複数のプロンプトをまとめて最適化する
複雑なエージェント システムでは、複数のプロンプトを連結している場合があります。これらすべてのプロンプトをまとめて GEPA に提供し、各プロンプトを考慮して最適化することができます。
依存関係をインストールする
Python
%pip install --upgrade mlflow databricks-sdk dspy openai
dbutils.library.restartPython()
基本的なプロンプトで2つの基本機能を設定する
Python
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
モデルをそのままテストする
Python
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)
Python
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]
新しいプロンプトをロードして再度テストします
プロンプトがどのように見えるかを確認し、それを予測関数に読み込んで、モデルのパフォーマンスがどのように異なるかを確認します。
Python
plan_prompt = mlflow.genai.load_prompt(f"prompts:/{plan_prompt_name}/4")
Markdown(plan_prompt.template)
Python
answer_prompt = mlflow.genai.load_prompt(f"prompts:/{answer_prompt_name}/4")
Markdown(answer_prompt.template)
Python
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)