▼Recently, there have been many live broadcasts,make an appointment to ensure you gain something
Today’s live broadcast:《Building Agent Industrial Applications with RAG and GPTs Practical Implementation》
—1—
LangGraph Technical Architecture Interpretation
LangGraph is a tool for building stateful and multi-role Agent applications. It is not a new framework independent of Langchain, but rather an extension library built on top of Langchain, capable of seamless cooperation with existing Langchain chains, LangChain Expression Language (LCEL), and more. LangGraph can coordinate multiple Chains, Agents, Tools, etc., to collaboratively complete input tasks, supporting more refined control of LLM calls, “loops,” and Agent processes..
LangGraph implements a new form of the previously black-boxed AgentExecutor calling process: the StateGraph. It precisely defines the details of LLM-based tasks (e.g., RAG, code generation, etc.) using a Graph (defining the nodes and edges of the graph), and ultimately compiles applications based on this graph. During task execution, a central state object is maintained, which continuously updates based on node transitions, and the properties contained in the state can be defined by the user.
Let’s analyze the official RAG application Graph case shown below to better understand LangGraph.
-
StateGraph: Represents the base class of the entire state graph. -
Nodes: Nodes. Once the graph is created, nodes can be added to it. Nodes are usually callable functions, runnable Chains, or Agents. There is a special node called END, entering this node indicates the end of execution. -
In the above diagram, inference function calls, calling the retriever, generating response content, question rewriting, etc., are all task nodes. -
Edges: Edges. Once nodes are created, edges need to be added to the graph, representing the relationship of transitioning from the previous node to the next node. Currently, there are three types of edges: -
Starting Edge: A special edge. Used to define the starting node of task execution, it has no previous node. -
Normal Edge: A regular edge. Represents immediately entering the next node after the previous node completes execution. For example, after obtaining results from calling Tools, it immediately enters the LLM inference node. -
Conditional Edge: A conditional edge. Represents that after the previous node completes execution, it needs to jump to a certain node based on conditions; thus, this edge requires not only upstream and downstream nodes but also a condition function to decide the downstream node based on the condition function’s return. -
In the above diagram, Check Relevance is a conditional edge; its upstream node is retrieving relevant documents, and the condition function checks whether the document is relevant. If relevant, it enters the downstream node [Generate Answer]; if not relevant, it enters the downstream node [Rewrite Input Question].
app = graph.compile()
—2—
Building Agent Applications with LangGraph
LangGraph essentially makes the currently black-boxed AgentExecutor transparent, allowing developers to define the internal structural details (in the form of a graph), thereby achieving more powerful functionalities. Therefore, LangGraph can be used to reimplement the original AgentExecutor, achieving a basic ReAct paradigm Agent application.
Pseudocode implementation is as follows:
Comments in the code explain the details of graph construction. Clearly, this is much more complex than simply using agentExecutor, but it also demonstrates the powerful control capabilities of LangGraph when constructing LLM applications: through the definition of the Graph, the processing of an LLM application can be designed in great detail to meet the needs of numerous complex scenarios in AI Agent industrial applications.
2. Practical case of reconstructing AI e-commerce middle platform industry with AI Agent
3. In-depth analysis of AI Agent technical architecture based on GPTs
Receive the “AI Large Model Technology Live Broadcast“
END