""" Generate realistic synthetic customer support and internal escalation documents. Features multi-format data (JSON, Slack, Logs, Emails), ambiguity, and VIP handling. """ import numpy as np import random import string import json from typing import Tuple, List, Dict from sklearn.feature_extraction.text import TfidfVectorizer from constants import CATEGORIES def get_task_config(difficulty: str) -> Dict: configs = { "easy": { "num_documents": 100, "num_categories": 5, "feature_dim": 100, "max_word_count": 200, "time_limit_per_action": None, "time_limit_per_episode": None, "ambiguity_rate": 0.0, "sla_seconds": None, }, "medium": { "num_documents": 500, "num_categories": 10, "feature_dim": 100, "max_word_count": 500, "time_limit_per_action": 2.0, "time_limit_per_episode": 1800, "ambiguity_rate": 0.15, "sla_seconds": 120.0, }, "hard": { "num_documents": 1000, "num_categories": 22, "feature_dim": 100, "max_word_count": 1000, "time_limit_per_action": 1.0, "time_limit_per_episode": 1200, "ambiguity_rate": 0.30, "sla_seconds": 60.0, } } return configs[difficulty] FIRST_NAMES = ["James", "Sarah", "Michael", "Emily", "David", "Jessica", "Robert", "Wei", "Fatima", "Carlos"] LAST_NAMES = ["Smith", "Johnson", "Williams", "Brown", "Jones", "Garcia", "Miller", "Chen", "Al-Rashid"] COMPANIES = ["Acme Corp", "TechForward Inc", "Global Solutions Ltd", "Stripe", "AWS", "Datadog"] def _ctx(): name = f"{random.choice(FIRST_NAMES)} {random.choice(LAST_NAMES)}" is_vip = random.random() < 0.15 # 15% chance of being a VIP enterprise client return { "name": name, "email": f"{name.lower().replace(' ', '.')}@{random.choice(['gmail.com', 'company.com', 'enterprise.io'])}", "company": random.choice(COMPANIES), "order_id": f"ORD-{random.randint(100000, 999999)}", "amount": f"${random.randint(50, 9999)}.{random.randint(0, 99):02d}", "date": f"2025-{random.randint(1,12):02d}-{random.randint(1,28):02d}", "is_vip": is_vip, "urgency_level": "CRITICAL" if is_vip else random.choice(["LOW", "MEDIUM", "HIGH"]) } class TaskDataGenerator: def __init__(self, difficulty: str, seed: int = None): self.difficulty = difficulty self.config = get_task_config(difficulty) if seed is not None: np.random.seed(seed) random.seed(seed) self.categories = list(CATEGORIES[difficulty].values()) # Simple vocabulary for TF-IDF training self.vectorizer = TfidfVectorizer(max_features=self.config["feature_dim"], lowercase=True, stop_words='english') self.vectorizer.fit(["invoice billing refund", "technical bug crash", "legal contract compliance", "hr payroll benefits complaint", "marketing operations finance strategy tax social"]) def _get_corpus_text(self) -> str: if not hasattr(self, "_hf_corpus"): try: from datasets import load_dataset ds = load_dataset("ag_news", split="test") self._hf_corpus = [x["text"] for x in ds] except Exception: self._hf_corpus = ["Please assist with process.", "Error encountered in system."] return random.choice(self._hf_corpus) def _generate_format(self, category: str, ctx: dict) -> str: """Generates content in various formats (JSON, Slack, Log, Email) with Real Corpus Injection""" fmt = random.choice(["json", "slack", "log", "email"]) cat_lower = category.lower() real_text = self._get_corpus_text() content = "" if fmt == "json": payload = { "event_id": f"evt_{random.randint(1000,9999)}", "source": "webhook_listener", "timestamp": ctx["date"], "customer": {"id": ctx["email"], "vip_tier": "Enterprise" if ctx["is_vip"] else "Standard"}, "payload": f"Category: {cat_lower}. Order {ctx['order_id']} value {ctx['amount']}. Extract: {real_text}", "metadata": {"raw_category_hint": category, "severity": ctx["urgency_level"]} } content = json.dumps(payload, indent=2) elif fmt == "slack": content = f"**Slack Transcript - Channel: #triage-incoming**\n" content += f"[{ctx['date']} 09:12 AM] @system_bot: New ticket from {ctx['name']} ({ctx['company']}). VIP: {ctx['is_vip']}\n" content += f"[{ctx['date']} 09:13 AM] {ctx['name']}: Hey team, we have an issue regarding {cat_lower}.\n" content += f"[{ctx['date']} 09:14 AM] {ctx['name']}: For context: {real_text}" elif fmt == "log": content = f"2025-01-01T12:00:00Z WARN [req_id={random.randint(100,999)}] subsystem={category.split('-')[0]} action=PROCESS userId={ctx['email']} msg=\"{real_text}\"" else: # Email vip_tag = "[VIP - URGENT] " if ctx['is_vip'] else "" content = f"Subject: {vip_tag}Inquiry regarding {cat_lower}\nFrom: {ctx['email']}\n\nTo whom it may concern,\nWe are reaching out about order {ctx['order_id']} value {ctx['amount']}.\n\nAdditional details: {real_text}\n\nRegards,\n{ctx['name']}" return content def generate_task_data(self) -> Tuple[List[Dict], np.ndarray]: documents = [] labels = [] num_docs = self.config["num_documents"] ambiguity_rate = self.config.get("ambiguity_rate", 0.0) for i in range(num_docs): ctx = _ctx() if random.random() < ambiguity_rate: # Deliberate heavy ambiguity: Mention multiple departments cat1 = random.choice(self.categories) cat2 = random.choice(self.categories) true_cat = random.choice([cat1, cat2]) content = self._generate_format(f"{cat1} and {cat2}", ctx) else: true_cat = random.choice(self.categories) content = self._generate_format(true_cat, ctx) features = self.vectorizer.transform([content]).toarray()[0] if np.max(features) > 0: features = 2 * features / (np.max(features) + 1e-8) - 1 doc = { "id": f"doc_{i:06d}", "content": content, "word_count": len(content.split()), "has_urgency_markers": ctx["is_vip"], # VIP flag passed via urgency "features": features.tolist(), "true_category": true_cat, "is_vip": ctx["is_vip"] } documents.append(doc) labels.append(self.categories.index(true_cat)) return documents, np.array(labels)