apunkt/MetaGPT
1
0
Fork 0
mirror of https://github.com/FoundationAgents/MetaGPT.git synced 2026-06-26 15:49:42 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

update code

Browse source
This commit is contained in:
stellahsr 2023-10-09 19:32:50 +08:00
parent f70ba27d35
commit 37a7a96bbe
13 changed files with 255 additions and 215 deletions
Download patch file Download diff file Expand all files Collapse all files

78
metagpt/actions/minecraft/design_curriculumn.py
View file

@ -21,11 +21,11 @@ class DesignTask(Action):
Action class for decomposing a task.
Refer to the code in the voyager/agents/curriculum.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.llm.model = "gpt-3.5-turbo"
async def decompose_task(self, query, events):
system_msgs = SystemMessage(
content=load_prompt("curriculum_task_decomposition")
@ -34,11 +34,11 @@ class DesignTask(Action):
events=events, chest_observation=""
) + HumanMessage(content=f"Final task: {query}")
logger.info(f"Curriculum Agent task decomposition\nFinal task: {query}")
rsp = await self._aask(prompt=prompt, system_msgs=system_msgs)
logger.info(f"Curriculum Agent task decomposition\n{rsp}")
return fix_and_parse_json(rsp)
def parse_llm_response(self, llm_resp):
task = ""
for line in llm_resp.split("\n"):
@ -46,14 +46,14 @@ class DesignTask(Action):
task = line[5:].replace(".", "").strip()
assert task, "Task not found in Curriculum Agent response"
return {"next_task": task}
async def generate_task(self, human_msg, system_msg, max_retries=5):
"""
Refer to the code in the voyager/agents/curriculum.py propose_next_ai_task() for implementation details.
Returns: task & context
"""
if max_retries == 0:
raise RuntimeError("Max retries reached, failed to propose task.")
curriculum = await self._aask(prompt=human_msg, system_msgs=system_msg)
@ -72,14 +72,14 @@ class DesignTask(Action):
system_msg=system_msg,
max_retries=max_retries - 1,
)
async def run(self, human_msg, system_msg, *args, **kwargs):
logger.info(f"run {self.__repr__()}")
# Call the language model to generate a response.
task = await self.generate_task(human_msg=human_msg, system_msg=system_msg)
return task
@ -88,49 +88,51 @@ class DesignCurriculum(Action):
Action class for designing curriculum-related questions.
Refer to the code in the voyager/agents/curriculum.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
# voyager vectordb using
@classmethod
def generate_qa(cls, events, chest_observation):
async def generate_qa(self, events, human_msg, system_msg):
"""
Generate qa for DesignTask's HumanMessage
"""
questions_new, _ = cls.generate_qa_step1(
events=events, chest_observation=chest_observation
questions_new = await self.generate_qa_step1(
events=events, human_msg=human_msg, system_msg=system_msg
)
logger.debug(f"Generate_qa_step1 result list is HERE: {questions_new}")
questions = []
answers = []
for question in questions_new:
if cls.qa_cache_questions_vectordb._collection.count() > 0:
if self.qa_cache_questions_vectordb._collection.count() > 0:
docs_and_scores = (
cls.qa_cache_questions_vectordb.similarity_search_with_score(
self.qa_cache_questions_vectordb.similarity_search_with_score(
question, k=1
)
)
if docs_and_scores and docs_and_scores[0][1] < 0.05:
question_cached = docs_and_scores[0][0].page_content
assert question_cached in cls.qa_cache
answer_cached = cls.qa_cache[question_cached]
assert question_cached in self.qa_cache
answer_cached = self.qa_cache[question_cached]
questions.append(question_cached)
answers.append(answer_cached)
continue
answer = cls.generate_qa_step2(question=question)
assert question not in cls.qa_cache
cls.qa_cache[question] = answer
cls.qa_cache_questions_vectordb.add_texts(
answer = await self.generate_qa_step2(question=question)
assert question not in self.qa_cache
self.qa_cache[question] = answer
self.qa_cache_questions_vectordb.add_texts(
texts=[question],
)
with open(f"{CKPT_DIR}/curriculum/qa_cache.json", "w") as f:
json.dump(cls.qa_cache, f)
cls.qa_cache_questions_vectordb.persist()
json.dump(self.qa_cache, f)
self.qa_cache_questions_vectordb.persist()
questions.append(question)
answers.append(answer)
assert len(questions_new) == len(questions) == len(answers)
logger.info(f"Curriculum Agent generate_qa Questions: {questions}")
logger.info(f"Curriculum Agent generate_qa Answers: {answers}")
return questions, answers
async def generate_qa_step1(self, events, human_msg, system_msg):
biome = events[-1][1]["status"]["biome"].replace("_", " ")
questions = [
@ -139,7 +141,7 @@ class DesignCurriculum(Action):
f"What are the mobs that I can find in the {biome} in Minecraft?",
]
qa_response = await self._aask(prompt=human_msg, system_msgs=system_msg)
try:
# Regex pattern to extract question and concept pairs
pattern = r"Question \d+: (.+)\nConcept \d+: (.+)"
@ -154,10 +156,10 @@ class DesignCurriculum(Action):
f"QA step 1 ask questions: {e}."
)
return questions
async def generate_qa_step2(self, question):
# Implement the logic for another specific step in generating questions and answers.
logger.info(f"Curriculum Agent Question: {question}")
# logger.info(f"Curriculum Agent generate_qa_step2 Question: {question}")
human_msg = HumanMessage(content=f"Question: {question}").content
system_msg = [
SystemMessage(
@ -165,16 +167,16 @@ class DesignCurriculum(Action):
).content
]
answer = await self._aask(prompt=human_msg, system_msgs=system_msg)
logger.info(f"Curriculum Agent {answer}")
# logger.info(f"Curriculum Agent generate_qa_step2 answer: {answer}")
return answer
async def get_context_from_task(self, task):
"""
Args: task
Returns: context: "Question: {question}\n{answer}"
if include ore in question, gpt will try to use tool with skill touch enhancement to mine
"""
question = (
f"How to {task.replace('_', ' ').replace(' ore', '').replace(' ores', '').replace('.', '').strip().lower()}"
f" in Minecraft?"
@ -192,14 +194,14 @@ class DesignCurriculum(Action):
self.qa_cache_questions_vectordb.persist()
context = f"Question: {question}\n{answer}"
return context
async def generate_context(self, task, max_retries=5):
"""
Refer to the code in the voyager/agents/curriculum.py propose_next_ai_task() for implementation details.
Returns: context
"""
if max_retries == 0:
raise RuntimeError("Max retries reached, failed to propose context.")
try:
@ -213,12 +215,12 @@ class DesignCurriculum(Action):
task=task,
max_retries=max_retries - 1,
)
async def run(self, task, human_msg, system_msg, *args, **kwargs):
logger.info(f"run {self.__repr__()}")
# Generate curriculum-related questions and answers.
# curriculum_qustion = await self.generate_qa_step1(events, human_msg, system_msg)
curriculum_context = await self.generate_context(task)
# Return the generated questions and answers.
return curriculum_context

17
metagpt/actions/minecraft/generate_actions.py
View file

@ -12,11 +12,11 @@ class GenerateActionCode(Action):
Action class for generating action code.
Refer to the code in the voyager/agents/action.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.llm.model = "gpt-4"
async def generate_code(self, human_msg, system_msg=[]):
"""
Generate action code logic.
@ -27,22 +27,23 @@ class GenerateActionCode(Action):
rsp = await self._aask(prompt=human_msg, system_msgs=system_msg)
parsed_result = parse_action_response(rsp)
# logger.info(f"parsed_result is HERE: {parsed_result}")
try:
return (
parsed_result["program_code"],
parsed_result["program_code"] + "\n" + parsed_result["exec_code"],
parsed_result["program_name"],
)
except:
logger.error(f"Failed to parse response: {parsed_result}")
return None, None
return None, None, None # TODO: midify to "", "", ""
async def run(self, human_msg, system_msg, *args, **kwargs):
logger.info(f"run {self.__repr__()}")
# Generate action code.
generated_code, program_name = await self.generate_code(
program_code, generated_code, program_name = await self.generate_code(
human_msg=human_msg, system_msg=system_msg
)
# Return the generated code.
return generated_code, program_name
return program_code, generated_code, program_name

25
metagpt/actions/minecraft/manage_skills.py
View file

@ -5,6 +5,8 @@
import os
import json
from traits.trait_types import self
from metagpt.logs import logger
from metagpt.actions.minecraft.player_action import PlayerActions as Action
from metagpt.const import CKPT_DIR
@ -15,11 +17,11 @@ class RetrieveSkills(Action):
Action class for retrieving skills.
Refer to the code in the voyager/agents/skill.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.llm.model = "gpt-3.5-turbo"
async def run(self, query, skills, *args, **kwargs):
# Implement the logic for retrieving skills here.
k = min(self.vectordb._collection.count(), self.retrieval_top_k)
@ -42,13 +44,13 @@ class AddNewSkills(Action):
Action class for adding new skills.
Refer to the code in the voyager/agents/skill.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.llm.model = "gpt-3.5-turbo"
async def run(
self, task, program_name, program_code, skills, skill_desp, *args, **kwargs
self, task, program_name, program_code, skills, skill_desp, *args, **kwargs
):
# Implement the logic for adding new skills here.
# TODO: Fix this
@ -75,12 +77,7 @@ class AddNewSkills(Action):
ids=[program_name],
metadatas=[{"name": program_name}],
)
# FIXME
# assert self.vectordb._collection.count() == len(
# skills
# ), "vectordb is not synced with skills.json"
with open(f"{CKPT_DIR}/skill/code/{dumped_program_name}.js", "w") as f:
f.write(program_code)
with open(f"{CKPT_DIR}/skill/description/{dumped_program_name}.txt", "w") as f:
@ -99,13 +96,13 @@ class GenerateSkillDescription(Action):
Action class for generating skill descriptions.
Refer to the code in the voyager/agents/skill.py for implementation details.
"""
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.llm.model = "gpt-3.5-turbo"
async def run(self, program_name, human_message, system_message, *args, **kwargs):
# Implement the logic for generating skill descriptions here.
rsp = await self._aask(prompt=human_message, system_msgs=system_message)
skill_description = f" // { rsp}"
skill_description = f" // {rsp}"
return f"async function {program_name}(bot) {{\n{skill_description}\n}}"

74
metagpt/actions/minecraft/player_action.py
View file

@ -2,58 +2,70 @@
# @Date : 2023/9/23 17:06
# @Author : stellahong (stellahong@fuzhi.ai)
# @Desc :
import json
from metagpt.actions import Action
from langchain.vectorstores import Chroma
from langchain.embeddings.openai import OpenAIEmbeddings
from metagpt.document_store import FaissStore
from metagpt.const import CKPT_DIR
from metagpt.config import CONFIG
class PlayerActions(Action):
def __init__(self, name="", context=None, llm=None):
super().__init__(name, context, llm)
self.skills = {}
self.skills = {} # for skills.json
self.qa_cache = {}
self.check_init = True
if (
CONFIG.resume
): # TODO: now for assert only, no update, cuz program using in step()
with open(f"{CKPT_DIR}/skill/skills.json", "r") as f:
self.skills = json.load(f)
with open(f"{CKPT_DIR}/curriculum/qa_cache.json", "r") as f:
self.qa_cache = json.load(f)
self.retrieval_top_k = 5
self.vectordb = Chroma(
collection_name="skill_vectordb",
embedding_function=OpenAIEmbeddings(),
persist_directory=f"{CKPT_DIR}/skill/vectordb",
)
self.qa_cache_questions_vectordb = Chroma(
collection_name="qa_cache_questions_vectordb",
embedding_function=OpenAIEmbeddings(),
persist_directory=f"{CKPT_DIR}/curriculum/vectordb",
)
# TODO: change to FaissStore
# self.qa_cache_questions_vectordb = FaissStore( {CKPT_DIR}/ 'curriculum/vectordb'
@classmethod
def set_skills(cls, skills):
cls.skills = skills
# Check if Skill Manager's vectordb right using
assert cls.vectordb._collection.count() == len(cls.skills), (
f"Skill Manager's vectordb is not synced with skills.json.\n"
f"There are {cls.vectordb._collection.count()} skills in vectordb but {len(cls.skills)} skills in skills.json.\n"
f"Did you set resume=False when initializing the manager?\n"
f"You may need to manually delete the vectordb directory for running from scratch."
)
# FIXME
if self.check_init:
# Check if Skill Manager's vectordb right using
assert self.vectordb._collection.count() >= len(self.skills), (
f"Skill Manager's vectordb is not synced with skills.json.\n"
f"There are {self.vectordb._collection.count()} skills in vectordb but {len(self.skills)} skills in skills.json.\n"
f"Did you set resume=False when initializing the manager?\n"
f"You may need to manually delete the vectordb directory for running from scratch."
)
# Check if Skill Manager's vectordb right using
assert self.qa_cache_questions_vectordb._collection.count() >= len(
self.qa_cache
), (
f"Curriculum Agent's qa cache question vectordb is not synced with qa_cache.json.\n"
f"There are {self.qa_cache_questions_vectordb._collection.count()} questions in vectordb "
f"but {len(self.qa_cache)} questions in qa_cache.json.\n"
f"Did you set resume=False when initializing the agent?\n"
f"You may need to manually delete the qa cache question vectordb directory for running from scratch.\n"
)
self.check_init = False
# TODO: change to FaissStore
# self.qa_cache_questions_vectordb = FaissStore( {CKPT_DIR}/ 'curriculum/vectordb'
@classmethod
def set_qa_cache(cls, qa_cache):
cls.qa_cache = qa_cache
# Check if qa_cache right using
# Check if Skill Manager's vectordb right using
assert cls.qa_cache_questions_vectordb._collection.count() == len(
cls.qa_cache
), (
f"Curriculum Agent's qa cache question vectordb is not synced with qa_cache.json.\n"
f"There are {cls.qa_cache_questions_vectordb._collection.count()} questions in vectordb "
f"but {len(cls.qa_cache)} questions in qa_cache.json.\n"
f"Did you set resume=False when initializing the agent?\n"
f"You may need to manually delete the qa cache question vectordb directory for running from scratch.\n"
)
"""Minecraft player info without any implementation details"""
async def run(self, *args, **kwargs):
raise NotImplementedError
raise NotImplementedError