dict[str, str] = dict()
|
||||
dependence: dict[str, list[str]] = dict()
|
||||
for line in planner_response.splitlines():
|
||||
if line.startswith("#E") and line[2].isdigit():
|
||||
e, tool_call = line.split(":", 1)
|
||||
e, tool_call = e.strip(), tool_call.strip()
|
||||
if len(e) == 3:
|
||||
dependence[e] = []
|
||||
evidences[e] = tool_call
|
||||
for var in re.findall(r"#E\d+", tool_call):
|
||||
if var in evidences:
|
||||
dependence[e].append(var)
|
||||
else:
|
||||
evidences[e] = "No evidence found"
|
||||
level = []
|
||||
while dependence:
|
||||
select = [i for i in dependence if not dependence[i]]
|
||||
if len(select) == 0:
|
||||
raise ValueError("Circular dependency detected.")
|
||||
level.append(select)
|
||||
for item in select:
|
||||
dependence.pop(item)
|
||||
for item in dependence:
|
||||
for i in select:
|
||||
if i in dependence[item]:
|
||||
dependence[item].remove(i)
|
||||
|
||||
return evidences, level
|
||||
|
||||
def _run_plugin(
|
||||
self,
|
||||
e: str,
|
||||
planner_evidences: dict[str, str],
|
||||
worker_evidences: dict[str, str],
|
||||
output=BaseScratchPad(),
|
||||
):
|
||||
"""
|
||||
Run a plugin for a given evidence.
|
||||
This function should also cumulate the cost and tokens.
|
||||
"""
|
||||
result = dict(e=e, plugin_cost=0, plugin_token=0, evidence="")
|
||||
tool_call = planner_evidences[e]
|
||||
if "[" not in tool_call:
|
||||
result["evidence"] = tool_call
|
||||
else:
|
||||
tool, tool_input = tool_call.split("[", 1)
|
||||
tool_input = tool_input[:-1]
|
||||
# find variables in input and replace with previous evidences
|
||||
for var in re.findall(r"#E\d+", tool_input):
|
||||
if var in worker_evidences:
|
||||
tool_input = tool_input.replace(var, worker_evidences.get(var, ""))
|
||||
try:
|
||||
selected_plugin = self._find_plugin(tool)
|
||||
if selected_plugin is None:
|
||||
raise ValueError("Invalid plugin detected")
|
||||
tool_response = selected_plugin(tool_input)
|
||||
result["evidence"] = get_plugin_response_content(tool_response)
|
||||
except ValueError:
|
||||
result["evidence"] = "No evidence found."
|
||||
finally:
|
||||
output.panel_print(
|
||||
result["evidence"], f"[green] Function Response of [blue]{tool}: "
|
||||
)
|
||||
return result
|
||||
|
||||
def _get_worker_evidence(
|
||||
self,
|
||||
planner_evidences: dict[str, str],
|
||||
evidences_level: list[list[str]],
|
||||
output=BaseScratchPad(),
|
||||
) -> Any:
|
||||
"""
|
||||
Parallel execution of plugins in DAG for speedup.
|
||||
This is one of core benefits of ReWOO agents.
|
||||
|
||||
Args:
|
||||
planner_evidences: A mapping from #E to tool call.
|
||||
evidences_level: A list of levels of evidences.
|
||||
Calculated from DAG of plugin calls.
|
||||
output: Output object, defaults to BaseOutput().
|
||||
Returns:
|
||||
A mapping from #E to tool call.
|
||||
"""
|
||||
worker_evidences: dict[str, str] = dict()
|
||||
plugin_cost, plugin_token = 0.0, 0.0
|
||||
with ThreadPoolExecutor() as pool:
|
||||
for level in evidences_level:
|
||||
results = []
|
||||
for e in level:
|
||||
results.append(
|
||||
pool.submit(
|
||||
self._run_plugin,
|
||||
e,
|
||||
planner_evidences,
|
||||
worker_evidences,
|
||||
output,
|
||||
)
|
||||
)
|
||||
if len(results) > 1:
|
||||
output.update_status(f"Running tasks {level} in parallel.")
|
||||
else:
|
||||
output.update_status(f"Running task {level[0]}.")
|
||||
for r in results:
|
||||
resp = r.result()
|
||||
plugin_cost += resp["plugin_cost"]
|
||||
plugin_token += resp["plugin_token"]
|
||||
worker_evidences[resp["e"]] = resp["evidence"]
|
||||
output.done()
|
||||
|
||||
return worker_evidences, plugin_cost, plugin_token
|
||||
|
||||
def _find_plugin(self, name: str):
|
||||
for p in self.plugins:
|
||||
if p.name == name:
|
||||
return p
|
||||
|
||||
def run(self, instruction: str, use_citation: bool = False) -> Document:
|
||||
"""
|
||||
Run the agent with a given instruction.
|
||||
"""
|
||||
logging.info(f"Running {self.name} with instruction: {instruction}")
|
||||
total_cost = 0.0
|
||||
total_token = 0
|
||||
|
||||
planner_llm = self._get_llms()["Planner"]
|
||||
solver_llm = self._get_llms()["Solver"]
|
||||
|
||||
planner = Planner(
|
||||
model=planner_llm,
|
||||
plugins=self.plugins,
|
||||
prompt_template=self.prompt_template.get("Planner", None),
|
||||
examples=self.examples.get("Planner", None),
|
||||
)
|
||||
solver = Solver(
|
||||
model=solver_llm,
|
||||
prompt_template=self.prompt_template.get("Solver", None),
|
||||
examples=self.examples.get("Solver", None),
|
||||
)
|
||||
|
||||
# Plan
|
||||
planner_output = planner(instruction)
|
||||
plannner_text_output = planner_output.text
|
||||
plan_to_es, plans = self._parse_plan_map(plannner_text_output)
|
||||
planner_evidences, evidence_level = self._parse_planner_evidences(
|
||||
plannner_text_output
|
||||
)
|
||||
|
||||
# Work
|
||||
worker_evidences, plugin_cost, plugin_token = self._get_worker_evidence(
|
||||
planner_evidences, evidence_level
|
||||
)
|
||||
worker_log = ""
|
||||
for plan in plan_to_es:
|
||||
worker_log += f"{plan}: {plans[plan]}\n"
|
||||
for e in plan_to_es[plan]:
|
||||
worker_log += f"{e}: {worker_evidences[e]}\n"
|
||||
|
||||
# Solve
|
||||
solver_output = solver(instruction, worker_log)
|
||||
solver_output_text = solver_output.text
|
||||
if use_citation:
|
||||
citation_pipeline = CitationPipeline(llm=solver_llm)
|
||||
citation = citation_pipeline(context=worker_log, question=instruction)
|
||||
else:
|
||||
citation = None
|
||||
|
||||
return Document(
|
||||
text=solver_output_text,
|
||||
metadata={
|
||||
"agent": "react",
|
||||
"cost": total_cost,
|
||||
"usage": total_token,
|
||||
"citation": citation,
|
||||
},
|
||||
)
|
||||
83
knowledgehub/agents/rewoo/planner.py
Normal file
83
knowledgehub/agents/rewoo/planner.py
Normal file
@@ -0,0 +1,83 @@
|
||||
from typing import Any, List, Optional, Union
|
||||
|
||||
from kotaemon.base import BaseComponent
|
||||
from kotaemon.llms import PromptTemplate
|
||||
|
||||
from ..base import BaseLLM, BaseTool
|
||||
from ..output.base import BaseScratchPad
|
||||
from .prompt import few_shot_planner_prompt, zero_shot_planner_prompt
|
||||
|
||||
|
||||
class Planner(BaseComponent):
|
||||
model: BaseLLM
|
||||
prompt_template: Optional[PromptTemplate] = None
|
||||
examples: Optional[Union[str, List[str]]] = None
|
||||
plugins: List[BaseTool]
|
||||
|
||||
def _compose_worker_description(self) -> str:
|
||||
"""
|
||||
Compose the worker prompt from the workers.
|
||||
|
||||
Example:
|
||||
toolname1[input]: tool1 description
|
||||
toolname2[input]: tool2 description
|
||||
"""
|
||||
prompt = ""
|
||||
try:
|
||||
for worker in self.plugins:
|
||||
prompt += f"{worker.name}[input]: {worker.description}\n"
|
||||
except Exception:
|
||||
raise ValueError("Worker must have a name and description.")
|
||||
return prompt
|
||||
|
||||
def _compose_fewshot_prompt(self) -> str:
|
||||
if self.examples is None:
|
||||
return ""
|
||||
if isinstance(self.examples, str):
|
||||
return self.examples
|
||||
else:
|
||||
return "\n\n".join([e.strip("\n") for e in self.examples])
|
||||
|
||||
def _compose_prompt(self, instruction) -> str:
|
||||
"""
|
||||
Compose the prompt from template, worker description, examples and instruction.
|
||||
"""
|
||||
worker_desctription = self._compose_worker_description()
|
||||
fewshot = self._compose_fewshot_prompt()
|
||||
if self.prompt_template is not None:
|
||||
if "fewshot" in self.prompt_template.placeholders:
|
||||
return self.prompt_template.populate(
|
||||
tool_description=worker_desctription,
|
||||
fewshot=fewshot,
|
||||
task=instruction,
|
||||
)
|
||||
else:
|
||||
return self.prompt_template.populate(
|
||||
tool_description=worker_desctription, task=instruction
|
||||
)
|
||||
else:
|
||||
if self.examples is not None:
|
||||
return few_shot_planner_prompt.populate(
|
||||
tool_description=worker_desctription,
|
||||
fewshot=fewshot,
|
||||
task=instruction,
|
||||
)
|
||||
else:
|
||||
return zero_shot_planner_prompt.populate(
|
||||
tool_description=worker_desctription, task=instruction
|
||||
)
|
||||
|
||||
def run(self, instruction: str, output: BaseScratchPad = BaseScratchPad()) -> Any:
|
||||
response = None
|
||||
output.info("Running Planner")
|
||||
prompt = self._compose_prompt(instruction)
|
||||
output.debug(f"Prompt: {prompt}")
|
||||
try:
|
||||
response = self.model(prompt)
|
||||
self.log_progress(".planner", response=response)
|
||||
output.info("Planner run successful.")
|
||||
except ValueError as e:
|
||||
output.error("Planner failed to retrieve response from LLM")
|
||||
raise ValueError("Planner failed to retrieve response from LLM") from e
|
||||
|
||||
return response
|
||||
119
knowledgehub/agents/rewoo/prompt.py
Normal file
119
knowledgehub/agents/rewoo/prompt.py
Normal file
@@ -0,0 +1,119 @@
|
||||
# flake8: noqa
|
||||
|
||||
from kotaemon.llms import PromptTemplate
|
||||
|
||||
zero_shot_planner_prompt = PromptTemplate(
|
||||
template="""You are an AI agent who makes step-by-step plans to solve a problem under the help of external tools.
|
||||
For each step, make one plan followed by one tool-call, which will be executed later to retrieve evidence for that step.
|
||||
You should store each evidence into a distinct variable #E1, #E2, #E3 ... that can be referred to in later tool-call inputs.
|
||||
|
||||
##Available Tools##
|
||||
{tool_description}
|
||||
|
||||
##Output Format (Replace '<...>')##
|
||||
#Plan1: <describe your plan here>
|
||||
#E1: <toolname>[<input here>] (eg. Search[What is Python])
|
||||
#Plan2: <describe next plan>
|
||||
#E2: <toolname>[<input here, you can use #E1 to represent its expected output>]
|
||||
And so on...
|
||||
|
||||
##Your Task##
|
||||
{task}
|
||||
|
||||
##Now Begin##
|
||||
"""
|
||||
)
|
||||
|
||||
one_shot_planner_prompt = PromptTemplate(
|
||||
template="""You are an AI agent who makes step-by-step plans to solve a problem under the help of external tools.
|
||||
For each step, make one plan followed by one tool-call, which will be executed later to retrieve evidence for that step.
|
||||
You should store each evidence into a distinct variable #E1, #E2, #E3 ... that can be referred to in later tool-call inputs.
|
||||
|
||||
##Available Tools##
|
||||
{tool_description}
|
||||
|
||||
##Output Format##
|
||||
#Plan1: <describe your plan here>
|
||||
#E1: <toolname>[<input here>]
|
||||
#Plan2: <describe next plan>
|
||||
#E2: <toolname>[<input here, you can use #E1 to represent its expected output>]
|
||||
And so on...
|
||||
|
||||
##Example##
|
||||
Task: What is the 4th root of 64 to the power of 3?
|
||||
#Plan1: Find the 4th root of 64
|
||||
#E1: Calculator[64^(1/4)]
|
||||
#Plan2: Raise the result from #Plan1 to the power of 3
|
||||
#E2: Calculator[#E1^3]
|
||||
|
||||
##Your Task##
|
||||
{task}
|
||||
|
||||
##Now Begin##
|
||||
"""
|
||||
)
|
||||
|
||||
|
||||
few_shot_planner_prompt = PromptTemplate(
|
||||
template="""You are an AI agent who makes step-by-step plans to solve a problem under the help of external tools.
|
||||
For each step, make one plan followed by one tool-call, which will be executed later to retrieve evidence for that step.
|
||||
You should store each evidence into a distinct variable #E1, #E2, #E3 ... that can be referred to in later tool-call inputs.
|
||||
|
||||
##Available Tools##
|
||||
{tool_description}
|
||||
|
||||
##Output Format (Replace '<...>')##
|
||||
#Plan1: <describe your plan here>
|
||||
#E1: <toolname>[<input>]
|
||||
#Plan2: <describe next plan>
|
||||
#E2: <toolname>[<input, you can use #E1 to represent its expected output>]
|
||||
And so on...
|
||||
|
||||
##Examples##
|
||||
{fewshot}
|
||||
|
||||
##Your Task##
|
||||
{task}
|
||||
|
||||
##Now Begin##
|
||||
"""
|
||||
)
|
||||
|
||||
zero_shot_solver_prompt = PromptTemplate(
|
||||
template="""You are an AI agent who solves a problem with my assistance. I will provide step-by-step plans(#Plan) and evidences(#E) that could be helpful.
|
||||
Your task is to briefly summarize each step, then make a short final conclusion for your task.
|
||||
|
||||
##My Plans and Evidences##
|
||||
{plan_evidence}
|
||||
|
||||
##Example Output##
|
||||
First, I <did something> , and I think <...>; Second, I <...>, and I think <...>; ....
|
||||
So, <your conclusion>.
|
||||
|
||||
##Your Task##
|
||||
{task}
|
||||
|
||||
##Now Begin##
|
||||
"""
|
||||
)
|
||||
|
||||
few_shot_solver_prompt = PromptTemplate(
|
||||
template="""You are an AI agent who solves a problem with my assistance. I will provide step-by-step plans and evidences that could be helpful.
|
||||
Your task is to briefly summarize each step, then make a short final conclusion for your task.
|
||||
|
||||
##My Plans and Evidences##
|
||||
{plan_evidence}
|
||||
|
||||
##Example Output##
|
||||
First, I <did something> , and I think <...>; Second, I <...>, and I think <...>; ....
|
||||
So, <your conclusion>.
|
||||
|
||||
##Example##
|
||||
{fewshot}
|
||||
|
||||
##Your Task##
|
||||
{task}
|
||||
|
||||
##Now Begin##
|
||||
"""
|
||||
)
|
||||
66
knowledgehub/agents/rewoo/solver.py
Normal file
66
knowledgehub/agents/rewoo/solver.py
Normal file
@@ -0,0 +1,66 @@
|
||||
from typing import Any, List, Optional, Union
|
||||
|
||||
from kotaemon.base import BaseComponent
|
||||
from kotaemon.llms import PromptTemplate
|
||||
|
||||
from ..base import BaseLLM
|
||||
from ..output.base import BaseScratchPad
|
||||
from .prompt import few_shot_solver_prompt, zero_shot_solver_prompt
|
||||
|
||||
|
||||
class Solver(BaseComponent):
|
||||
model: BaseLLM
|
||||
prompt_template: Optional[PromptTemplate] = None
|
||||
examples: Optional[Union[str, List[str]]] = None
|
||||
|
||||
def _compose_fewshot_prompt(self) -> str:
|
||||
if self.examples is None:
|
||||
return ""
|
||||
if isinstance(self.examples, str):
|
||||
return self.examples
|
||||
else:
|
||||
return "\n\n".join([e.strip("\n") for e in self.examples])
|
||||
|
||||
def _compose_prompt(self, instruction, plan_evidence) -> str:
|
||||
"""
|
||||
Compose the prompt from template, plan&evidence, examples and instruction.
|
||||
"""
|
||||
fewshot = self._compose_fewshot_prompt()
|
||||
if self.prompt_template is not None:
|
||||
if "fewshot" in self.prompt_template.placeholders:
|
||||
return self.prompt_template.populate(
|
||||
plan_evidence=plan_evidence, fewshot=fewshot, task=instruction
|
||||
)
|
||||
else:
|
||||
return self.prompt_template.populate(
|
||||
plan_evidence=plan_evidence, task=instruction
|
||||
)
|
||||
else:
|
||||
if self.examples is not None:
|
||||
return few_shot_solver_prompt.populate(
|
||||
plan_evidence=plan_evidence, fewshot=fewshot, task=instruction
|
||||
)
|
||||
else:
|
||||
return zero_shot_solver_prompt.populate(
|
||||
plan_evidence=plan_evidence, task=instruction
|
||||
)
|
||||
|
||||
def run(
|
||||
self,
|
||||
instruction: str,
|
||||
plan_evidence: str,
|
||||
output: BaseScratchPad = BaseScratchPad(),
|
||||
) -> Any:
|
||||
response = None
|
||||
output.info("Running Solver")
|
||||
output.debug(f"Instruction: {instruction}")
|
||||
output.debug(f"Plan Evidence: {plan_evidence}")
|
||||
prompt = self._compose_prompt(instruction, plan_evidence)
|
||||
output.debug(f"Prompt: {prompt}")
|
||||
try:
|
||||
response = self.model(prompt)
|
||||
output.info("Solver run successful.")
|
||||
except ValueError:
|
||||
output.error("Solver failed to retrieve response from LLM")
|
||||
|
||||
return response
|
||||
Reference in New Issue
Block a user