LangGraph_ State Memory篇 環境建置 安裝套件包 在本演練中， 我們將使用以下套件： %pip install --upgrade --quiet langchain %pip install --upgrade --quiet langchain-openai %pip install --upgrade --quiet langgraph %pip install --upgrade --quiet langchain-community [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m50.4/50.4 kB[0m [31m1.1 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m990.6/990.6 kB[0m [31m19.4 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m384.0/384.0 kB[0m [31m12.9 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m140.2/140.2 kB[0m [31m5.9 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m141.9/141.9 kB[0m [31m7.1 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m48.2/48.2 kB[0m [31m3.1 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m360.7/360.7 kB[0m [31m15.7 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m1.1/1.1 MB[0m [31m42.1 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m75.6/75.6 kB[0m [31m2.7 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m77.9/77.9 kB[0m [31m6.3 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m318.9/318.9 kB[0m [31m20.9 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m58.3/58.3 kB[0m [31m3.9 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m81.4/81.4 kB[0m [31m3.5 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m2.3/2.3 MB[0m [31m38.3 MB/s[0m eta [36m0:00:00[0m [2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m49.2/49.2 kB[0m [31m2.3 MB/s[0m eta [36m0:00:00[0m [?25h 我們需要設置 環境變數 OPENAI_API_KEY ，可以直接完成，如下所示： import os import getpass from google.colab import userdata os.environ["OPENAI_API_KEY"] = userdata.get('OPENAI_API_KEY') os.environ["TAVILY_API_KEY"] = userdata.get('TAVILY_API_KEY') 建立小型 Graph from langgraph.graph import StateGraph, END from typing import TypedDict, Annotated import operator Define a simple 2 node graph with the following state: lnode: last node scratch: a scratchpad location count : a counter that is incremented each step 定義狀態 class AgentState(TypedDict): lnode: str scratch: str count: Annotated[int, operator.add] 定義節點函數 def node1(state: AgentState): print(f"node1, count:{state['count']}") return {"lnode": "node_1", "count": 1, } def node2(state: AgentState): print(f"node2, count:{state['count']}") return {"lnode": "node_2", "count": 1, } 設立條件邊 def should_continue(state): return state["count"] < 3 編譯整張 Graph 圖論應用說明: 節點 (Nodes): Node1 和 Node2 代表系統中的兩個狀態或處理步驟 邊 (Edges): 連接節點的邊表示狀態轉移或處理流程 有向圖 (Directed Graph): 邊是有方向的,表示狀態轉移的方向 條件轉移: 使用條件邊實現了基於條件的狀態轉移 循環 (Cycle): Node1 -> Node2 -> Node1 形成了一個潛在的循環 終止狀態: END 代表流程的結束點 from langgraph.checkpoint.memory import MemorySaver # 初始化 StateGraph builder = StateGraph(AgentState) # 添加節點 builder.add_node("Node1", node1) builder.add_node("Node2", node2) # 添加邊 builder.add_edge("Node1", "Node2") builder.add_conditional_edges( "Node2", should_continue, {True: "Node1", False: END} ) builder.set_entry_point("Node1") # Set up memory memory = MemorySaver() graph = builder.compile(checkpointer=memory) 視覺化 GRaph from IPython.display import Image, display try: display(Image(graph.get_graph(xray=True).draw_mermaid_png())) except: pass 與 GRaph 互動 thread = { "configurable": { "thread_id": "1" } } graph.invoke({"count":0, "scratch":"hi"},thread) node1, count:0 node2, count:1 node1, count:2 node2, count:3 {'lnode': 'node_2', 'scratch': 'hi', 'count': 4} 查看狀態(state) Get the current state. Note the values which are the AgentState. Note the config and the thread_ts. You will be using those to refer to snapshots below. import rich result = graph.get_state(thread) rich.print(result) StateSnapshot( values={'lnode': 'node_2', 'scratch': 'hi', 'count': 4}, next=(), config={ 'configurable': { 'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bcde-67c2-8004-03e814a1fc10' } }, metadata={'source': 'loop', 'writes': {'Node2': {'lnode': 'node_2', 'count': 1}}, 'step': 4}, created_at='2024-08-09T07:05:44.508382+00:00', parent_config={ 'configurable': { 'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bccf-6fed-8003-870412c94ed8' } } ) 查看歷史狀態 有些文件是寫說 thread_ts, parent_ts，根據官方文件 LangGraph v0.2，已經分別改名為 checkpoint_id 以及 parent_checkpoint_id for state in graph.get_state_history(thread): rich.print(state, "\n") 保存歷史資訊成 list 後續方便 states = [] for state in graph.get_state_history(thread): states.append(state.config) print(f"{state.config}, count: {state.values['count']}") {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561e0-cb40-6be9-8004-f21497772532'}}, count: 4 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561e0-cb33-6d1b-8003-d7c00eb02a41'}}, count: 3 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561e0-cb25-662d-8002-463f51af9a58'}}, count: 2 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bcde-67c2-8004-03e814a1fc10'}}, count: 4 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bccf-6fed-8003-870412c94ed8'}}, count: 3 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bcc2-654d-8002-cee129f19add'}}, count: 2 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bc9c-6d50-8001-e9bee1517c9b'}}, count: 1 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bc87-6d50-8000-05126d1312e2'}}, count: 0 {'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bc82-6bdb-bfff-cb96ef83cd2d'}}, count: 0 取得特定狀態資訊 specific_state = graph.get_state(states[-3]) rich.print(specific_state) StateSnapshot( values={'lnode': 'node_1', 'scratch': 'hi', 'count': 1}, next=('Node2',), config={ 'configurable': { 'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bc9c-6d50-8001-e9bee1517c9b' } }, metadata={'source': 'loop', 'writes': {'Node1': {'lnode': 'node_1', 'count': 1}}, 'step': 1}, created_at='2024-08-09T07:05:44.481474+00:00', parent_config={ 'configurable': { 'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1ef561dc-bc87-6d50-8000-05126d1312e2' } } ) 將該狀態用於invoke 中以回到過去。 請注意，它使用 states[-3] 作為 current_state 並繼續到 node2。 graph.invoke(None, states[-3]) node1, count:2 node2, count:3 {'lnode': 'node_2', 'scratch': 'hi', 'count': 4}