import json
import os
import sys
import time
import copy
import re
from pathlib import Path
from typing import List, Literal, Optional, Tuple, TypedDict, Dict
import numpy as np
from tqdm import tqdm
# Get the path from environment variable
prj_root_path = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.append(prj_root_path)
from code_interpreter.JuypyterClient import JupyterNotebook
from code_interpreter.BaseCodeInterpreter import BaseCodeInterpreter
from utils.const import *
from prompt.gpt4_prompt import CODE_INTERPRETER_SYSTEM_PROMPT
# from prompt.gpt4_prompt import CODE_INTERPRETER_SYSTEM_PROMPT
from colorama import init, Fore, Style, Back
from rich.markdown import Markdown
import base64
import openai
from retrying import retry
import requests
import logging
from termcolor import colored
# load from key file
with open("./openai_api_key.txt") as f:
OPENAI_API_KEY = key = f.read()
openai.api_key = OPENAI_API_KEY
from utils.cleaner import clean_error_msg
def remove_string(s):
pattern = r"\d{4}-\d{2}-\d{2} \d{2}:\d{2}:\d{2}\.\d{6}:.*LD_LIBRARY_PATH: /usr/local/nvidia/lib:/usr/local/nvidia/lib64\n"
return re.sub(pattern, "", s)
def clean_the_dialog(dialog, question):
question_idx = 0
for idx, item in enumerate(dialog):
if item["content"] == question:
question_idx = idx
filtered_dialog = dialog[question_idx:]
user_qinit_dict = filtered_dialog[0]
answer_fuse_str = "\n".join([i["content"].strip() for i in filtered_dialog[1::2]])
final_dialog_dict = [
{"role": "user", "content": user_qinit_dict["content"]},
{"role": "assistant", "content": answer_fuse_str},
]
return final_dialog_dict
@retry(
stop_max_attempt_number=7,
wait_exponential_multiplier=1000,
wait_exponential_max=10000,
)
def get_embedding(text, model="text-embedding-ada-002"):
global counter
headers = {
"Authorization": f"Bearer {OPENAI_API_KEY}", # Make sure to replace with your OpenAI API key
"Content-Type": "application/json",
}
payload = {"input": text, "model": model}
response = requests.post(
"https://api.openai.com/v1/embeddings", headers=headers, json=payload
)
if response.status_code != 200:
raise Exception(f"Request failed with status {response.status_code}")
return np.array(response.json()["data"][0]["embedding"])
class QueryRetrospect:
def __init__(
self,
data_directory="./gpt_data_gen_retrospect/",
embeddings_path="./gpt_data_gen_retrospect/embeddings.npy",
):
self.data_directory = data_directory
self.embeddings_path = embeddings_path
self.data = []
self.embeddings = []
if os.path.exists(embeddings_path):
print("++ Embedding Exists!")
self.embeddings = np.load(embeddings_path)
for fname in [i for i in os.listdir(data_directory) if i.endswith(".json")]:
with open(
os.path.join(data_directory, fname),
"r",
encoding="utf-8",
errors="replace",
) as f:
self.data.append(json.load(f))
else:
only_files = [
f
for f in os.listdir(data_directory)
if os.path.isfile(os.path.join(data_directory, f))
and f.endswith(".json")
]
for fname in tqdm(only_files):
with open(
os.path.join(data_directory, fname), "r", encoding="cp1252"
) as f:
data_point = json.load(f)
self.data.append(data_point)
self.embeddings.append(
get_embedding(data_point["execution_result"])
)
self.embeddings = np.array(self.embeddings)
self.save_embeddings()
print(f"++ Embedding Saved! {self.embeddings.shape}")
def save_embeddings(self):
np.save(self.embeddings_path, self.embeddings)
def __call__(self, query, top_k=3, VERBOSE: bool = False):
query_embedding = get_embedding(query)
similarities = np.dot(self.embeddings, query_embedding)
top_indices = similarities.argsort()[-top_k:][::-1]
return [self.data[i]["retrospection"] for i in top_indices]
class QueryRetrospectPrefix:
def __init__(
self,
model="gpt-4",
data_directory="./eval/gpt_mbpp_output",
embeddings_path="./eval/gpt_mbpp_output/embeddings.npy",
):
self.data_directory = data_directory
self.embeddings_path = embeddings_path
self.data = []
self.embeddings = []
if os.path.exists(embeddings_path):
print("++ Embedding Exists!")
self.embeddings = np.load(embeddings_path)
for fname in [i for i in os.listdir(data_directory) if i.endswith(".json")]:
with open(
os.path.join(data_directory, fname),
"r",
encoding="utf-8",
errors="replace",
) as f:
self.data.append(json.load(f))
else:
only_files = [
f
for f in os.listdir(data_directory)
if os.path.isfile(os.path.join(data_directory, f))
and f.endswith(".json")
]
for fname in tqdm(only_files):
with open(
os.path.join(data_directory, fname), "r", encoding="cp1252"
) as f:
data_point = json.load(f)
print(f'Processing "{data_point[1]["content"]}" ...')
self.data.append(data_point)
self.embeddings.append(get_embedding(data_point[1]["content"]))
self.embeddings = np.array(self.embeddings)
self.save_embeddings()
print(f"++ Embedding Saved! {self.embeddings.shape}")
self.model = model
self.dialog = [
{
"role": "system",
"content": "You are retrospection GPT. retrospect from the given data.",
},
{
"role": "user",
"content": 'Current Question:\n\nWrite a Python function to solve the following task:\n\nfrom typing import List\n\ndef cum_sum(numbers: List[int]) -> List[int]:\n """\n From a given list of integers, generate a list representing the cumulative sum of elements at each index.\n >>> cum_sum([1, 2, 3, 4])\n [1, 3, 6, 10]\n """\n\nRetrieved Trajectories : \nIn a past interaction, a function named running_average was provided to calculate the running average of a list of numbers.\n\n```python\ndef running_average(numbers: List[int]) -> List[float]:\n total = 0\n averages = []\n for i, num in enumerate(numbers):\n total += num\n averages.append(total / (i+1))\n return averages\n\nprint(running_average([1,2,3,4])) # expected [1.0, 1.5, 2.0, 2.5]\n```\n```RESULT\n[1.0, 1.5, 2.0, 2.5]\n```\nThe output is expected. \n\n',
},
{
"role": "assistant",
"content": "From previous similar questions :\nThe `running_average` function highlights an important concept of maintaining a running or cumulative value (total) as one iterates over the list. This is directly applicable to the cum_sum problem.\n\nApplication to the Question:\nFor the cum_sum function, one needs to maintain a cumulative total of the elements as we traverse through the list. The running_average function is most closely related since it involves accumulating a total and storing intermediate results. By adapting this logic (i.e., excluding the division operation to compute the average), one can easily derive the cumulative sum solution.",
},
]
self.response = ""
@retry(
stop_max_attempt_number=7,
wait_exponential_multiplier=1000,
wait_exponential_max=10000,
)
def ChatCompletion(self):
try:
self.response = openai.ChatCompletion.create(
model=self.model, messages=self.dialog, temperature=0.2, top_p=0.9
)
except Exception as e:
print(f"error while OPENAI api call {e} {self.response}")
def save_embeddings(self):
np.save(self.embeddings_path, self.embeddings)
def __call__(self, query, top_k=3, VERBOSE: bool = False):
query_embedding = get_embedding(query)
similarities = np.dot(self.embeddings, query_embedding)
top_indices = similarities.argsort()[-top_k:][::-1]
top_i = top_indices[0]
prior_traj = self.data[top_i][-1]["content"]
ask_dict = {
"role": "user",
"content": f"Current Question:\n\n{query}\n\nRetrieved Trajectories :\n{prior_traj}",
}
# print(f"From prior experience:\n{prior_traj}\n\nCurrent Question:\n{query}\n")
self.dialog.append(ask_dict)
self.ChatCompletion()
return self.response["choices"][0]["message"]["content"]
class RetrospectiveGPTCodeInterpreter(BaseCodeInterpreter):
def __init__(self, model="gpt-4"):
self.model = model
self.dialog = [
# {"role": "system", "content": CODE_INTERPRETER_SYSTEM_PROMPT },
{
"role": "system",
"content": CODE_INTERPRETER_SYSTEM_PROMPT,
},
# {"role": "user", "content": "How can I use BeautifulSoup to scrape a website and extract all the URLs on a page?"},
# {"role": "assistant", "content": "I think I need to use beatifulsoup to find current korean president,"}
]
# self.dialog += few_shot_4
self.response = None
assert os.path.isfile(
"./openai_api_key.txt"
), "The openai_api_key.txt file could not be found. Please make sure it is in the same directory as this script, and that it contains your OpenAI API key."
# load from key file
with open("./openai_api_key.txt") as f:
OPENAI_API_KEY = f.read()
openai.api_key = OPENAI_API_KEY
self.nb = JupyterNotebook()
out = self.nb.add_and_run(TOOLS_CODE) # tool import
# retrospections
self.retrospector = QueryRetrospectPrefix()
def get_response_content(self):
if self.response:
return self.response["choices"][0]["message"]["content"]
else:
return None
@retry(
stop_max_attempt_number=7,
wait_exponential_multiplier=1000,
wait_exponential_max=10000,
)
def ChatCompletion(self):
try:
self.response = openai.ChatCompletion.create(
model=self.model, messages=self.dialog, temperature=0.2, top_p=0.9
)
except Exception as e:
print(f"error while OPENAI api call {e}")
def save_dialog(self, path: str = "./output/dialog.json"):
with open(path, "w") as f:
json.dump(self.dialog, f)
print(f" ++Dialog saved to [{path}]")
def close(self):
"""
close jupyter notebook, and this class instance
"""
self.nb.close()
def chat(
self,
user_message: str,
VERBOSE: bool = False,
MAX_TRY: int = 6,
code_exec_prefix: str = "",
feedback_prompt: str = "",
append_result: bool = True,
use_retrospect: bool = True,
):
prefix_retrospection = self.retrospector(query=user_message)
self.dialog.append(
{"role": "user", "content": f"{prefix_retrospection}\n\n{user_message}"}
)
init_feedback = copy.deepcopy(feedback_prompt)
code_block_output = ""
attempt = 0
img_data = None
if VERBOSE:
print(
"###Retrospection : "
+ Fore.BLUE
+ Back.WHITE
+ Style.BRIGHT
+ prefix_retrospection
+ Style.RESET_ALL
)
print(
"###User : " + Fore.BLUE + Style.BRIGHT + user_message + Style.RESET_ALL
)
print("\n###Assistant : ")
for i in range(MAX_TRY):
# GPT response
self.ChatCompletion()
# Get code block
generated_text = self.get_response_content()
generated_code_blocks = self.extract_code_blocks(generated_text)
# execute code
if len(generated_code_blocks) > 0:
# Find the position of the first code block in the last answer
first_code_block_pos = (
generated_text.find(generated_code_blocks[0])
if generated_code_blocks
else -1
)
text_before_first_code_block = (
generated_text
if first_code_block_pos == -1
else generated_text[:first_code_block_pos]
)
if VERBOSE:
print(Fore.GREEN + text_before_first_code_block + Style.RESET_ALL)
if VERBOSE:
print(
Fore.YELLOW
+ generated_code_blocks[0]
+ "\n```\n"
+ Style.RESET_ALL
)
code_block_output, error_flag = self.execute_code_and_return_output(
generated_code_blocks[0]
)
code_block_output = f"{code_block_output}"
if code_block_output is not None:
code_block_output = code_block_output.strip()
code_block_output = remove_string(code_block_output)
if len(code_block_output) > 500:
code_block_output = (
code_block_output[:200] + "⋯(skip)⋯" + code_block_output[-200:]
)
code_block_output_str = f"\n```RESULT\n{code_block_output}\n```\n"
if append_result:
gen_final = f"{text_before_first_code_block}{generated_code_blocks[0]}\n```{code_block_output_str}"
if VERBOSE:
print(
Fore.LIGHTBLACK_EX + code_block_output_str + Style.RESET_ALL
)
else:
gen_final = (
f"{text_before_first_code_block}{generated_code_blocks[0]}\n```"
)
self.dialog.append(
{
"role": "assistant",
"content": gen_final,
}
)
feedback_prompt = f"{init_feedback}\nif you accomplish the instruction just say <done>\nIf not keep going."
if VERBOSE:
print(Fore.MAGENTA + feedback_prompt + Style.RESET_ALL)
feedback_dict = {
"role": "user",
"content": feedback_prompt,
}
self.dialog.append(feedback_dict)
else:
if "<done>" in generated_text:
generated_text = generated_text.split("<done>")[0].strip()
if len(generated_text) <= 0:
break
if VERBOSE:
print(Fore.GREEN + generated_text + Style.RESET_ALL)
self.dialog.append(
{
"role": "assistant",
"content": f"{generated_text}",
}
)
break
self.dialog = [self.dialog[0]] + clean_the_dialog(
self.dialog, question=f"{prefix_retrospection}\n\n{user_message}"
) # delete retrospections after generation step
return self.dialog[-1]
if __name__ == "__main__":
import pickle
import random
from tqdm import tqdm
# python3 -m code_interpreter.RetrospectiveGPTCodeInterpreter
retro_interpreter = RetrospectiveGPTCodeInterpreter(model="gpt-4")
instruction = """
Write a Python script to solve the following problem:
def get_row(lst, x):
\"\"\"
You are given a 2 dimensional data, as a nested lists,
which is similar to matrix, however, unlike matrices,
each row may contain a different number of columns.
Given lst, and integer x, find integers x in the list,
and return list of tuples, [(x1, y1), (x2, y2) ...] such that
each tuple is a coordinate - (row, columns), starting with 0.
Sort coordinates initially by rows in ascending order.
Also, sort coordinates of the row by columns in descending order.
Examples:
get_row([
[1,2,3,4,5,6],
[1,2,3,4,1,6],
[1,2,3,4,5,1]
], 1) == [(0, 0), (1, 4), (1, 0), (2, 5), (2, 0)]
get_row([], 1) == []
get_row([[], [1], [1, 2, 3]], 3) == [(2, 2)]
\"\"\"
Ensure the solution is verified by printing the expected output.
"""
# instruction = "Can you make a image of astraunaut in the garden?"
# example
retro_interpreter.chat(
user_message=instruction,
MAX_TRY=5,
use_retrospect=True,
feedback_prompt="Ensure the output matches the expected result, taking into account any corner cases. If discrepancies arise, pinpoint where you went wrong. Then, refine the code to achieve the desired outcome.",
VERBOSE=True,
)