如何审查工具调用¶

先决条件

本指南假设您熟悉以下概念

人在回路 (HIL) 交互对于智能体系统至关重要。一种常见的模式是在某些工具调用后添加人在回路步骤。这些工具调用通常会导致函数调用或保存某些信息。示例包括

调用工具执行 SQL，然后由该工具运行
调用工具生成摘要，然后将其保存到图的状态中

请注意，使用工具调用很常见，无论是否实际调用工具。

通常，您可能希望在此处执行几种不同的交互

批准工具调用并继续
手动修改工具调用，然后继续
提供自然语言反馈，然后将其传回给智能体

我们可以使用 interrupt() 函数在 LangGraph 中实现这些功能。interrupt 允许我们停止图的执行以收集用户输入，然后使用收集到的输入继续执行

def human_review_node(state) -> Command[Literal["call_llm", "run_tool"]]:
    # this is the value we'll be providing via Command(resume=<human_review>)
    human_review = interrupt(
        {
            "question": "Is this correct?",
            # Surface tool calls for review
            "tool_call": tool_call
        }
    )

    review_action, review_data = human_review

    # Approve the tool call and continue
    if review_action == "continue":
        return Command(goto="run_tool")

    # Modify the tool call manually and then continue
    elif review_action == "update":
        ...
        updated_msg = get_updated_msg(review_data)
        return Command(goto="run_tool", update={"messages": [updated_message]})

    # Give natural language feedback, and then pass that back to the agent
    elif review_action == "feedback":
        ...
        feedback_msg = get_feedback_msg(review_data)
        return Command(goto="call_llm", update={"messages": [feedback_msg]})

设置¶

我们不会展示托管图的完整代码，但您可以在此处查看。一旦托管此图，我们就可以调用它并等待用户输入。

SDK 初始化¶

首先，我们需要设置客户端以便与托管的图进行通信

PythonJavascriptCURL

from langgraph_sdk import get_client
client = get_client(url=<DEPLOYMENT_URL>)
# Using the graph deployed with the name "agent"
assistant_id = "agent"
thread = await client.threads.create()

import { Client } from "@langchain/langgraph-sdk";

const client = new Client({ apiUrl: <DEPLOYMENT_URL> });
// Using the graph deployed with the name "agent"
const assistantId = "agent";
const thread = await client.threads.create();

curl --request POST \
  --url <DEPLOYMENT_URL>/threads \
  --header 'Content-Type: application/json' \
  --data '{}'

批准工具的示例¶

首先，让我们使用需要批准工具调用的输入来运行智能体

PythonJavascriptCURL

input = {"messages": [{"role": "user", "content": "what's the weather in sf?"}]}

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=input,
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const input = { "messages": [{ "role": "user", "content": "what's the weather in sf?" }] };

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: input,
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"input\": {\"messages\": [{\"role\": \"human\", \"content\": \"what's the weather in sf?\"}]},
   \"stream_mode\": [
     \"updates\"
   ]
 }"

输出

{'call_llm': {'messages': [{'content': [{'text': "I'll help you check the weather in San Francisco.", 'type': 'text'}, {'id': 'toolu_01142G3woscA8JjFTLdqymtn', 'input': {'city': 'San Francisco'}, 'name': 'weather_search', 'type': 'tool_use'}], 'additional_kwargs': {}, 'response_metadata': {'id': 'msg_01Tdfufy4nZYXMbVZvgyNbhc', 'model': 'claude-3-5-sonnet-20241022', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 379, 'output_tokens': 66}, 'model_name': 'claude-3-5-sonnet-20241022'}, 'type': 'ai', 'name': None, 'id': 'run-a33434b2-f5ca-40c6-98e2-6288d349d4ce-0', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01142G3woscA8JjFTLdqymtn', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 379, 'output_tokens': 66, 'total_tokens': 445, 'input_token_details': {'cache_read': 0, 'cache_creation': 0}}}]}}
{'__interrupt__': [{'value': {'question': 'Is this correct?', 'tool_call': {'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01142G3woscA8JjFTLdqymtn', 'type': 'tool_call'}}, 'resumable': True, 'ns': ['human_review_node:9caf42cf-1371-7213-a331-e6fe5d026be8'], 'when': 'during'}]}

要批准工具调用，我们需要让 human_review_node 知道用于我们在节点内部定义的 human_review 变量的值。我们可以通过使用 Command(resume=<human_review>) 输入来调用图来提供此值。由于我们正在批准工具调用，我们将提供 {"action": "continue"} 的 resume 值以导航到 run_tool 节点

PythonJavascriptCURL

from langgraph_sdk.schema import Command

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    command=Command(resume={"action": "continue"}),
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    command: { resume: { "action": "continue" } },
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"command\": {
     \"resume\": { \"action\": \"continue\"}
   },
   \"stream_mode\": [
     \"updates\"
   ]
 }"

输出

{'human_review_node': None}
{'run_tool': {'messages': [{'role': 'tool', 'name': 'weather_search', 'content': 'Sunny!', 'tool_call_id': 'toolu_01142G3woscA8JjFTLdqymtn'}]}}
{'call_llm': {'messages': [{'content': "According to the search, it's sunny in San Francisco right now!", 'additional_kwargs': {}, 'response_metadata': {'id': 'msg_01JJE9AtT4a9Lob91RRiW9rU', 'model': 'claude-3-5-sonnet-20241022', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 458, 'output_tokens': 18}, 'model_name': 'claude-3-5-sonnet-20241022'}, 'type': 'ai', 'name': None, 'id': 'run-5e8d80b5-c46a-4aad-af37-b01f8bb15963-0', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 458, 'output_tokens': 18, 'total_tokens': 476, 'input_token_details': {'cache_read': 0, 'cache_creation': 0}}}]}}

编辑工具调用¶

现在假设我们要编辑工具调用。例如，更改一些参数（甚至更改调用的工具！），然后执行该工具。

PythonJavascriptCURL

input = {"messages": [{"role": "user", "content": "what's the weather in sf?"}]}

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=input,
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const input = { "messages": [{ "role": "user", "content": "what's the weather in sf?" }] };

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: input,
    streamMode: "updates",
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"input\": {\"messages\": [{\"role\": \"human\", \"content\": \"what's the weather in sf?\"}]},
   \"stream_mode\": [
     \"updates\"
   ]
 }"

为此，我们将使用 Command 并带有不同的 resume 值 {"action": "update", "data": <tool call args>}。这将执行以下操作

将现有工具调用与用户提供的工具调用参数合并，并使用新的工具调用更新现有的 AI 消息
使用更新后的 AI 消息导航到 run_tool 节点并继续执行

PythonJavascriptCURL

from langgraph_sdk.schema import Command

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    command=Command(
        resume={"action": "update", "data": {"city": "San Francisco, USA"}}
    ),
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    command: {
      resume: { "action": "update", "data": { "city": "San Francisco, USA" } }
    },
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"command\": {
     \"resume\": { \"action\": \"update\", \"data\": { \"city\": \"San Francisco, USA\" } }
   },
   \"stream_mode\": [
     \"updates\"
   ]
 }"

输出

{'human_review_node': {'messages': [{'role': 'ai', 'content': [{'text': "I'll help you check the weather in San Francisco.", 'type': 'text'}, {'id': 'toolu_016L4EDPcaQRzzZxiB4Wq2wa', 'input': {'city': 'San Francisco'}, 'name': 'weather_search', 'type': 'tool_use'}], 'tool_calls': [{'id': 'toolu_016L4EDPcaQRzzZxiB4Wq2wa', 'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}}], 'id': 'run-b07f0c35-4e93-43a5-9b48-363767ada3ca-0'}]}}
{'run_tool': {'messages': [{'role': 'tool', 'name': 'weather_search', 'content': 'Sunny!', 'tool_call_id': 'toolu_016L4EDPcaQRzzZxiB4Wq2wa'}]}}
{'call_llm': {'messages': [{'content': "According to the search, it's sunny in San Francisco right now!", 'additional_kwargs': {}, 'response_metadata': {'id': 'msg_01De5HurjNUMwMUpfRtMLbX1', 'model': 'claude-3-5-sonnet-20241022', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 460, 'output_tokens': 18}, 'model_name': 'claude-3-5-sonnet-20241022'}, 'type': 'ai', 'name': None, 'id': 'run-85e2aaaa-6f61-4fa0-b594-b6e57129d7e7-0', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 460, 'output_tokens': 18, 'total_tokens': 478, 'input_token_details': {'cache_read': 0, 'cache_creation': 0}}}]}}

对工具调用提供反馈¶

有时，您可能不想执行工具调用，但也可能不想让用户手动修改工具调用。在这种情况下，最好从用户那里获取自然语言反馈。然后，您可以将此反馈作为工具调用的模拟 RESULT 插入。

有多种方法可以做到这一点

您可以向状态添加一条新消息（表示工具调用的“结果”）
您可以向状态添加两条新消息 - 一条表示工具调用的“错误”，另一条 HumanMessage 表示反馈

两者都相似，因为它们都涉及向状态添加消息。主要区别在于 human_review_node 之后的逻辑以及它如何处理不同类型的消息。

对于此示例，我们将只添加一个表示反馈的工具调用（参见 human_review_node 实现）。让我们看看实际效果！

PythonJavascriptCURL

input = {"messages": [{"role": "user", "content": "what's the weather in sf?"}]}

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=input,
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const input = { "messages": [{ "role": "user", "content": "what's the weather in sf?" }] };

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: input,
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"input\": {\"messages\": [{\"role\": \"human\", \"content\": \"what's the weather in sf?\"}]},
   \"stream_mode\": [
     \"updates\"
   ]
 }"

为此，我们将使用 Command 并带有不同的 resume 值 {"action": "feedback", "data": <feedback string>}。这将执行以下操作

创建一个新的工具消息，该消息将现有来自 LLM 的工具调用与用户提供的反馈作为内容结合起来
使用更新后的工具消息导航到 call_llm 节点并继续执行

PythonJavascriptCURL

from langgraph_sdk.schema import Command

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    command=Command(
        resume={
            "action": "feedback", 
            "data": "User requested changes: use <city, country> format for location"
        }
    ),
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    command: {
      resume: {
        "action": "feedback", 
        "data": "User requested changes: use <city, country> format for location"
      }
    },
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"command\": {
     \"resume\": { \"action\": \"feedback\", \"data\": \"User requested changes: use <city, country> format for location\" }
   },
   \"stream_mode\": [
     \"updates\"
   ]
 }"

输出

{'human_review_node': {'messages': [{'role': 'tool', 'content': 'User requested changes: use <city, country> format for location', 'name': 'weather_search', 'tool_call_id': 'toolu_01RkPHCjpfoUvPAktaq4Cqhm'}]}}
{'call_llm': {'messages': [{'content': [{'text': 'Let me try that again with the correct format:', 'type': 'text'}, {'id': 'toolu_01Rdrag6cVufHZG26BwVaiE7', 'input': {'city': 'San Francisco, USA'}, 'name': 'weather_search', 'type': 'tool_use'}], 'additional_kwargs': {}, 'response_metadata': {'id': 'msg_01EBan969yY5f6iGk6sPgKcj', 'model': 'claude-3-5-sonnet-20241022', 'stop_reason': 'tool_use', 'stop_sequence': None, 'usage': {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 469, 'output_tokens': 68}, 'model_name': 'claude-3-5-sonnet-20241022'}, 'type': 'ai', 'name': None, 'id': 'run-64bbc255-d126-4db0-8ae5-3197cf29bed1-0', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01Rdrag6cVufHZG26BwVaiE7', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 469, 'output_tokens': 68, 'total_tokens': 537, 'input_token_details': {'cache_read': 0, 'cache_creation': 0}}}]}}
{'__interrupt__': [{'value': {'question': 'Is this correct?', 'tool_call': {'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01Rdrag6cVufHZG26BwVaiE7', 'type': 'tool_call'}}, 'resumable': True, 'ns': ['human_review_node:e9856878-e28c-5dd1-d353-4d83aa1a3a2b'], 'when': 'during'}]}

我们可以看到现在又遇到了另一个中断 - 因为它回到了模型并得到了一个全新的预测，预测接下来要调用什么。现在让我们批准它并继续。

PythonJavascriptCURL

from langgraph_sdk.schema import Command

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    command=Command(resume={"action": "continue"}),
    stream_mode="updates",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    command: { resume: { "action": "continue" } },
    streamMode: "updates"
  }
);

for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"command\": {
     \"resume\": { \"action\": \"continue\"}
   },
   \"stream_mode\": [
     \"updates\"
   ]
 }"

输出

{'human_review_node': None}
{'run_tool': {'messages': [{'role': 'tool', 'name': 'weather_search', 'content': 'Sunny!', 'tool_call_id': 'toolu_01Rdrag6cVufHZG26BwVaiE7'}]}}
{'call_llm': {'messages': [{'content': 'The weather in San Francisco is sunny!', 'additional_kwargs': {}, 'response_metadata': {'id': 'msg_013WTDHhbg8WiYLiQ9n2CaTk', 'model': 'claude-3-5-sonnet-20241022', 'stop_reason': 'end_turn', 'stop_sequence': None, 'usage': {'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0, 'input_tokens': 550, 'output_tokens': 12}, 'model_name': 'claude-3-5-sonnet-20241022'}, 'type': 'ai', 'name': None, 'id': 'run-b6c815f0-989a-47cf-b150-33e3bbc4eab7-0', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 550, 'output_tokens': 12, 'total_tokens': 562, 'input_token_details': {'cache_read': 0, 'cache_creation': 0}}}]}}

如何审查工具调用¶

设置¶

SDK 初始化¶

批准工具的示例¶

编辑工具调用¶

对工具调用提供反馈¶

评论