När fine-tuning överträffar prompting

# Prompting is sufficient for most well-defined tasks
system_prompt = '''
You are a customer support agent for TechCorp. Your tone is friendly but professional.
Always:
1. Acknowledge the customer's issue in the first sentence
2. Provide step-by-step solutions with numbered lists
3. End with 'Is there anything else I can help you with?'
Never: reveal pricing, discuss competitors, or make promises about future features.
'''

# With clear instructions, GPT-4o handles this well - no fine-tuning needed
# Before investing in fine-tuning, prove prompting is insufficient

# Problem: prompting gives 90% format compliance - 10% failures cause downstream errors
# Prompt approach (unreliable)
system = 'Always respond with JSON: {"category": "...", "priority": 1-5, "tags": [...]}'
# 1 in 10 responses adds explanation text, omits a field, or uses strings for priority

# Fine-tuned approach: train on 500 examples of perfect output
# Training example format:
train_example = {
    'messages': [
        {'role': 'system', 'content': 'Classify customer support tickets.'},
        {'role': 'user', 'content': 'My order is late and I need it for tomorrow.'},
        {'role': 'assistant', 'content': '{"category": "shipping", "priority": 4, "tags": ["late_delivery", "urgent"]}'}
    ]
}
# After fine-tuning: 99.5%+ format compliance with minimal system prompt

# Example: Internal code style with dozens of company-specific conventions
# Too many rules to fit in a prompt effectively:

# Company conventions (partial list of 200+):
# - Use AppException instead of RuntimeError
# - Repositories are named FooRepository not FooRepo
# - Service methods use handle_verb_noun naming not do_action
# - Config values go through AppConfig.get(), never os.environ directly
# - ... 196 more conventions

# Prompting: you can include ~20 conventions before the model starts ignoring them
# Fine-tuning: train on 1000 examples of compliant vs. non-compliant code
# Result: model learns ALL conventions and applies them automatically

COST_PER_1K_TOKENS_INPUT = 0.0050  # gpt-4o
DAILY_REQUESTS = 10_000_000

# Base model with detailed prompt
base_prompt_tokens = 2000
daily_input_tokens_base = DAILY_REQUESTS * base_prompt_tokens
daily_cost_base = (daily_input_tokens_base / 1000) * COST_PER_1K_TOKENS_INPUT

# Fine-tuned model with short prompt
fine_tuned_prompt_tokens = 50
daily_input_tokens_ft = DAILY_REQUESTS * fine_tuned_prompt_tokens
daily_cost_ft = (daily_input_tokens_ft / 1000) * COST_PER_1K_TOKENS_INPUT

print(f'Base model daily input cost: ${daily_cost_base:,.2f}')
print(f'Fine-tuned model daily input cost: ${daily_cost_ft:,.2f}')
print(f'Monthly savings: ${(daily_cost_base - daily_cost_ft) * 30:,.2f}')
# Base: $100,000/day. Fine-tuned: $2,500/day. Savings: ~$2.9M/month

# Latency comparison (approximate)

# Base model with 2000-token prompt:
# - Input tokens processed: 2000 + 50 (user query) = 2050
# - Response: often starts with 'Sure! Here is...' (5-10 unnecessary tokens)
# - TTFT: ~800ms (more tokens to process)

# Fine-tuned model with 50-token prompt:
# - Input tokens processed: 50 + 50 (user query) = 100
# - Response: starts directly with the answer (no preamble)
# - TTFT: ~100ms (few tokens to process)

# For classification tasks (short outputs), this is a 5-8x latency improvement
# For generation tasks, improvement is less dramatic but still significant

print('Fine-tuning trades upfront training cost for per-request latency+cost savings')

def estimate_fine_tuning_feasibility(num_examples: int, task_type: str) -> str:
    if num_examples < 50:
        return 'Insufficient data. Use few-shot prompting with these examples instead.'
    
    if task_type == 'format_adherence' and num_examples >= 100:
        return 'Fine-tuning recommended. Format consistency issues are hard to solve with prompting.'
    
    if task_type == 'style_matching' and num_examples >= 300:
        return 'Fine-tuning recommended. Consistent style requires enough examples to learn the distribution.'
    
    if task_type == 'domain_knowledge' and num_examples >= 500:
        return 'Fine-tuning recommended if knowledge is truly proprietary.'
    
    return 'Continue with advanced prompting (chain-of-thought, structured output) and revisit fine-tuning when you have more data.'

fine_tuning_total_cost = {
    'data_collection_and_QA': '$5,000-$50,000',  # human annotation or LLM-generated
    'training_compute': '$50-$5,000',             # depends on model size and data volume
    'evaluation_pipeline': '$500-$2,000',         # building eval harness
    'deployment_infra': '$200-$2,000/month',      # serving the custom model
    'maintenance': '$1,000-$5,000/year',          # retraining when things change
    'opportunity_cost': 'weeks to months',        # time to build vs. prompt iteration
}

# Compare to prompting costs:
prompting_costs = {
    'data_needed': None,  # no training data required
    'infra': '$0 (uses existing API)',
    'maintenance': 'update prompts when needed',
    'time_to_production': 'hours to days'
}

# Decision guide: RAG vs Fine-tuning vs Prompting

def choose_approach(requirements: dict) -> str:
    if requirements.get('knowledge_changes_frequently'):  # pricing, news, live data
        return 'RAG - knowledge must be updatable at query time'
    
    if requirements.get('needs_consistent_format') and requirements.get('high_volume'):
        return 'Fine-tuning - format adherence + cost savings at scale'
    
    if requirements.get('proprietary_domain_vocabulary'):
        return 'Fine-tuning - model needs to learn new terminology'
    
    if requirements.get('low_volume') or requirements.get('still_exploring'):
        return 'Prompting - fastest iteration, lowest cost'
    
    if requirements.get('combination_needed'):  # most production systems
        return 'Fine-tuning for style/format + RAG for dynamic knowledge'