webrlvr/random_sample/generate_tasks.py
2025-07-01 10:58:17 +00:00

474 lines
20 KiB
Python

import os
import random
import json
import mysql.connector
import argparse
from openai import OpenAI
from dotenv import load_dotenv
# --- Configuration ---
load_dotenv()
server_address = "localhost"
MYSQL_CONFIG = {
"host": server_address,
"port": "23306",
"user": "mcpuser",
"password": "StrongPass123!",
"database": "magentodb"
}
OPENAI_CONFIG = {
"api_key": os.getenv("OPENAI_API_KEY"),
"base_url": os.getenv("OPENAI_BASE_URL"),
"model": "gpt-4o"
}
# --- Prompt Template ---
# This is a carefully engineered prompt to guide the LLM's output.
PROMPT_TEMPLATE = """
You are an expert database analyst and a creative test case designer for e-commerce web applications.
Your goal is to generate realistic administrative tasks that can be solved by a Web Agent navigating an admin panel.
I will provide you with the following context:
1. **Full Database Schema**: A list of `CREATE TABLE` statements for the core tables of a Magento e-commerce platform.
2. **Sampled Data**: A JSON object containing 5 random rows of data from 5 randomly selected core tables. This data is REAL and should be used to inspire specific, answerable questions.
## Your Task
Based on the provided schema and sample data, create a JSON object containing a single key, "questions", which holds an array of up to 10 unique task objects.
### Requirements for Each Question:
- **Web Agent Solvable**: The task must represent a realistic action an administrator would perform in a web UI (e.g., "Find all orders for customer X", "Update the stock for product Y", "Approve a pending review").
- **Grounded in Data**: The questions should be specific, using names, IDs, or values from the provided **Sampled Data** to make them concrete.
- **Utilize Schema**: You can formulate questions that require joining tables, even if not all tables were sampled. The full schema is your guide.
### Output Format
The final output MUST be a single, valid JSON object. Do not include any other text, explanations, or markdown formatting like ```json.
The JSON object must have one key: "questions", containing a JSON array of task objects.
Each object in the array must contain exactly three keys: `question`, `answer`, and `sql`.
- **`question`**: (string) A natural language description of the task for a web agent.
- **`answer`**: (string, integer, float, or list) The precise and concise answer to the question, derived by running the SQL query against the database.
- **`sql`**: (string) The exact, runnable MySQL query that was used to find the answer.
### Output Format Example
```json
{{
"questions": [
{{
"question": "What is the email address for customer with ID 5?",
"answer": "customer5@example.com",
"sql": "SELECT email FROM customer_entity WHERE entity_id = 5;"
}},
{{
"question": "Find the total quantity of item with SKU 'ABC-123' in the cart.",
"answer": 3,
"sql": "SELECT SUM(qty) FROM quote_item WHERE sku = 'ABC-123';"
}}
]
}}
```
---
### Full Database Schema
{schema_context}
---
### Sampled Data
Here is the sample data from randomly selected tables. Use this to make your questions specific.
{sampled_data_str}
---
Now, generate the JSON object based on these instructions.
"""
# This is a new prompt to evaluate results and generate a corrected answer.
SEMANTIC_EVALUATION_PROMPT_TEMPLATE = """
You are a precise data analyst. Your task is to evaluate if a SQL query's result adequately answers a given natural language question. You will then either refine the answer, or completely rephrase the question if the result set is large.
I will provide you with a JSON object containing:
1. `question`: The original question asked.
2. `sql`: The SQL query that was executed.
3. `sql_result`: The actual data returned by executing the SQL query.
4. `row_count`: The number of rows in `sql_result`.
## Your Task
Analyze the inputs and respond with a JSON object. You have three cases. The `new_answer` field MUST always be an array of strings.
### Data Analysis and Refinement Rules
1. **Analyze SQL and Question Intent**: Look at the SQL query (`SELECT`, `COUNT`, `DISTINCT`, etc.) and the natural language `question` to understand the user's goal. Is the goal to count things, list unique items, or retrieve specific related data points?
2. **Handle Duplicates and Merge Data**:
- **De-duplication**: If the `question` implies a list of unique items (e.g., "List the cities..." or "What are the unique order statuses?"), you MUST de-duplicate the values in `sql_result` to form the `new_answer`. For example, if `sql_result` is `[["pending"], ["shipped"], ["pending"]]`, the `new_answer` should be `["pending", "shipped"]`.
- **Data Merging**: If the `sql_result` contains multiple rows related to the same entity (e.g., different attributes of one product), combine the relevant information into a concise `new_answer`. For instance, if the question is "What are the name and price of product 'XYZ'?" and `sql_result` is `[["Product XYZ", 99.99]]`, the `new_answer` is `["Product XYZ", "99.99"]`. If the result was `[["Product XYZ", "Red"], ["Product XYZ", "Blue"]]` for a question about colors, `new_answer` could be `["Red", "Blue"]`. Extract only the information that directly answers the question.
After applying these rules, select one of the three cases below for your response format.
### Case 1: Large Result Set (Question Transformation)
If `row_count` is greater than 10 AND the original `question` does NOT already ask for a count (e.g., it is not phrased like "How many..."), you must transform the question.
Respond with:
```json
{{
"can_answer": true,
"new_question": "How many items were found?",
"new_answer": ["42"]
}}
```
- `can_answer`: (boolean) Must be `true`.
- `new_question`: (string) A rephrased question that asks for the quantity of items. For example, if the original question was "List all products", the new question should be "How many products were found?".
- `new_answer`: (array of strings) An array containing the `row_count` as a single string element.
### Case 2: Standard Answer (No Transformation)
If Case 1 does not apply, but the `sql_result` still provides a clear answer to the original `question` (after applying the refinement rules), respond with:
```json
{{
"can_answer": true,
"new_answer": ["value1", "value2", ...]
}}
```
- `can_answer`: (boolean) Must be `true`.
- `new_answer`: (array of strings) An array containing all the essential parts of the answer extracted and refined from `sql_result`. Every value from the result set that contributes to the answer should be included as a string in the array. This ensures answer completeness.
- **Example 1**: If `question` is "What is the status of order 123?" and `sql_result` is `[["processing"]]`, `new_answer` should be `["processing"]`.
- **Example 2**: If `question` is "List emails for pending customers" and `sql_result` is `[["test@a.com"], ["test@b.com"]]`, `new_answer` should be `["test@a.com", "test@b.com"]`.
- **Example 3**: If `question` is "Get product name and price for SKU 'XYZ'" and `sql_result` is `[["My Product", 19.99]]`, `new_answer` should be `["My Product", "19.99"]`.
### Case 3: The question CANNOT be answered
If the `sql_result` is empty, irrelevant, or insufficient to answer the question, respond with:
```json
{{
"can_answer": false,
"reason": "..."
}}
```
- `can_answer`: (boolean) Must be `false`.
- `reason`: (string) A brief explanation for why the question cannot be answered.
---
### Evaluation Data
{task_data_json}
---
Now, provide your evaluation as a JSON object.
"""
def get_db_connection():
"""Establishes a connection to the MySQL database."""
try:
conn = mysql.connector.connect(**MYSQL_CONFIG)
return conn
except mysql.connector.Error as err:
print(f"Error connecting to MySQL: {err}")
return None
def get_full_schema(cursor, tables):
"""Fetches the CREATE TABLE statements for all core tables."""
schema_parts = []
for table_name in tables:
try:
cursor.execute(f"SHOW CREATE TABLE `{table_name}`")
result = cursor.fetchone()
if result:
schema_parts.append(result[1]) # result[1] is the CREATE TABLE statement
except mysql.connector.Error as err:
print(f"Warning: Could not get schema for table {table_name}: {err}")
return "\n\n".join(schema_parts)
def get_random_tables_and_samples(cursor, tables, num_tables=5, num_samples=5):
"""Selects random tables and samples random rows from them."""
selected_tables = random.sample(tables, num_tables)
sampled_data = {}
for table_name in selected_tables:
try:
# Use ORDER BY RAND() for random sampling. Can be slow on very large tables.
query = f"SELECT * FROM `{table_name}` ORDER BY RAND() LIMIT {num_samples}"
cursor.execute(query)
rows = cursor.fetchall()
if not rows:
sampled_data[table_name] = []
continue
columns = [desc[0] for desc in cursor.description]
# Convert rows (tuples) to a list of dictionaries
sampled_rows = []
for row in rows:
row_dict = {}
for i, col_value in enumerate(row):
# Handle bytes by decoding, fall back to string representation
if isinstance(col_value, bytes):
try:
row_dict[columns[i]] = col_value.decode('utf-8')
except UnicodeDecodeError:
row_dict[columns[i]] = str(col_value)
else:
row_dict[columns[i]] = col_value
sampled_rows.append(row_dict)
sampled_data[table_name] = sampled_rows
except mysql.connector.Error as err:
print(f"Warning: Could not sample data from table {table_name}: {err}")
sampled_data[table_name] = f"Error: {err}"
return sampled_data
def generate_questions(client, schema_context, sampled_data, verbose=False):
"""Generates questions by calling the OpenAI API."""
if not client:
raise ValueError("OpenAI client not provided.")
sampled_data_str = json.dumps(sampled_data, indent=2, default=str)
prompt = PROMPT_TEMPLATE.format(
schema_context=schema_context,
sampled_data_str=sampled_data_str
)
if verbose:
print("\n--- Generation Prompt ---")
print(prompt)
print("-------------------------\n")
try:
response = client.chat.completions.create(
model=OPENAI_CONFIG["model"],
messages=[
{"role": "system", "content": "You are a helpful assistant designed to output JSON."},
{"role": "user", "content": prompt}
],
temperature=0.7,
response_format={"type": "json_object"},
)
content = response.choices[0].message.content
if verbose:
print("\n--- GPT-4o Raw Generation Response ---")
print(content)
print("--------------------------------------\n")
data = json.loads(content)
# The prompt asks for {"questions": [...]}, so we extract the list.
if isinstance(data, dict) and "questions" in data and isinstance(data["questions"], list):
return data["questions"]
elif isinstance(data, list):
# Fallback in case the model returns a list directly
print("Warning: Model returned a raw list instead of an object with a 'questions' key.")
return data
else:
print(f"Warning: Failed to find a 'questions' list in the model's output. Got: {content}")
return None
except Exception as e:
print(f"Error calling OpenAI API or parsing JSON: {e}")
return None
def load_existing_tasks(filepath):
"""Loads tasks from a JSON file if it exists."""
if not os.path.exists(filepath):
return []
try:
with open(filepath, 'r') as f:
content = f.read()
if not content: # Handle empty file
return []
return json.loads(content)
except (json.JSONDecodeError, FileNotFoundError):
print(f"Warning: Could not read or parse {filepath}. Starting with an empty list.")
return []
def evaluate_and_refine_tasks(tasks, client, verbose=False):
"""
Uses an LLM to evaluate if a SQL result answers the question and refines the answer.
"""
if not tasks:
return []
final_validated_tasks = []
print("\nPerforming semantic evaluation and answer refinement with GPT-4o...")
for task in tasks:
# Prepare data for the prompt, excluding the original 'answer'
task_data_for_prompt = {
"question": task["question"],
"sql": task["sql"],
"sql_result": task["sql_result"],
"row_count": task["row_count"]
}
task_data_json = json.dumps(task_data_for_prompt, indent=2, default=str)
prompt = SEMANTIC_EVALUATION_PROMPT_TEMPLATE.format(task_data_json=task_data_json)
if verbose:
print("\n--- Evaluation Prompt ---")
print(prompt)
print("-------------------------\n")
try:
print(f" - Evaluating question: \"{task['question'][:80]}...\"")
response = client.chat.completions.create(
model=OPENAI_CONFIG["model"],
messages=[
{"role": "system", "content": "You are a helpful assistant designed to output JSON."},
{"role": "user", "content": prompt}
],
temperature=0.0, # We want deterministic evaluation
response_format={"type": "json_object"},
)
content = response.choices[0].message.content
if verbose:
print("\n--- GPT-4o Raw Evaluation Response ---")
print(content)
print("----------------------------------------\n")
evaluation_result = json.loads(content)
if evaluation_result.get("can_answer") is True and "new_answer" in evaluation_result:
# Task is valid. Update the answer with the refined one from the LLM.
task['answer'] = evaluation_result['new_answer']
# If the LLM provides a new question, update it.
if 'new_question' in evaluation_result:
task['question'] = evaluation_result['new_question']
print(f" - Question was rephrased: \"{task['question']}\"")
task['sql_execute_result'] = task.pop('sql_result')
task.pop('row_count', None) # Clean up temp key
final_validated_tasks.append(task)
print(f" - Evaluation PASSED. New answer: {json.dumps(task['answer'])}")
else:
reason = evaluation_result.get('reason', 'No reason provided.')
print(f" - Evaluation FAILED. Filtering task.")
print(f" - Reason: {reason}")
print(f" - Question: {task['question']}")
print(f" - SQL: {task['sql']}")
sql_result_str = json.dumps(task['sql_result'], indent=2, default=str)
print(f" - SQL Result: {sql_result_str}")
except Exception as e:
print(f" - An error occurred during semantic evaluation for task, filtering it out: {e}")
print(f" - Question: {task.get('question', 'N/A')}")
print(f" - SQL: {task.get('sql', 'N/A')}")
return final_validated_tasks
def main():
"""Main function to run the script."""
parser = argparse.ArgumentParser(description="Generate and validate e-commerce admin tasks.")
parser.add_argument(
"--target-count",
type=int,
required=True,
help="The total number of questions to generate."
)
parser.add_argument(
"--output-file",
type=str,
default="generated_tasks.json",
help="The file to save the generated tasks to (in JSON format)."
)
parser.add_argument(
"-v", "--verbose",
action="store_true",
help="Enable verbose output, including prompts and raw LLM responses."
)
args = parser.parse_args()
# Load existing tasks from the output file
all_tasks = load_existing_tasks(args.output_file)
print(f"Found {len(all_tasks)} existing valid tasks in '{args.output_file}'.")
# Connect to DB and set up client
conn = get_db_connection()
if not conn:
return
cursor = conn.cursor()
if not OPENAI_CONFIG["api_key"]:
print("Error: OPENAI_API_KEY environment variable not set.")
return
client = OpenAI(api_key=OPENAI_CONFIG["api_key"], base_url=OPENAI_CONFIG["base_url"])
try:
# Load core tables and schema once
try:
with open('core_tables.json', 'r') as f:
core_tables = json.load(f)
except FileNotFoundError:
print("Error: core_tables.json not found. Please create it.")
return
print("Fetching full database schema...")
schema_context = get_full_schema(cursor, core_tables)
# Start the generation loop
round_num = 1
while len(all_tasks) < args.target_count:
print(f"\n--- Starting Generation Round {round_num} ---")
print(f"Goal: {args.target_count} | Current: {len(all_tasks)} | Needed: {args.target_count - len(all_tasks)}")
# Get random samples for this round
print("Sampling data from 5 random tables...")
sampled_data = get_random_tables_and_samples(cursor, core_tables, num_tables=5, num_samples=5)
if args.verbose:
print("\n--- Sampled Data ---")
print(json.dumps(sampled_data, indent=2, default=str))
print("--------------------\n")
# Generate questions
print("Generating questions with GPT-4o...")
generated_tasks = generate_questions(client, schema_context, sampled_data, verbose=args.verbose)
# Execute SQL for generated tasks
tasks_for_evaluation = []
if generated_tasks:
print("\nExecuting SQL for generated tasks...")
for task in generated_tasks:
if not isinstance(task, dict) or not all(k in task for k in ['sql', 'answer', 'question']):
print(f"Filtering task due to malformed structure: {task}")
continue
try:
cursor.execute(task['sql'])
sql_result = cursor.fetchall()
# Create a new dict for evaluation, excluding the original 'answer'.
tasks_for_evaluation.append({
'question': task['question'],
'sql': task['sql'],
'sql_result': sql_result,
'row_count': len(sql_result)
})
except mysql.connector.Error as err:
print(f"Filtering task due to SQL error: {err} on SQL: {task['sql']}")
# Perform semantic evaluation and get validated tasks
validated_tasks = evaluate_and_refine_tasks(tasks_for_evaluation, client, verbose=args.verbose)
# Append new tasks and save to file
if validated_tasks:
all_tasks.extend(validated_tasks)
with open(args.output_file, 'w') as f:
json.dump(all_tasks, f, indent=2, default=str)
print("\n--- Round Summary ---")
print(f"Generated {len(validated_tasks)} new valid tasks in this round.")
print(f"Progress: {len(all_tasks)} / {args.target_count} tasks.")
else:
print("\n--- Round Summary ---")
print("No new valid tasks were generated in this round. Retrying...")
round_num += 1
finally:
# Close the database connection
if conn.is_connected():
cursor.close()
conn.close()
print("\nDatabase connection closed.")
print(f"\nTarget of {args.target_count} tasks reached. Final output saved to {args.output_file}.")
if __name__ == "__main__":
main()