如何从先前状态进行重放和分支

使用 LangGraph Cloud，你可以返回到任何先前状态，既可以重新运行图以重现测试期间发现的问题，也可以从先前状态以不同于最初的方式进行分支。在本指南中，我们将快速举例说明如何重新运行过去的状态以及如何从先前状态进行分支。

配置

我们不会展示所托管图的完整代码，但如果你想查看，可以点击此处。一旦该图完成托管，我们就可以调用它并等待用户输入。

SDK 初始化

首先，我们需要配置客户端，以便与托管的图进行通信：

PythonJavascriptCURL

from langgraph_sdk import get_client
client = get_client(url=<部署 URL>)
# 使用以 "agent" 名称部署的图
assistant_id = "agent"
thread = await client.threads.create()

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

const client = new Client({ apiUrl: <部署 URL> });
// 使用以 "agent" 名称部署的图
const assistantId = "agent";
const thread = await client.threads.create();

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

重放状态

初始调用

在重放状态之前，我们需要创建可供重放的状态！为了做到这一点，让我们用一条简单的消息来调用我们的图：

PythonJavascriptCURL

input = {"messages": [{"role": "user", "content": "Please search 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": "Please search 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\": \"Please search the weather in SF\"}]},
   \"stream_mode\": [
     \"updates\"
   ]
 }" | \
 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：

{'agent': {'messages': [{'content': [{'text': "Certainly! I'll use the search function to look up the current weather in San Francisco for you. Let me do that now.", 'type': 'text'}, {'id': 'toolu_011vroKUtWU7SBdrngpgpFMn', 'input': {'query': 'current weather in San Francisco'}, 'name': 'search', 'type': 'tool_use'}], 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-ee639877-d97d-40f8-96dc-d0d1ae22d203', 'example': False, 'tool_calls': [{'name': 'search', 'args': {'query': 'current weather in San Francisco'}, 'id': 'toolu_011vroKUtWU7SBdrngpgpFMn'}], 'invalid_tool_calls': [], 'usage_metadata': None}]}}
{'action': {'messages': [{'content': '["I looked up: current weather in San Francisco. Result: It\'s sunny in San Francisco, but you better look out if you\'re a Gemini 😈."]', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'search', 'id': '7bad0e72-5ebe-4b08-9b8a-b99b0fe22fb7', 'tool_call_id': 'toolu_011vroKUtWU7SBdrngpgpFMn'}]}}
{'agent': {'messages': [{'content': "Based on the search results, I can provide you with information about the current weather in San Francisco:\n\nThe weather in San Francisco is currently sunny. This is great news for outdoor activities and enjoying the city's beautiful sights.\n\nIt's worth noting that the search result included an unusual comment about Geminis, which isn't typically part of a weather report. This might be due to the search engine including some astrological information or a joke in its results. However, for the purpose of answering your question about the weather, we can focus on the fact that it's sunny in San Francisco.\n\nIf you need any more specific information about the weather in San Francisco, such as temperature, wind speed, or forecast for the coming days, please let me know, and I'd be happy to search for that information for you.", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-dbac539a-33c8-4f0c-9e20-91f318371e7c', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': None}]}}

现在让我们获取状态列表，并从第三个状态（就在调用工具之前）开始调用：

PythonJavascriptCURL

states = await client.threads.get_history(thread['thread_id'])

# We can confirm that this state is correct by checking the 'next' attribute and seeing that it is the tool call node
state_to_replay = states[2]
print(state_to_replay['next'])

const states = await client.threads.getHistory(thread['thread_id']);

// We can confirm that this state is correct by checking the 'next' attribute and seeing that it is the tool call node
const stateToReplay = states[2];
console.log(stateToReplay['next']);

curl --request GET --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/history | jq -r '.[2].next'

输出：

['action']

要从某个状态重新运行，我们首先需要向线程状态发出一个空更新。然后，我们需要按如下方式传入生成的 checkpoint_id：

PythonJavascriptCURL

state_to_replay = states[2]
updated_config = await client.threads.update_state(
    thread["thread_id"],
    {"messages": []},
    checkpoint_id=state_to_replay["checkpoint_id"]
)
async for chunk in client.runs.stream(
    thread["thread_id"],
    assistant_id, # graph_id
    input=None,
    stream_mode="updates",
    checkpoint_id=updated_config["checkpoint_id"]
):
    if chunk.data and chunk.event != "metadata": 
        print(chunk.data)

const stateToReplay = states[2];
const config = await client.threads.updateState(thread["thread_id"], { values: {"messages": [] }, checkpointId: stateToReplay["checkpoint_id"] });
const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: null,
    streamMode: "updates",
    checkpointId: config["checkpoint_id"]
  }
);
for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl --request GET --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/history | jq -c '
    .[2] as $state_to_replay |
    {
        values: { messages: .[2].values.messages[-1] },
        checkpoint_id: $state_to_replay.checkpoint_id
    }' | \
curl --request POST \
    --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state \
    --header 'Content-Type: application/json' \
    --data @- | jq .checkpoint_id | \
curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"checkpoint_id\": \"$1\",
   \"stream_mode\": [
     \"updates\"
   ]
 }" | \
 sed 's/\r$//' | \
 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：

{'action': {'messages': [{'content': '["I looked up: current weather in San Francisco. Result: It\'s sunny in San Francisco, but you better look out if you\'re a Gemini 😈."]', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'search', 'id': 'eba650e5-400e-4938-8508-f878dcbcc532', 'tool_call_id': 'toolu_011vroKUtWU7SBdrngpgpFMn'}]}}
{'agent': {'messages': [{'content': "Based on the search results, I can provide you with information about the current weather in San Francisco:\n\nThe weather in San Francisco is currently sunny. This is great news if you're planning any outdoor activities or simply want to enjoy a pleasant day in the city.\n\nIt's worth noting that the search result included an unusual comment about Geminis, which doesn't seem directly related to the weather. This appears to be a playful or humorous addition to the weather report, possibly from the source where this information was obtained.\n\nIs there anything else you'd like to know about the weather in San Francisco or any other information you need?", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-bc6dca3f-a1e2-4f59-a69b-fe0515a348bb', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': None}]}}

正如我们所见，图从工具节点重新启动，输入与我们最初的图运行相同。

从先前状态分支

利用 LangGraph 的检查点功能，你能做的不仅仅是重放过去的状态。你可以从先前的位置进行分支，让智能体探索其他轨迹，或者让用户对工作流中的更改进行“版本控制”。

下面展示如何在特定时间点编辑状态。我们来更新状态，更改工具的输入：

PythonJavascriptCURL

# 现在让我们获取状态中的最后一条消息
# 这是包含我们要更新的工具调用的消息
last_message = state_to_replay['values']['messages'][-1]

# 现在让我们更新该工具调用的参数
last_message['tool_calls'][0]['args'] = {'query': 'current weather in SF'}

config = await client.threads.update_state(thread['thread_id'],{"messages":[last_message]},checkpoint_id=state_to_replay['checkpoint_id'])

// 现在让我们获取状态中的最后一条消息
// 这是包含我们要更新的工具调用的消息
let lastMessage = stateToReplay['values']['messages'][-1];

// 现在让我们更新该工具调用的参数
lastMessage['tool_calls'][0]['args'] = { 'query': 'current weather in SF' };

const config = await client.threads.updateState(thread['thread_id'], { values: { "messages": [lastMessage] }, checkpointId: stateToReplay['checkpoint_id'] });

curl -s --request GET --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/history | \
jq -c '
    .[2] as $state_to_replay |
    .[2].values.messages[-1].tool_calls[0].args.query = "current weather in SF" |
    {
        values: { messages: .[2].values.messages[-1] },
        checkpoint_id: $state_to_replay.checkpoint_id
    }' | \
curl --request POST \
    --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state \
    --header 'Content-Type: application/json' \
    --data @-

现在我们可以使用这个新配置重新运行我们的图，从 new_state 开始，它是 state_to_replay 的一个分支：

PythonJavascriptCURL

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

const streamResponse = client.runs.stream(
  thread["thread_id"],
  assistantId,
  {
    input: null,
    streamMode: "updates",
    checkpointId: config['checkpoint_id'],
  }
);
for await (const chunk of streamResponse) {
  if (chunk.data && chunk.event !== "metadata") {
    console.log(chunk.data);
  }
}

curl -s --request GET --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/state | \
jq -c '.checkpoint_id' | \
curl --request POST \
 --url <DEPLOYMENT_URL>/threads/<THREAD_ID>/runs/stream \
 --header 'Content-Type: application/json' \
 --data "{
   \"assistant_id\": \"agent\",
   \"checkpoint_id\": \"$1\",
   \"stream_mode\": [
     \"updates\"
   ]
 }" | \
 sed 's/\r$//' | \
 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：

{'action': {'messages': [{'content': '["I looked up: current weather in SF. Result: It\'s sunny in San Francisco, but you better look out if you\'re a Gemini 😈."]', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'tool', 'name': 'search', 'id': '2baf9941-4fda-4081-9f87-d76795d289f1', 'tool_call_id': 'toolu_011vroKUtWU7SBdrngpgpFMn'}]}}
{'agent': {'messages': [{'content': "Based on the search results, I can provide you with information about the current weather in San Francisco (SF):\n\nThe weather in San Francisco is currently sunny. This means it's a clear day with plenty of sunshine. \n\nIt's worth noting that the specific temperature wasn't provided in the search result, but sunny weather in San Francisco typically means comfortable temperatures. San Francisco is known for its mild climate, so even on sunny days, it's often not too hot.\n\nThe search result also included a playful reference to astrological signs, mentioning Gemini. However, this is likely just a joke or part of the search engine's presentation and not related to the actual weather conditions.\n\nIs there any specific information about the weather in San Francisco you'd like to know more about? I'd be happy to perform another search if you need details on temperature, wind conditions, or the forecast for the coming days.", 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': 'run-a83de52d-ed18-4402-9384-75c462485743', 'example': False, 'tool_calls': [], 'invalid_tool_calls': [], 'usage_metadata': None}]}}

正如我们所见，搜索查询从“San Francisco”更改为“SF”，正如我们所期望的那样！

Comments