📝 LangChain

Description

Definition: LangChain is a framework for building applications and agents around language models. It provides interfaces and integrations for models, tools, and agent behavior. The application supplies its model, instructions, tools, and data; LangChain is not itself a language model. See the LangChain overview.
Example: a support assistant can retrieve a return policy, include it in a model request, and produce an answer with a source reference.

Workflow overview: the diagram separates execution control, application components, and tracing. It is redrawn from the supplied ByteMonk image.
LangGraph: controls workflow state, node execution, and branches. A node may call a language model, retrieve documents, execute a tool, or run ordinary application code.
LangChain: supplies model and tool abstractions and higher-level agent construction. Components such as prompt formatting and retrieval can be used inside workflow nodes. Its high-level agents use LangGraph underneath.
LangSmith: records instrumented runs for inspection and supports evaluation. The diagram's “Trace logger,” “Output recorder,” and “Decision tracker” labels describe recorded information, not three required standalone services. A trace exposes recorded actions and outputs, not a model's unobservable internal reasoning.
Reading the arrows: solid lines represent execution or data flow; dashed lines represent tracing events. The figure illustrates responsibilities rather than a mandatory deployment topology. LangGraph can be used without LangChain, and LangSmith can instrument applications built with other frameworks. See the LangGraph overview and LangSmith observability documentation.

Choose by responsibility: these tools address complementary application concerns.

Tool	Main responsibility	Use it when
LangChain	Model and tool integration; higher-level agent construction	The application needs reusable integrations and an agent loop
LangGraph	Stateful orchestration with explicit transitions and branches	The workflow needs custom control flow, persistence, or human review
LangSmith	Tracing, debugging, datasets, and evaluation	Developers need to inspect execution and compare application quality

LangSmith also supports prompt management and deployment; this diagram focuses on its observability role. See the official ecosystem overview.

Illustrative input and output: the following is a designed workflow, not a recorded model run.

Step	Concrete data or action
User request	“How long do I have to return headphones?”
Retrieval	Document `D17`: “Headphones may be returned within 30 days.”
Workflow decision	Evidence is available, so continue to answer generation
Model input	The question, document `D17`, and an instruction to cite evidence
Expected answer	“Headphones can be returned within 30 days. [D17]”
Trace	Retrieval result, selected route, model request and response, duration, and errors if any

LangGraph explains state, nodes, edges, and workflow control.
LangSmith explains tracing and evaluation.
LlamaIndex covers data ingestion and retrieval integration.
Retrieval-Augmented Generation (RAG) explains grounding answers in retrieved documents.