Que signifie le contexte dans la formulation de requêtes pour l’IA ?

Contexte : les informations générales fournies au modèle

Ce qui se passe sans contexte

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

no_context = 'What should I do about the database issue?'

with_context = (
    'We are running PostgreSQL 15 on AWS RDS. '
    'Query response time spiked from 20ms to 800ms after a schema migration yesterday. '
    'The migration added 3 new indexes on the orders table (500M rows). '
    'No other infrastructure changes were made. '
    'What should I investigate first to diagnose the slowdown?'
)

for label, prompt in [('NO CONTEXT', no_context), ('WITH CONTEXT', with_context)]:
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=200,
        messages=[{'role': 'user', 'content': prompt}]
    )
    print(f'--- {label} ---')
    print(response.content[0].text[:300])
    print()

Les quatre types de contexte

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

# Prompt with all 4 context types explicitly labelled
response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{
        'role': 'user',
        'content': (
            'Domain context: B2B SaaS, fintech, invoice reconciliation.\n'
            'Audience context: mid-market CFOs with basic Excel skills, no coding.\n'
            'Goal context: write a one-pager that convinces them to book a demo.\n'
            'Constraint context: max 300 words, no technical jargon, no pricing mentioned.\n\n'
            'Task: Write the one-pager.'
        )
    }]
)
print(response.choices[0].message.content)

Contexte du domaine

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

# Same question, different domain context
domains = [
    ('Software Engineering', 'We build Python microservices.'),
    ('Customer Service', 'We run a 50-agent call center for an e-commerce brand.'),
    ('Surgical Team', 'We are a hospital OR team implementing WHO checklists.'),
]

for domain, context in domains:
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=80,
        messages=[{
            'role': 'user',
            'content': f'Context: {context}\n\nQuestion: How do I handle errors gracefully? (1 sentence answer)'
        }]
    )
    print(f'[{domain}]: {response.content[0].text.strip()}')
    print()

Contexte de l'audience

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

audiences = [
    'a 10-year-old who loves video games',
    'a first-year computer science university student',
    'a senior software architect with 15 years of experience'
]

for audience in audiences:
    response = client.chat.completions.create(
        model='gpt-4o',
        max_tokens=80,
        messages=[{
            'role': 'user',
            'content': f'Explain recursion in 2 sentences for {audience}.'
        }]
    )
    print(f'Audience: {audience}')
    print(response.choices[0].message.content.strip())
    print()

Contexte de l'objectif

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

goals = [
    'Internal documentation for our own engineering team',
    'A sales one-pager to send cold prospects who know nothing about our product',
    'A support article for existing customers who are confused about the feature'
]

for goal in goals:
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=100,
        messages=[{
            'role': 'user',
            'content': (
                f'Goal: {goal}\n\n'
                f'Task: Write 2 sentences about our new AI-powered search feature '
                f'that finds relevant documents from a knowledge base.'
            )
        }]
    )
    print(f'Goal: {goal}')
    print(response.content[0].text.strip())
    print()

Contexte des contraintes

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{
        'role': 'user',
        'content': (
            'Suggest 3 ways to cache API responses in a Python backend.\n\n'
            'Constraint context:\n'
            '- Python 3.11, stdlib only (no Redis, no Memcached, no third-party libraries)\n'
            '- The backend is a single process (no distributed cache needed)\n'
            '- Cache must expire after 5 minutes automatically\n'
            '- Solution must work on a server without internet access'
        )
    }]
)
print(response.choices[0].message.content)

Le modèle devine quand le contexte manque

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

# Demonstrate: same question, very different useful answers with context
without_context = 'How should I structure my data?'

with_full_context = (
    'Domain: mobile gaming backend, 10M daily active users.\n'
    'Technology: Python FastAPI + PostgreSQL + Redis.\n'
    'Goal: store player inventory items (weapon skins, power-ups) with fast reads.\n'
    'Constraint: reads happen 50x more than writes; schema changes are expensive.\n\n'
    'How should I structure my data? Give 2 options with tradeoffs.'
)

response = client.messages.create(
    model='claude-opus-4-5',
    max_tokens=300,
    messages=[{'role': 'user', 'content': with_full_context}]
)
print(response.content[0].text)

Quand chaque type de contexte est le plus important

import openai

client = openai.OpenAI(api_key='sk-your-key-here')

# Context selection for a creative task
response = client.chat.completions.create(
    model='gpt-4o',
    messages=[{
        'role': 'user',
        'content': (
            'Domain: luxury skincare brand (eco-conscious, premium, women 35-55).\n'
            'Goal: Valentine\'s Day campaign headline — will run on Instagram and in email.\n'
            'Constraint: must not mention price, discount, or sale. Max 8 words.\n\n'
            'Generate 5 headline options.'
        )
    }]
)
print(response.choices[0].message.content)

Contexte et pollution de l'instruction

import anthropic

client = anthropic.Anthropic(api_key='sk-ant-your-key-here')

# Polluted context (too much irrelevant info)
polluted = (
    'I started my company in 2019. We have 12 employees. We use Slack. '
    'Our office is in Berlin. We had a good Q3. Our CEO is named Thomas. '
    'We have a dog-friendly office. We use Python for our backend. '
    'Last year we moved to a new CRM. We sponsor a local football team.\n\n'
    'Write a 2-sentence company bio for our website.'
)

# Clean context (only relevant info)
clean = (
    'Company: B2B SaaS, Berlin, founded 2019, 12 employees. '
    'Product: Python-based CRM for mid-market sales teams.\n\n'
    'Write a 2-sentence company bio for our website.'
)

for label, prompt in [('POLLUTED', polluted), ('CLEAN', clean)]:
    response = client.messages.create(
        model='claude-opus-4-5',
        max_tokens=100,
        messages=[{'role': 'user', 'content': prompt}]
    )
    print(f'--- {label} ---')
    print(response.content[0].text.strip())
    print()

Faire du contexte une pratique standard

Context: [domain, technology, role, situation]
Audience: [who reads/uses the output]
Goal: [what this will accomplish]
Constraints: [what must/must not be included]
Task: [the actual instruction]

# Standard context template implementation
def build_prompt(context, audience, goal, constraints, task):
    sections = []
    if context:     sections.append(f'Context: {context}')
    if audience:    sections.append(f'Audience: {audience}')
    if goal:        sections.append(f'Goal: {goal}')
    if constraints: sections.append(f'Constraints: {constraints}')
    sections.append(f'Task: {task}')
    return '\n'.join(sections)

prompt = build_prompt(
    context='Python open-source project, MIT license, 2,000 GitHub stars',
    audience='New contributors who know Python but are unfamiliar with our codebase',
    goal='Help them submit their first pull request within 30 minutes of reading',
    constraints='Max 400 words. No command-line flags beyond git basics. No Docker.',
    task='Write a Getting Started contributing guide.'
)
print(prompt)

Vérification des connaissances

Le contexte dans les instructions à l'IA — récapitulatif