여러 프롬프트를 함께 최적화

복잡한 에이전트 시스템에서 여러 프롬프트를 함께 연결했을 수 있습니다. GEPA가 각 프롬프트를 고려하고 최적화할 수 있도록 이러한 모든 프롬프트를 함께 제공할 수 있습니다.

종속성 설치

%pip install --upgrade mlflow databricks-sdk dspy openai
dbutils.library.restartPython()

기본 프롬프트를 사용하여 두 가지 기본 함수 설정

import mlflow
from mlflow.genai.scorers import Correctness
from databricks.sdk import WorkspaceClient
from mlflow.genai.optimize import GepaPromptOptimizer

w = WorkspaceClient()
openai_client = w.serving_endpoints.get_open_ai_client()

catalog = ""
schema = ""

plan_prompt_name = f"{catalog}.{schema}.plan"
answer_prompt_name = f"{catalog}.{schema}.answer"

# Register multiple prompts
plan_prompt = mlflow.genai.register_prompt(
    name=plan_prompt_name,
    template="Make a plan to classify {{query}}.",
)
answer_prompt = mlflow.genai.register_prompt(
    name=answer_prompt_name,
    template="classify {{query}} following the plan: {{plan}}",
)

def predict_fn(query: str) -> str:
    plan_prompt = mlflow.genai.load_prompt(f"prompts:/{plan_prompt_name}/1")
    completion = openai_client.chat.completions.create(
        model="databricks-gpt-5",  # strong model
        messages=[{"role": "user", "content": plan_prompt.format(query=query)}],
    )
    plan = completion.choices[0].message.content

    answer_prompt = mlflow.genai.load_prompt(f"prompts:/{answer_prompt_name}/1")
    completion = openai_client.chat.completions.create(
        model="databricks-gpt-5-nano",  # cost efficient model
        messages=[
            {
                "role": "user",
                "content": answer_prompt.format(query=query, plan=plan),
            }
        ],
    )
    return completion.choices[0].message.content

모델을 있는 그대로 테스트

from IPython.display import Markdown

output = predict_fn("The emergence of HIV as a chronic condition means that people living with HIV are required to take more responsibility for the self-management of their condition , including making physical , emotional and social adjustments .")

Markdown(output)

dataset = [
    {
        "inputs": {"query": "The emergence of HIV as a chronic condition means that people living with HIV are required to take more responsibility for the self-management of their condition , including making physical , emotional and social adjustments ."},
        "outputs": {"response": "BACKGROUND"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    },
    {
        "inputs": {"query": "This paper describes the design and evaluation of Positive Outlook , an online program aiming to enhance the self-management skills of gay men living with HIV ."},
        "outputs": {"response": "BACKGROUND"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    },
    {
        "inputs": {"query": "This study is designed as a randomised controlled trial in which men living with HIV in Australia will be assigned to either an intervention group or usual care control group ."},
        "outputs": {"response": "METHODS"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    },
    {
        "inputs": {"query": "The intervention group will participate in the online group program ` Positive Outlook ' ."},
        "outputs": {"response": "METHODS"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    },
    {
        "inputs": {"query": "The program is based on self-efficacy theory and uses a self-management approach to enhance skills , confidence and abilities to manage the psychosocial issues associated with HIV in daily life ."},
        "outputs": {"response": "METHODS"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    },
    {
        "inputs": {"query": "Participants will access the program for a minimum of 90 minutes per week over seven weeks ."},
        "outputs": {"response": "METHODS"},
        "expectations": {"expected_facts": ["Classification label must be 'CONCLUSIONS', 'RESULTS', 'METHODS', 'OBJECTIVE', 'BACKGROUND'"]}
    }
]


# Optimize both
result = mlflow.genai.optimize_prompts(
    predict_fn=predict_fn,
    train_data=dataset,
    prompt_uris=[plan_prompt.uri, answer_prompt.uri],
    optimizer=GepaPromptOptimizer(reflection_model="databricks:/databricks-gemini-2-5-pro"),
    scorers=[Correctness(model="databricks:/databricks-claude-sonnet-4-5")],
)

# Access optimized prompts
optimized_plan = result.optimized_prompts[0]
optimized_answer = result.optimized_prompts[1]

새 프롬프트 로드 및 다시 테스트

프롬프트의 모양을 확인하고 예측 함수에 로드하여 모델이 수행하는 성능이 어떻게 다른지 확인합니다.

plan_prompt = mlflow.genai.load_prompt(f"prompts:/{plan_prompt_name}/4")
Markdown(plan_prompt.template)

answer_prompt = mlflow.genai.load_prompt(f"prompts:/{answer_prompt_name}/4")
Markdown(answer_prompt.template)

from IPython.display import Markdown


def predict_fn(query: str) -> str:
    plan_prompt = mlflow.genai.load_prompt(f"prompts:/{plan_prompt_name}/4")
    completion = openai_client.chat.completions.create(
        model="databricks-gpt-5",  # strong model
        messages=[{"role": "user", "content": plan_prompt.format(query=query)}],
    )
    plan = completion.choices[0].message.content

    answer_prompt = mlflow.genai.load_prompt(f"prompts:/{answer_prompt_name}/4")
    completion = openai_client.chat.completions.create(
        model="databricks-gpt-5-nano",  # cost efficient model
        messages=[
            {
                "role": "user",
                "content": answer_prompt.format(query=query, plan=plan),
            }
        ],
    )
    return completion.choices[0].message.content

output = predict_fn("Participants will access the program for a minimum of 90 minutes per week over seven weeks")

Markdown(output)

GEPA 및 GPT-OSS 20B를 사용한 프롬프트 최적화

노트북 받기