Build Agentic RAG Systems

An archived LangGraph example (no longer maintained - see LangChain's consolidated documentation for current guidance) implementing an Agentic RAG pattern: a retrieval agent that decides for itself whether to call a retriever tool, rather than always retrieving on every query. Three blog posts are indexed and wrapped as a retriever tool; the agent (call_model) can choose to call that tool (action), gets the tool output appended to its message state, and is called again with that new context.

Use this as a reference pattern for building a retrieval agent that decides on its own whether retrieval is even necessary for a given question, instead of a fixed retrieve-then-generate pipeline. It is explicitly archived and no longer updated, so treat it as a study reference rather than a production dependency - check LangChain's current consolidated documentation for up-to-date APIs and patterns.

State: a list of messages, appended to by each node as the graph runs. Nodes: call_model (the agent, which decides whether to call the retriever tool) and action (executes the retriever tool call). Conditional edges route from the agent to either the tool-calling action or onward, based on whether the agent decided to call a function. Output: the final message state after the agent has optionally retrieved and incorporated context.

Built with LangGraph and LangChain, and integrates with LangSmith for tracing, debugging, and monitoring the graph during development.

Developers building an agent that should retrieve from an index only when it decides retrieval is actually useful for the question at hand, rather than retrieving unconditionally on every turn.