code/data_processing/generate_fixed_sql_data.py

import argparse
import os
import json
from tqdm import tqdm
from planner import get_answer_llamacpp, get_answer_openai
from openai import OpenAI
from dotenv import load_dotenv
from multiprocessing import Pool, Manager

# Set up argument parser
parser = argparse.ArgumentParser()
parser.add_argument('--validator_select', type=str, default='data/multi-agents/validator/gpt-4o-mini-validator_select_bird_with_evidence_train.jsonl')
parser.add_argument('--validator_condition', type=str, default='data/multi-agents/validator/gpt-4o-mini-validator_condition_bird_with_evidence_train.jsonl')
parser.add_argument('--validator_join', type=str, default='data/multi-agents/validator/gpt-4o-mini-validator_join_bird_with_evidence_train.jsonl')
parser.add_argument('--validator_order', type=str, default='data/multi-agents/validator/gpt-4o-mini-validator_order_bird_with_evidence_train.jsonl')
parser.add_argument('--output_file', type=str, default='./data/multi-agents/fixed/gpt-4o-mini-fixed-bird_with_evidence_train.jsonl')
parser.add_argument('--endpoint_type', type=str, default='openai', choices=['openai', 'vllm'])
args = parser.parse_args()

# Define FixAgent class
class FixAgent:
    def __init__(self, prompt_template, endpoint_type='llamacpp'):
        self.prompt_template = prompt_template

        load_dotenv()

        if endpoint_type == 'llamacpp':
            self.get_answer = get_answer_llamacpp
        elif endpoint_type == 'vllm':
            # self.get_answer = get_answer_vllm
            client = OpenAI(
                base_url="http://localhost:8003/v1",
                api_key="no-key",
            )
            self.get_answer = lambda x: get_answer_openai(client, x, model='Qwen/Qwen2.5-14B-Instruct/')
        elif endpoint_type == 'openai':
            client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
            self.get_answer = lambda x: get_answer_openai(client, x)

    def generate(self, sample, feedback_select, feedback_condition, feedback_join, feedback_order):
        prompt = self.prompt_template.format(
            schema=sample['schema_sequence'], 
            question=sample['question'],
            evidence=sample['evidence'],
            sql_query=sample['sql'],
            execution_response=sample['pred_result'],
            feedback_select=feedback_select,
            feedback_condition=feedback_condition,
            feedback_join=feedback_join,
            feedback_order=feedback_order
        )
        answer = self.get_answer([{"role": "user", "content": prompt}])
        return answer


# Define the prompt template
PROMPT = """You are a SQL tutor that helps fixing the SQL query generated by a student. Given a database schema and a question with external knowledge. Generate Fixed SQL query based on the feedback. Write the SQL query directly, do not add more thoughts.

{schema}

Question: {question}
External knowledge: {evidence}

Generated SQL query from student with the execution response.
SQL query: {sql_query}

Execution response [written in pandas format]:
{execution_response}

The feedback for the SQL query:
{feedback_select}

{feedback_condition}

{feedback_join}

{feedback_order}

FIXED SQL:"""

# Define input files
input_files = [
    args.validator_select,
    args.validator_condition,
    args.validator_join,
    args.validator_order
]

# Read data from input files
input_data = [[], [], [], []]
for i, input_file in enumerate(input_files):
    with open(input_file, 'r') as f:
        for line in f:
            input_data[i].append(json.loads(line))

# for each input_data, get subset of data where its question can be found in all input_data, take validator_condition as a reference
# first get intersection of questions from all input_data
question_select = [sample['question'] for sample in input_data[0]]
question_condition = [sample['question'] for sample in input_data[1]]
question_join = [sample['question'] for sample in input_data[2]]
question_order = [sample['question'] for sample in input_data[3]]

subset_questions = set(question_select) & set(question_condition) & set(question_join) & set(question_order)

# then get subset of each input_data where its question can be found in subset_questions
for i in range(len(input_data)):
    input_data[i] = [sample for sample in input_data[i] if sample['question'] in subset_questions]

# rearrange input_data to have the same order of questions as validator_condition
# Build a mapping from question to sample for input_data[0], [2], and [3]
question_to_sample = {}
for i in [0, 2, 3]:
    question_to_sample[i] = {sample['question']: sample for sample in input_data[i]}

# Rearrange input_data[0], [2], and [3] to follow the order of input_data[1]
ordered_questions = [sample['question'] for sample in input_data[1]]
for i in [0, 2, 3]:
    input_data[i] = [question_to_sample[i][question] for question in ordered_questions]

# print length of each input_data
for i in range(len(input_data)):
    print(f"Length of input_data[{i}]: {len(input_data[i])}")

# Ensure questions are correctly aligned across all data
for i in tqdm(range(len(input_data[0])), desc="Checking input alignment"):
    question_0 = input_data[0][i]['question']
    question_1 = input_data[1][i]['question']
    question_2 = input_data[2][i]['question']
    question_3 = input_data[3][i]['question']
    assert question_0 == question_1 == question_2 == question_3

# Load already processed questions if output file exists
processed_questions = set()
if os.path.exists(args.output_file):
    with open(args.output_file, 'r') as f:
        for line in f:
            processed_sample = json.loads(line)
            processed_questions.add(processed_sample['question'])

# Ensure output directory exists
os.makedirs(os.path.dirname(args.output_file), exist_ok=True)

# Main function to handle multiprocessing
def main():
    # Filter out already processed samples and create a list of data to process
    indices_to_process = [i for i in range(len(input_data[0])) if input_data[0][i]['question'] not in processed_questions]
    data_to_process = []
    for i in indices_to_process:
        sample_select = input_data[0][i]
        if sample_select['question'] in processed_questions:
            continue
        sample_condition = input_data[1][i]
        sample_join = input_data[2][i]
        sample_order = input_data[3][i]
        data_to_process.append((sample_select, sample_condition, sample_join, sample_order))

    with Manager() as manager:
        lock = manager.Lock()
        output_file_path = args.output_file
        with Pool(processes=8, initializer=init_process, initargs=(output_file_path, lock, args.endpoint_type)) as pool:
            list(tqdm(
                pool.imap_unordered(process_sample, data_to_process),
                total=len(data_to_process),
                desc="Generating Fixed SQL"
            ))

# Initialize shared resources in worker processes
def init_process(output_file_path_arg, lock_arg, endpoint_type_arg):
    global output_file_path
    global lock
    global fixed_sql_agent
    output_file_path = output_file_path_arg
    lock = lock_arg
    fixed_sql_agent = FixAgent(PROMPT, endpoint_type=endpoint_type_arg)

# Define the function to be executed by each process
def process_sample(samples):
    global output_file_path
    global lock
    global fixed_sql_agent
    sample_select, sample_condition, sample_join, sample_order = samples

    select_correct = sample_select['feedback_conclude']
    condition_correct = sample_condition['feedback_conclude']
    join_correct = sample_join['feedback_conclude']
    order_correct = sample_order['feedback_conclude'] if sample_order['validator_order'] is not None else True

    print(select_correct, condition_correct, join_correct, order_correct)

    if select_correct and condition_correct and join_correct and order_correct:
        return None  # Skip if all feedbacks are correct

    feedback_select = sample_select['validator_select']
    feedback_condition = sample_condition['validator_condition']
    feedback_join = sample_join['validator_join']
    feedback_order = sample_order['validator_order']

    if feedback_select is None:
        feedback_select = "SELECT.\nNone"
    if feedback_condition is None:
        feedback_condition = "CONDITION.\nNone"
    if feedback_join is None:
        feedback_join = "JOIN.\nNone"
    if feedback_order is None:
        feedback_order = "ORDER BY.\nNone"

    sample_select['validator_condition'] = sample_condition['validator_condition']
    sample_select['validator_join'] = sample_join['validator_join']
    sample_select['validator_order'] = sample_order['validator_order']

    # Generate fixed SQL
    fixed_sql = fixed_sql_agent.generate(sample_select, feedback_select, feedback_condition, feedback_join, feedback_order)
    sample_select['fixed_sql'] = fixed_sql

    # Output the result directly to the output file using the lock
    result = json.dumps(sample_select)
    with lock:
        with open(output_file_path, 'a') as output_fp:
            output_fp.write(result + '\n')

# Execute the main function
if __name__ == "__main__":
    main()