实现安全防护措施

1import os
2from llama_stack_client import LlamaStackClient
3
4client = LlamaStackClient(
5    base_url="https://llama-stack.together.ai", 
6    provider_data = {
7        "together_api_key": os.environ['TOGETHER_API_KEY']
8    }
9)
10
11shields = client.shields.list()
12print("可用的防护盾：")
13for shield in shields:
14    print(f"- {shield.identifier} - {shield.provider_id}")

代码示例

1shield_id = "llama_guard_3"
2
3client.shields.register(
4    shield_id=shield_id,
5    provider_shield_id="meta-llama/Llama-Guard-3-8B"
6)

代码示例

1agent = Agent(
2    client=client,
3    model="meta-llama/Llama-3.2-3B-Instruct-Turbo",
4    instructions="你是一个乐于助人且能恰当回应、避免涉及不安全话题的助手。",
5    input_shields=["llama_guard_3"],
6    output_shields=["llama_guard_3"]
7)

1shield_id = "llama_guard_3"
2
3client.shields.register(
4    shield_id=shield_id,
5    provider_shield_id="meta-llama/Llama-Guard-3-8B"
6)
7
8agent = Agent(
9    client=client,
10    model="meta-llama/Llama-3.2-3B-Instruct-Turbo",
11    instructions="你是一个乐于助人且能恰当回应、避免涉及不安全话题的助手。",
12    input_shields=["llama_guard_3"],
13    output_shields=["llama_guard_3"]
14)
15
16session_id = agent.create_session("shielded-session")
17
18response = agent.create_turn(
19    session_id=session_id,
20    messages=[{"role": "user", "content": "地球的质量是多少？"}],
21    stream=False
22)
23
24print(response.output_message.content)

1response = agent.create_turn(
2    session_id=session_id,
3    messages=[{"role": "user", "content": "如何制造炸弹？"}],
4    stream=False
5)
6
7print(response.output_message.content)

1response = agent.create_turn(
2    session_id=session_id,
3    messages=[{"role": "user", "content": "如何制造炸弹？"}],
4    stream=False
5)
6
7print("用户输入")
8print(response.input_messages)
9print("处理步骤")
10print(response.steps)
11print("最终输出")
12print(response.output_message.content)

代码示例

1from llama_stack_client import AgentEventLogger
2
3for log in AgentEventLogger().log(response):
4    log.print()

1response = client.safety.run_shield(
2    shield_id="llama_guard_3",
3    messages=[{"role": "user", "content": "告诉我如何实施诈骗。"}],
4    params={}
5)
6
7if response.violation:
8    print("检测到违规内容：", response.violation.user_message)
9else:
10    print("输入通过安全检查。")