변수 치환 기법

파이썬의 템플릿 렌더링 네 가지 방식

접근법 1: 파이썬 f 문자열

import openai

client = openai.OpenAI(api_key='sk-...')

def generate_linkedin_post(company, topic, tone, word_count):
    prompt = (
        f'Write a LinkedIn post for {company} about {topic}. '
        f'Tone: {tone}. '
        f'Length: {word_count} words. '
        'Professional but conversational. '
        'End with one question to engage readers. '
        'No hashtags. Active voice.'
    )

    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.choices[0].message.content

print(generate_linkedin_post(
    company='DataStream Analytics',
    topic='how AI is changing data pipelines',
    tone='enthusiastic but grounded',
    word_count=180
))

접근법 2: str.format()

import openai

client = openai.OpenAI(api_key='sk-...')

# Template stored as a module-level constant or loaded from a file
SUPPORT_REPLY_TEMPLATE = '''You are a customer support agent for {company_name}.

Respond to this customer message:
---
{customer_message}
---

Tone: {tone}.
Keep the response under {max_words} words.
Do not offer refunds unless the customer explicitly asks.
Always close by asking if there is anything else you can help with.'''

def generate_support_reply(company, message, tone='empathetic and helpful', max_words=150):
    prompt = SUPPORT_REPLY_TEMPLATE.format(
        company_name=company,
        customer_message=message,
        tone=tone,
        max_words=max_words
    )

    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.choices[0].message.content

접근법 3: string.Template

from string import Template
import openai

client = openai.OpenAI(api_key='sk-...')

# $ placeholders — safe with code that contains curly braces
BASE_TEMPLATE = Template(
    'Write a $format_type for $audience about $topic. '
    'Tone: $tone. Length: $word_count words. '
    'Active voice. No jargon.'
)

def generate(format_type, audience, topic, tone='professional', word_count=200):
    prompt = BASE_TEMPLATE.substitute(
        format_type=format_type,
        audience=audience,
        topic=topic,
        tone=tone,
        word_count=word_count
    )

    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.choices[0].message.content

# Partial fill example — safe_substitute leaves $word_count as-is
partial = BASE_TEMPLATE.safe_substitute(
    format_type='blog post', audience='developers', topic='API design'
)
print(partial)  # $tone and $word_count remain as placeholders

접근법 4: Jinja2 기초

from jinja2 import Template
import openai

client = openai.OpenAI(api_key='sk-...')

# Jinja2 uses {{ }} for variables and {% %} for logic
JINJA_PROMPT = Template('''
Write a {{content_type}} for {{audience}} about {{topic}}.
Tone: {{tone}}.
{% if include_examples %}
Include {{example_count}} concrete examples.
{% endif %}
{% if word_count %}
Length: {{word_count}} words.
{% else %}
Aim for 200-300 words.
{% endif %}
Active voice. No jargon.
''')

def generate(content_type, audience, topic, tone, include_examples=False, example_count=2, word_count=None):
    prompt = JINJA_PROMPT.render(
        content_type=content_type,
        audience=audience,
        topic=topic,
        tone=tone,
        include_examples=include_examples,
        example_count=example_count,
        word_count=word_count
    )

    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.choices[0].message.content

템플릿의 Jinja2 반복문

from jinja2 import Template
import openai

client = openai.OpenAI(api_key='sk-...')

MULTI_PRODUCT_TEMPLATE = Template('''
Write a product comparison for {{audience}}.
Compare the following products:

{% for product in products %}
- {{product.name}}: {{product.description}}
{% endfor %}

Structure: one paragraph per product, then a 2-sentence recommendation.
Tone: {{tone}}. Active voice. No bullet points in paragraphs.
''')

products = [
    {'name': 'Asana', 'description': 'project management with timeline views'},
    {'name': 'Linear', 'description': 'developer-focused issue tracking'},
    {'name': 'Monday.com', 'description': 'visual work management for teams'}
]

prompt = MULTI_PRODUCT_TEMPLATE.render(
    audience='startup founders',
    products=products,
    tone='direct and practical'
)

response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=[{'role': 'user', 'content': prompt}]
)
print(response.choices[0].message.content)

Jinja2 필터

파일에서 템플릿 불러오기

from jinja2 import Environment, FileSystemLoader
import openai

client = openai.OpenAI(api_key='sk-...')

# Load all templates from the 'prompts/' directory
env = Environment(loader=FileSystemLoader('prompts/'))

def render_template(template_name, variables):
    '''Load and render a .j2 template file with the given variables.'''
    template = env.get_template(template_name)
    return template.render(**variables)

def generate_from_file(template_name, variables):
    prompt = render_template(template_name, variables)

    response = client.chat.completions.create(
        model='gpt-4o-mini',
        messages=[{'role': 'user', 'content': prompt}]
    )
    return response.choices[0].message.content

# Usage: load prompts/blog_post.j2 and fill with variables
result = generate_from_file('blog_post.j2', {
    'topic': 'API rate limiting strategies',
    'audience': 'backend engineers',
    'tone': 'technical and direct',
    'word_count': 500
})
print(result)

적절한 방식 선택하기

템플릿 안전성: 주입 공격

템플릿 렌더링 테스트

def test_template_render():
    test_cases = [
        {'topic': 'cloud security', 'audience': 'CTOs', 'tone': 'formal', 'word_count': 300},
        {'topic': 'ML pipelines', 'audience': 'data scientists', 'tone': 'technical', 'word_count': 500},
        # Edge cases
        {'topic': '', 'audience': 'developers', 'tone': 'casual', 'word_count': 100},  # empty topic
        {'topic': 'AI' * 100, 'audience': 'all', 'tone': 'brief', 'word_count': 50},  # very long topic
    ]

    TEMPLATE = 'Write a {word_count}-word {tone} article about {topic} for {audience}. Active voice.'

    for i, case in enumerate(test_cases):
        try:
            rendered = TEMPLATE.format(**case)
            assert len(rendered) > 0, 'Empty render'
            print(f'Case {i+1} OK: {len(rendered)} chars')
        except (KeyError, AssertionError) as e:
            print(f'Case {i+1} FAILED: {e}')

test_template_render()

지식 확인: 치환 기법

복습: 변수 치환 기법