审查工具调用¶

人工参与 (HIL) 交互对于 agentic 系统至关重要。一种常见的模式是在某些工具调用之后添加一些人工参与步骤。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

执行 SQL 的工具调用，然后将由该工具运行
生成摘要的工具调用，然后将摘要保存到图的状态

请注意，无论是否实际调用工具，使用工具调用都很常见。

您通常可能希望在此处进行一些不同的交互

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，然后将其传递回代理而不是继续

我们可以使用 LangGraph 中的断点来实现这一点：断点允许我们在特定步骤之前中断图的执行。在此断点处，我们可以手动更新图状态，采取上述三种选项之一

设置¶

我们不会展示我们托管的图的完整代码，但如果您想查看，可以在这里查看。一旦托管了此图，我们就可以调用它并等待用户输入。

SDK 初始化¶

首先，我们需要设置我们的客户端，以便我们可以与托管图进行通信

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 '{}'

无审查示例¶

让我们看一个不需要审查的示例（因为没有调用工具）

PythonJavascriptCURL

input = { 'messages':[{ "role":"user", "content":"hi!" }] }

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

const input = { "messages": [{ "role": "user", "content": "hi!" }] };

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

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\": \"hi!\"}]},
   \"stream_mode\": [
     \"updates\"
   ],
   \"interrupt_before\": [\"action\"]
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': 'hi!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '39c51f14-2d5c-4690-883a-d940854b1845', 'example': False}]}
{'messages': [{'content': 'hi!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '39c51f14-2d5c-4690-883a-d940854b1845', 'example': False}, {'content': [{'text': "Hello! Welcome. How can I assist you today? Is there anything specific you'd like to know or any information you're looking for?", 'type': 'text', 'index': 0}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-d65e07fb-43ff-4d98-ab6b-6316191b9c8b', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 355, 'output_tokens': 31, 'total_tokens': 386}}]}

如果我们检查状态，我们可以看到它已完成

PythonJavascriptCURL

state = await client.threads.get_state(thread["thread_id"])

print(state['next'])

const state = await client.threads.getState(thread["thread_id"]);

console.log(state.next);

curl --request GET \
    --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state | jq -c '.next'

输出

[]

批准工具示例¶

现在让我们看看批准工具调用是什么样的。请注意，我们不需要将中断传递给我们的流式调用，因为该图（在这里定义）在 human_review_node 之前已经使用中断编译。

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,
):
    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,
  }
);

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?\"}]}
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '54e19d6e-89fa-44fb-b92c-12e7dd4ddf08', 'example': False}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '54e19d6e-89fa-44fb-b92c-12e7dd4ddf08', 'example': False}, {'content': [{'text': "Certainly! I can help you check the weather in San Francisco. To get this information, I'll use the weather search function. Let me do that for you right away.", 'type': 'text', 'index': 0}, {'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-45a6b6c3-ac69-42a4-8957-d982203d6392', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 90, 'total_tokens': 450}}]}

如果我们现在检查，我们可以看到它正在等待人工审查

PythonJavascriptCURL

state = await client.threads.get_state(thread["thread_id"])

print(state['next'])

const state = await client.threads.getState(thread["thread_id"]);

console.log(state.next);

curl --request GET \
    --url <DELPOYMENT_URL>/threads/<THREAD_ID>/state | jq -c '.next'

输出

['human_review_node']

要批准工具调用，我们可以继续线程，无需编辑。为此，我们只需创建一个新的运行，无需输入。

PythonJavascriptCURL

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=None,
    stream_mode="values",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

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

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\"
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '54e19d6e-89fa-44fb-b92c-12e7dd4ddf08', 'example': False}, {'content': [{'text': "Certainly! I can help you check the weather in San Francisco. To get this information, I'll use the weather search function. Let me do that for you right away.", 'type': 'text', 'index': 0}, {'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-45a6b6c3-ac69-42a4-8957-d982203d6392', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 90, 'total_tokens': 450}}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '826cd0f2-9cc6-46f0-b7df-daa6a05d13d2', 'tool_call_id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'artifact': None, 'status': 'success'}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '54e19d6e-89fa-44fb-b92c-12e7dd4ddf08', 'example': False}, {'content': [{'text': "Certainly! I can help you check the weather in San Francisco. To get this information, I'll use the weather search function. Let me do that for you right away.", 'type': 'text', 'index': 0}, {'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-45a6b6c3-ac69-42a4-8957-d982203d6392', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 90, 'total_tokens': 450}}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '826cd0f2-9cc6-46f0-b7df-daa6a05d13d2', 'tool_call_id': 'toolu_015yrR3GMDXe6X8m2p9CsEDN', 'artifact': None, 'status': 'success'}, {'content': [{'text': "\n\nGreat news! The weather in San Francisco is sunny today. It's a beautiful day in the city by the bay. Is there anything else you'd like to know about the weather or any other information I can help you with?", 'type': 'text', 'index': 0}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-5d5fd0f1-a939-447e-801a-9aaa812322d3', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 464, 'output_tokens': 50, 'total_tokens': 514}}]}

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="values",
):
    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: "values",
  }
);

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?\"}]}
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': 'cec11391-84da-464b-bd2a-bd4f0d93b9ee', 'example': False}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': 'cec11391-84da-464b-bd2a-bd4f0d93b9ee', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01SunSpDurNfcnXppWLPrtjC', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-6326da9f-6061-4e12-8586-482e32ab4cab', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01SunSpDurNfcnXppWLPrtjC', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}]}

为此，我们首先需要更新状态。我们可以通过传入一条消息来完成此操作，该消息具有我们要覆盖的消息的相同 ID。这将具有替换旧消息的效果。请注意，这仅在由于我们使用的reducer 可以替换具有相同 ID 的消息时才有可能 - 在此处阅读更多相关信息。

PythonJavascriptCURL

# To get the ID of the message we want to replace, we need to fetch the current state and find it there.
state = await client.threads.get_state(thread['thread_id'])
print("Current State:")
print(state['values'])
print("\nCurrent Tool Call ID:")
current_content = state
current_id = state
tool_call_id = state
print(tool_call_id)

# We now need to construct a replacement tool call.
# We will change the argument to be `San Francisco, USA`
# Note that we could change any number of arguments or tool names - it just has to be a valid one
new_message = {
    "role": "assistant", 
    "content": current_content,
    "tool_calls": [
        {
            "id": tool_call_id,
            "name": "weather_search",
            "args": {"city": "San Francisco, USA"}
        }
    ],
    # This is important - this needs to be the same as the message you replacing!
    # Otherwise, it will show up as a separate message
    "id": current_id
}
await client.threads.update_state(
    # This is the config which represents this thread
    thread['thread_id'], 
    # This is the updated value we want to push
    {"messages": [new_message]}, 
    # We push this update acting as our human_review_node
    as_node="human_review_node"
)

print("\nResuming Execution")
# Let's now continue executing from here
async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=None,
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const state = await client.threads.getState(thread.thread_id);
console.log("Current State:");
console.log(state.values);

console.log("\nCurrent Tool Call ID:");
const lastMessage = state.values.messages[state.values.messages.length - 1];
const currentContent = lastMessage.content;
const currentId = lastMessage.id;
const toolCallId = lastMessage.tool_calls[0].id;
console.log(toolCallId);

// Construct a replacement tool call
const newMessage = {
  role: "assistant",
  content: currentContent,
  tool_calls: [
    {
      id: toolCallId,
      name: "weather_search",
      args: { city: "San Francisco, USA" }
    }
  ],
  // Ensure the ID is the same as the message you're replacing
  id: currentId
};

await client.threads.updateState(
  thread.thread_id,  // Thread ID
  {
    values: { "messages": [newMessage] },  // Updated message
    asNode: "human_review_node"
  }  // Acting as human_review_node
);

console.log("\nResuming Execution");
// Continue executing from here
const streamResponseResumed = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: null,
  }
);

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

curl --request POST \
--url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state \
--header 'Content-Type: application/json' \
--data "{
    \"values\": { \"messages\": [$(curl --request GET \
        --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state |
        jq -c '{
        role: "assistant",
        content: .values.messages[-1].content,
        tool_calls: [
            {
            id: .values.messages[-1].tool_calls[0].id,
            name: "weather_search",
            args: { city: "San Francisco, USA" }
            }
        ],
        id: .values.messages[-1].id
        }')
    ]},
    \"as_node\": \"human_review_node\"
}" && echo "Resuming Execution" && curl --request POST \
--url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
--header 'Content-Type: application/json' \
--data '{
"assistant_id": "agent"
}' | \
sed 's/\r$//' | \
awk '
/^event:/ {
    if (data_content != "" && event_type != "metadata") {
        print data_content "\n"
    }
    sub(/^event: /, "", $0)
    event_type = $0
    data_content = ""
}
/^data:/ {
    sub(/^data: /, "", $0)
    data_content = $0
}
END {
    if (data_content != "" && event_type != "metadata") {
        print data_content "\n"
    }
}
'

输出

Current State:
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '8713d1fa-9b26-4eab-b768-dafdaac70590', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-ede13f26-daf5-4d8f-817a-7611075bbcf1', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}]}

Current Tool Call ID:
toolu_01VzagzsUGZsNMwW1wHkcw7h

Resuming Execution
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '8713d1fa-9b26-4eab-b768-dafdaac70590', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-ede13f26-daf5-4d8f-817a-7611075bbcf1', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': None}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '7fc7d463-66bf-4555-9929-6af483de169b', 'tool_call_id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'artifact': None, 'status': 'success'}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '8713d1fa-9b26-4eab-b768-dafdaac70590', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-ede13f26-daf5-4d8f-817a-7611075bbcf1', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': None}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '7fc7d463-66bf-4555-9929-6af483de169b', 'tool_call_id': 'toolu_01VzagzsUGZsNMwW1wHkcw7h', 'artifact': None, 'status': 'success'}, {'content': [{'text': "\n\nBased on the search result, the weather in San Francisco is sunny! It's a beautiful day in the city by the bay. Is there anything else you'd like to know about the weather or any other information I can help you with?", 'type': 'text', 'index': 0}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-d90ce97a-39f9-4330-985e-67c5f351a0c5', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 455, 'output_tokens': 52, 'total_tokens': 507}}]}

向工具调用提供反馈¶

有时，您可能不想执行工具调用，但您也可能不想要求用户手动修改工具调用。在这种情况下，最好从用户那里获得自然语言反馈。然后，您可以将这些反馈作为工具调用的模拟结果插入。

有多种方法可以做到这一点

您可以向状态添加新消息（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 表示反馈两者类似之处在于它们都涉及向状态添加消息。主要区别在于 human_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，我们将仅添加一个表示反馈的工具调用。让我们看看实际操作！

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,
):
    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,
  }
);

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?\"}]}
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': 'c80f13d0-674d-4233-b6a0-3940509d3cf3', 'example': False}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': 'c80f13d0-674d-4233-b6a0-3940509d3cf3', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_016XyTdFA8NuPWeLyZPSzoM3', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-4911ac27-3d7c-4edf-a3ca-c2908e3922eb', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_016XyTdFA8NuPWeLyZPSzoM3', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}]}

为此，我们首先需要更新状态。我们可以通过传入一条消息来完成此操作，该消息具有我们要响应的工具调用的相同工具调用 ID。请注意，这是一个与上述不同* 的 ID

PythonJavascriptCURL

# To get the ID of the message we want to replace, we need to fetch the current state and find it there.
state = await client.threads.get_state(thread['thread_id'])
print("Current State:")
print(state['values'])
print("\nCurrent Tool Call ID:")
tool_call_id = state
print(tool_call_id)

# We now need to construct a replacement tool call.
# We will change the argument to be `San Francisco, USA`
# Note that we could change any number of arguments or tool names - it just has to be a valid one
new_message = {
    "role": "tool", 
    # This is our natural language feedback
    "content": "User requested changes: pass in the country as well",
    "name": "weather_search",
    "tool_call_id": tool_call_id
}
await client.threads.update_state(
    # This is the config which represents this thread
    thread['thread_id'], 
    # This is the updated value we want to push
    {"messages": [new_message]}, 
    # We push this update acting as our human_review_node
    as_node="human_review_node"
)

print("\nResuming execution")
# Let's now continue executing from here
async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=None,
    stream_mode="values",
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const state = await client.threads.getState(thread.thread_id);
console.log("Current State:");
console.log(state.values);

console.log("\nCurrent Tool Call ID:");
const lastMessage = state.values.messages[state.values.messages.length - 1];
const toolCallId = lastMessage.tool_calls[0].id;
console.log(toolCallId);

// Construct a replacement tool call
const newMessage = {
  role: "tool",
  content: "User requested changes: pass in the country as well",
  name: "weather_search",
  tool_call_id: toolCallId,
};

await client.threads.updateState(
  thread.thread_id,  // Thread ID
  {
    values: { "messages": [newMessage] },  // Updated message
    asNode: "human_review_node"
  }  // Acting as human_review_node
);

console.log("\nResuming Execution");
// Continue executing from here
const streamResponseEdited = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: null,
    streamMode: "values",
    interruptBefore: ["action"],
  }
);

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

curl --request POST \
--url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state \
--header 'Content-Type: application/json' \
--data "{
    \"values\": { \"messages\": [$(curl --request GET \
        --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state |
        jq -c '{
        role: "tool",
        content: "User requested changes: pass in the country as well",
        name: "get_weather",
        tool_call_id: .values.messages[-1].id.tool_calls[0].id
        }')
    ]},
    \"as_node\": \"human_review_node\"
}" && echo "Resuming Execution" && curl --request POST \
--url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
--header 'Content-Type: application/json' \
--data '{
"assistant_id": "agent"
}' | \
sed 's/\r$//' | \
awk '
/^event:/ {
    if (data_content != "" && event_type != "metadata") {
        print data_content "\n"
    }
    sub(/^event: /, "", $0)
    event_type = $0
    data_content = ""
}
/^data:/ {
    sub(/^data: /, "", $0)
    data_content = $0
}
END {
    if (data_content != "" && event_type != "metadata") {
        print data_content "\n"
    }
}
'

输出

Current State:
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '3b2bbc38-d11b-49eb-80c0-c24a40dab5a8', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-c5a50900-abf5-4885-9cdb-da2bf0d892ac', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}]}

Current Tool Call ID:
toolu_01NNw18j57GEGPZvsa9f1wvX

Resuming execution
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '3b2bbc38-d11b-49eb-80c0-c24a40dab5a8', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-c5a50900-abf5-4885-9cdb-da2bf0d892ac', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}, {'content': 'User requested changes: pass in the country as well', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '787288be-213c-4fd3-8503-4a009bdb1b00', 'tool_call_id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'artifact': None, 'status': 'success'}, {'content': [{'text': '\n\nI apologize for the oversight. It seems the function requires additional information. Let me try again with a more specific request.', 'type': 'text', 'index': 0}, {'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco, USA"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-5c355a56-cfe3-4046-b49f-f5b09fc397ef', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 461, 'output_tokens': 83, 'total_tokens': 544}}]}

我们可以看到我们现在进入了另一个断点 - 因为它返回到模型并获得了对要调用的内容的全新预测。现在让我们批准这一个并继续

PythonJavascriptCURL

async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id,
    input=None,
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

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

for await (const chunk of streamResponseResumed) {
  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\"
 }" | \
 sed 's/\r$//' | \
 awk '
 /^event:/ {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
     sub(/^event: /, "", $0)
     event_type = $0
     data_content = ""
 }
 /^data:/ {
     sub(/^data: /, "", $0)
     data_content = $0
 }
 END {
     if (data_content != "" && event_type != "metadata") {
         print data_content "\n"
     }
 }
 '

输出

{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '3b2bbc38-d11b-49eb-80c0-c24a40dab5a8', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-c5a50900-abf5-4885-9cdb-da2bf0d892ac', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}, {'content': 'User requested changes: pass in the country as well', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '787288be-213c-4fd3-8503-4a009bdb1b00', 'tool_call_id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'artifact': None, 'status': 'success'}, {'content': [{'text': '\n\nI apologize for the oversight. It seems the function requires additional information. Let me try again with a more specific request.', 'type': 'text', 'index': 0}, {'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco, USA"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-5c355a56-cfe3-4046-b49f-f5b09fc397ef', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 461, 'output_tokens': 83, 'total_tokens': 544}}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '3b857482-bca2-4a73-a9ab-1f35a3e43e5f', 'tool_call_id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'artifact': None, 'status': 'success'}]}
{'messages': [{'content': "what's the weather in sf?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'human', 'name': None, 'id': '3b2bbc38-d11b-49eb-80c0-c24a40dab5a8', 'example': False}, {'content': [{'text': 'To get the weather information for San Francisco, I can use the weather_search function. Let me do that for you.', 'type': 'text', 'index': 0}, {'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-c5a50900-abf5-4885-9cdb-da2bf0d892ac', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco'}, 'id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 360, 'output_tokens': 80, 'total_tokens': 440}}, {'content': 'User requested changes: pass in the country as well', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '787288be-213c-4fd3-8503-4a009bdb1b00', 'tool_call_id': 'toolu_01NNw18j57GEGPZvsa9f1wvX', 'artifact': None, 'status': 'success'}, {'content': [{'text': '\n\nI apologize for the oversight. It seems the function requires additional information. Let me try again with a more specific request.', 'type': 'text', 'index': 0}, {'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'input': {}, 'name': 'weather_search', 'type': 'tool_use', 'index': 1, 'partial_json': '{"city": "San Francisco, USA"}'}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'tool_use', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-5c355a56-cfe3-4046-b49f-f5b09fc397ef', 'example': False, 'tool_calls': [{'name': 'weather_search', 'args': {'city': 'San Francisco, USA'}, 'id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 461, 'output_tokens': 83, 'total_tokens': 544}}, {'content': 'Sunny!', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'weather_search', 'id': '3b857482-bca2-4a73-a9ab-1f35a3e43e5f', 'tool_call_id': 'toolu_01YAbLBoKozJyRQnB8LUMpXC', 'artifact': None, 'status': 'success'}, {'content': [{'text': "\n\nGreat news! The weather in San Francisco is sunny today. Is there anything else you'd like to know about the weather or any other information I can help you with?", 'type': 'text', 'index': 0}], 'additional_kwargs': {}, 'response_metadata': {'stop_reason': 'end_turn', 'stop_sequence': None}, 'type': 'ai', 'name': None, 'id': 'run-6a857bb1-f65b-4b86-93d6-c025e003c777', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': {'input_tokens': 557, 'output_tokens': 38, 'total_tokens': 595}}]}