AI创想

标题: LangGraph使用PostgreSaver持久化，初始化与异步实现的一些细节 [打印本页]

---

作者: AI小编    时间: 昨天 23:30
标题: LangGraph使用PostgreSaver持久化，初始化与异步实现的一些细节
作者：CSDN博客
博主环境：

1. python==3.12.9
2. langgraph==0.5.4
3. langgraph-checkpoint-postgres==2.0.23
4. psycopg==3.2.10

复制代码

一、初始化

初始化的两种方式

方法一：直接对PostgreSaver类进行初始化

1. from psycopg import Connection
2. from langgraph.checkpoint.postgres import PostgresSaver
4. db_uri = "postgres://<user>:<password>@<host>:<port>/<database>"
5. conn = Connection.connect(db_uri)
6. conn.autocommit = True    # 关键：设置连接为自动提交模式，否则会报错psycopg.errors.ActiveSqlTransaction: CREATE INDEX CONCURRENTLY 无法在事物块中运行
7. checkpointer = PostgresSaver(conn)
8. checkpointer.setup()

复制代码

方法二：使用PostgreSaver.from_conn_string方法。

1. from langgraph.checkpoint.postgres import PostgresSaver
3. db_uri = "postgres://<user>:<password>@<host>:<port>/<database>"
4. with PostgresSaver.from_conn_string(db_uri) as checkpointer:
5.     checkpointer.setup()

复制代码

初始化时指定schema

langgraph.checkpoint.postgres初始化时默认在public schema下创建四张表

    """CREATE TABLE IF NOT EXISTS checkpoint_migrations (    v INTEGER PRIMARY KEY);""",    """CREATE TABLE IF NOT EXISTS checkpoints (    thread_id TEXT NOT NULL,    checkpoint_ns TEXT NOT NULL DEFAULT '',    checkpoint_id TEXT NOT NULL,    parent_checkpoint_id TEXT,    type TEXT,    checkpoint JSONB NOT NULL,    metadata JSONB NOT NULL DEFAULT '{}',    PRIMARY KEY (thread_id, checkpoint_ns, checkpoint_id));""",    """CREATE TABLE IF NOT EXISTS checkpoint_blobs (    thread_id TEXT NOT NULL,    checkpoint_ns TEXT NOT NULL DEFAULT '',    channel TEXT NOT NULL,    version TEXT NOT NULL,    type TEXT NOT NULL,    blob BYTEA,    PRIMARY KEY (thread_id, checkpoint_ns, channel, version));""",    """CREATE TABLE IF NOT EXISTS checkpoint_writes (    thread_id TEXT NOT NULL,    checkpoint_ns TEXT NOT NULL DEFAULT '',    checkpoint_id TEXT NOT NULL,    task_id TEXT NOT NULL,    idx INTEGER NOT NULL,    channel TEXT NOT NULL,    type TEXT,    blob BYTEA NOT NULL,    PRIMARY KEY (thread_id, checkpoint_ns, checkpoint_id, task_id, idx));""",

如果要指定schema，可以在db_uri中加上option

1. db_uri = "postgres://<user>:<password>@<host>:<port>/<database>?options=-csearch_path%3D<your_schema>,public"
3. # 假设所有参数都为test
4. # db_uri = "postgres://<user>:<password>@<host>:<port>/<database>?options=-csearch_path%3Dtest,public"

复制代码

二、异步实现

如果使用图流程的异步输出，需要搭配使用AsyncPostgresSaver，否则会报错 "Synchronous calls to AsyncPostgresSaver are only allowed from a different thread. From the main thread, use the async interface."

1. from langgraph.graph import StateGraph
3. workflow = StateGraph(MainState)
4. workflow.add_node("node1", execute_node_func)
6. # ......
9. # 使用异步执行模式
10. async def generate_async_response(*args, **kwargs):
11.     async with AsyncPostgresSaver.from_conn_string(db_uri) as async_checkpointer:
12.         # 编译工作流
13.         graph = workflow.compile(checkpointer=async_checkpointer)
15.         # 忽略其他一些细节
16.         inputs = {
17.             # 入参
18.         }
19.         config = {
20.             "configurable": {
21.                 "thread_id": "123456",
22.             }
23.         }
25.         # 调用流式输出
26.         async for output in graph.astream(inputs, config, subgraphs=True, stream_mode="updates"):
27.             # do_something
28.             pass

复制代码

注意，如果如果使用了AsyncPostgresSaver，graph的同步方法不能使用了，比如get_state：

1. # 原本的方法不可用了
2. def is_thread_interrupted(graph, thread_id: str):
3.     state = graph.get_state({"configurable": {"thread_id": thread_id}})
4.     return state.next not in [(), ("__end__",)]
7. # 要换成这个
8. async def is_thread_interrupted(graph, thread_id: str):
9.     state = await graph.aget_state({"configurable": {"thread_id": thread_id}})
10.     return state.next not in [(), ("__end__",)]
14. # 对应的使用
15. async def generate_async_response(*args, **kwargs):
16.     async with AsyncPostgresSaver.from_conn_string(db_uri) as async_checkpointer:
17.         # 编译工作流
18.         graph = workflow.compile(checkpointer=async_checkpointer)
20.         inputs = {
21.             # 入参
22.         }
23.         config = {
24.             "configurable": {
25.                 "thread_id": "123456",
26.             }
27.         }
29.         if await is_thread_interrupted(graph, thread_id=thread_id):
30.             inputs = Command(resume=inputs)
32.         # 调用流式输出
33.         async for output in graph.astream(inputs, config, subgraphs=True, stream_mode="updates"):
34.             # do_something
35.             pass

复制代码

原文地址：https://blog.csdn.net/baobaodarenya/article/details/151855700

---

欢迎光临 AI创想 (https://www.llms-ai.com/)

Powered by Discuz! X3.4