Generieren von Antworten mit der ChatCompletions-API

Abgeschlossen

Die OpenAI ChatCompletions-API wird häufig für generative KI-Modelle und -Plattformen verwendet. Obwohl die Antwort-API für die neue Projektentwicklung empfohlen wird, treten wahrscheinlich Szenarien auf, in denen die ChatCompletions-API für die Codewartung plattformübergreifender Kompatibilität nützlich ist.

Senden eines Prompts

Die ChatCompletions-API verwendet Sammlungen von Nachrichtenobjekten im JSON-Format, um Eingabeaufforderungen zu kapseln:

completion = openai_client.chat.completions.create(
    model="gpt-4o",  # Your model deployment name
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "When was Microsoft founded?"}
    ]
)

print(completion.choices[0].message.content)

Beibehalten des Unterhaltungskontexts

Im Gegensatz zur Antwort-API stellt die ChatCompletins-API kein feature für die statusbehaftete Antwortnachverfolgung bereit. Um den Unterhaltungskontext beizubehalten, müssen Sie Code schreiben, um vorherige Eingabeaufforderungen und Antworten manuell nachzuverfolgen.

# Initial messages
conversation_messages=[
    {
        "role": "system",
        "content": "You are a helpful AI assistant that answers questions and provides information."
    }
]

# Add the first user message
conversation_messages.append(
    {"role": "user",
    "content": "When was Microsoft founded?"}
)

# Get a completion
completion = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=conversation_messages
)
assistant_message = completion.choices[0].message.content
print("Assistant:", assistant_text)

# Append the response to the conversation
conversation_messages.append(
    {"role": "assistant", "content": assistant_text}
)

# Add the next user message
conversation_messages.append(
    {"role": "user",
    "content": "Who founded it?"}
)

# Get a completion
completion = openai_client.chat.completions.create(
    model="gpt-4o",
    messages=conversation_messages
)
assistant_message = completion.choices[0].message.content
print("Assistant:", assistant_text)

# and so on...

In einer echten Anwendung wird die Unterhaltung wahrscheinlich in einer Schleife implementiert; So:

# Initial messages
conversation_messages=[
    {
        "role": "system",
        "content": "You are a helpful AI assistant that answers questions and provides information."
    }
]

# Loop until the user wants to quit
print("Assistant: Enter a prompt (or type 'quit' to exit)")
while True:
    input_text = input('\nYou: ')
    if input_text.lower() == "quit":
        print("Assistant: Goodbye!")
        break

    # Add the user message
    conversation_messages.append(
        {"role": "user",
        "content": input_text}
    )

    # Get a completion
    completion = openai_client.chat.completions.create(
        model="gpt-4o",
        messages=conversation_messages
    )
    assistant_message = completion.choices[0].message.content
    print("\nAssistant:", assistant_message)
    
    # Append the response to the conversation
    conversation_messages.append(
        {"role": "assistant", "content": assistant_message}
    )

Die Ausgabe dieses Beispiels sieht in etwa so aus:

Assistant: Enter a prompt (or type 'quit' to exit)

You: When was Microsoft founded?

Assistant: Microsoft was founded on April 4, 1975 in Albuquerque, New Mexico, USA.

You: Who founded it?

Assistant: Microsoft was founded by Bill Gates and Paul Allen.

You: quit

Assistant: Goodbye!

Jeder neue Benutzer Prompt und Abschluss wird der Kommunikation hinzugefügt, und die gesamte Historie der Kommunikation wird in jeder Runde gesendet.

Die ChatCompletions-API ist zwar nicht so umfassend wie die Antwort-API, aber im generativen KI-Modellökosystem gut etabliert, daher ist es hilfreich, mit der API vertraut zu sein.