开始使用
本主题提供了创建、填充和发布数据集的端到端工作流程。
数据集架构
每个数据集schemaType在创建时都声明一个。 AgentCore 在接受每个示例之前，会根据声明的架构对其进行验证。支持两种架构类型：
-
AGENTCORE_EVALUATION_PREDEFINED_V1 — 用于根据预先写好的对话回合测试代理。必填字段：scenario_id，turns（非空列表；每回合必须包含input）。
-
AGENTCORE_EVALUATION_SIMULATION_V1 — 用于生成合成对话。必填字段：scenario_id、actor_profile（带有必填context和的对象goal）、input。
有关完整的架构字段定义、示例和实况映射，请参阅数据集架构。
End-to-end 工作流程
以下示例演示了完整的数据集生命周期：创建、添加示例、列出示例、发布版本和清理。
例
- AgentCore CLI
-
-
# 1. Create dataset
agentcore add dataset --name my_eval_dataset \
--schema-type AGENTCORE_EVALUATION_PREDEFINED_V1
# 2. Add your scenarios to the JSONL file
# File: agentcore/datasets/my_eval_dataset.jsonl
# 3. Deploy to create the dataset and sync examples
agentcore deploy
# 4. Publish version 1
agentcore dataset publish-version --name my_eval_dataset
# 5. Check status (shows versions and example count)
agentcore status --type dataset
# 6. Download a published version to local file
agentcore dataset download --name my_eval_dataset --version 1
# 7. Cleanup
agentcore remove dataset --name my_eval_dataset
agentcore deploy
- AgentCore SDK
-
-
from bedrock_agentcore.evaluation import DatasetClient
client = DatasetClient(region_name="us-west-2")
# 1. Create dataset (polls until ACTIVE)
ds = client.create_dataset_and_wait(
datasetName="my_eval_dataset",
schemaType="AGENTCORE_EVALUATION_PREDEFINED_V1",
source={
"inlineExamples": {
"examples": [
{
"scenario_id": "TC-01",
"turns": [{"input": "What is my balance?", "expected_response": "Your balance is $50."}],
"assertions": ["Response includes a dollar amount"],
}
]
}
},
)
dataset_id = ds["datasetId"]
print(f"Created: {dataset_id}, status={ds['status']}")
# 2. Add more examples
ds = client.add_examples_and_wait(
datasetId=dataset_id,
source={
"inlineExamples": {
"examples": [
{"scenario_id": "TC-02", "turns": [{"input": "Transfer $100", "expected_response": "Transfer complete."}]}
]
}
},
)
print(f"Example count: {ds['exampleCount']}")
# 3. List examples
resp = client.list_dataset_examples(datasetId=dataset_id)
for example in resp["examples"]:
print(f" {example['exampleId']}: {example['scenario_id']}")
# 4. Publish version 1
ds = client.create_dataset_version_and_wait(datasetId=dataset_id)
print(f"Published, draftStatus: {ds.get('draftStatus')}")
# 5. List versions
resp = client.list_dataset_versions(datasetId=dataset_id)
for v in resp["versions"]:
print(f" Version {v['datasetVersion']}: {v['exampleCount']} examples")
# 6. Cleanup
client.delete_dataset_and_wait(datasetId=dataset_id)
