Ingénierie et perfectionnement itératifs des prompts

Ingénierie des prompts avec l'API OpenAI

Fouad Trad

Machine Learning Engineer

Ingénierie itérative des prompts

Aucun prompt ne peut être parfait dès le début
L'ingénierie des prompts est un processus itératif :
- Créer une invite
- L'intégrer au modèle
- Observer et analyser les résultats
- Réitérer afin d'améliorer le prompt

Image representing a cycle. A blue circle with four arrows feeding into each other.

Affiner les prompts

Prompt initial

prompt = "Generate an Excel sheet containing
five student names and their grades"

print(get_response(prompt))

I'm sorry, but as a text-based AI, I am 
unable to directly provide an Excel sheet. 
However, I can help you generate a sample
representation of the data you requested.

Prompt affiné

prompt = "Generate a table that I can copy 
to Excel, containing five student names and 
their grades"
print(get_response(prompt))

| Student Name | Grade |
|--------------|-------|
| Student 1    |  90   |
| Student 2    |  85   |
| Student 3    |  95   |
| Student 4    |  88   |
| Student 5    |  92   |

Exemple : analyse d'une fonction Python

code = '''
def calculate_rectangle_area(length, width):
    area = length * width
    return area
'''

Image showing a rectangle and the formula to compute its area given its width and length. Area = length*width.

Exemple : prompt initial

prompt = f"""
  Analyze the code delimited by triple backticks with one sentence
  ```{code}```.
"""
print(get_response(prompt))

The code calculates the area of a rectangle based on its length and width.

Exemple : affinement du prompt

Nous modifions le prompt pour obtenir le langage de programmation

prompt = f"""
  Analyze the code delimited by triple backticks and provide its programming 
  language with one sentence 
  ```{code}```.
"""
print(get_response(prompt))

The provided code is a function written in Python that calculates the area of a 
rectangle based on its length and width.

Exemple : affinement du prompt

Nous modifions le prompt pour obtenir un résultat structuré

prompt = f"""
  For the function delimited by triple backticks, provide in a structured format
  the following:
  - description: one sentence short description
  - language: the programming language used
  - input: the inputs to the function
  - output: the output returned by the function
  ```{code}```.
"""
print(get_response(prompt))

Exemple : affinement du prompt

description: This function calculates the area of a rectangle.

language: Python

input:
 - length: The length of the rectangle.
 - width: The width of the rectangle.

output:
 - area: The calculated area of the rectangle, which is the product of the length 
 and width.

Affinement du prompt few-shot

Classification des descriptions météorologiques

Prompt initial

prompt = """
Clear skies and a gentle breeze. -> Sunny
Heavy rain and thunderstorms expected. -> Rainy
Fresh snowfall with freezing temperatures. -> 
"""
print(get_response(prompt))

Snowy

Affinement du prompt few-shot

Classification des descriptions météorologiques

Prompt initial

prompt = """
Clear skies and a gentle breeze. -> Sunny
Heavy rain and thunderstorms expected. -> Rainy
The wind of change brought a refreshing breeze to the company's operations. -> 
"""
print(get_response(prompt))

Windy

Affinement du prompt few-shot

Prompt affiné

prompt = """
Clear skies and a gentle breeze. -> Sunny
Heavy rain and thunderstorms expected. -> Rainy
The political climate in the country was stormy -> Unknown
The wind of change brought a refreshing breeze to the company's operations. -> 
"""
print(get_response(prompt))

Unknown

Amélioration des divers types de prompts

Few-shot prompts : affiner les exemples
Prompts multi-étapes : affiner les étapes de guidage
Prompts par chaîne de pensée et cohérence propre : affiner la description du problème

Image showing three rows with toggles with a hand adjusting the toggles to reflect refinement.

Passons à la pratique !

Ingénierie des prompts avec l'API OpenAI