Start here
LangChain in plain language
The framework helps connect application parts that would otherwise need glue code. The current LangChain documentation describes an agent harness assembled from a model, tools, prompts and middleware, with agents built on LangGraph.
The decision is not “framework or no AI”. For a small application, a direct provider SDK may be enough. Adopt a framework when its integration and orchestration support make the application easier to maintain, then inspect what happens beneath the abstraction.
How it works
Define a narrow application
Choose the task and expected result before choosing integrations. Start with a test set that can reveal incorrect answers.
Connect the required parts
Configure the model and a small tool set. Keep credentials outside prompts and define clear inputs and outputs for custom tools.
Inspect the execution trace
Review actual requests, tool results and final answers. Pin dependencies and rerun tests when models or libraries change.
Illustrative example · Fictional data
Building a fictional stock assistant
- Input
- A small catalogue lists notebook N-12 as unavailable and notebook N-18 as available. The assistant has one read-only stock lookup tool.
- Process
- Ask whether N-12 can ship. Inspect whether the application actually calls the tool and uses its returned status rather than relying on the model’s prior answer.
- Expected output
- An answer that N-12 is unavailable, with no invented dispatch date. A substitute is suggested only if its details were separately checked.
- How to check it
- Return an error instead of stock data. The assistant should say it could not check, not turn an error into “out of stock”.
Where it helps
- Applications needing reusable model and tool integrations.
- Agent prototypes that need a configurable execution harness.
- Teams comparing orchestration choices using a common task and test set.
LangChain, LangGraph and a model SDK
- Model SDK
- Calls a provider directly. Often sufficient for a small, well-defined application.
- LangChain
- Provides integrations and a configurable agent harness. Useful when these abstractions fit the application.
- LangGraph
- Provides lower-level orchestration for stateful workflows and agents, including explicit control of execution paths.
From documents to a checked application response
A retrieval application needs several responsibilities: load documents, split them into useful units, create or retrieve representations, search for relevant context, assemble model input, generate a response and validate the result. LangChain supplies integrations and abstractions for parts of this work; your application still defines the data contract and success criteria.
A prompt template combines stable instructions with task data. A retriever returns candidate context. An output parser or schema validator checks the response structure, but cannot prove its facts. Trace the retrieved passages, prompt/version, model response and validation result so a wrong answer can be investigated.
LangChain, LangGraph and LlamaIndex
Compare frameworks through a small representative workflow rather than a universal winner. LangChain offers model/tool integrations and agent/application abstractions; LangGraph exposes lower-level stateful orchestration; LlamaIndex provides data-oriented tools for retrieval and agent applications. Their capabilities overlap and evolve.
For Python-based learning, be comfortable with functions, classes, data structures, exceptions and API requests; JavaScript ecosystems also exist. Learn the underlying model and retrieval calls before depending on a framework. Logging and evaluation are essential capabilities, but no particular commercial tracing product is mandatory.
Related reading and learning pathways
Common pitfalls
- A framework does not automatically make answers accurate or permissions correct.
- Older tutorials may use different APIs; match documentation to installed package versions.
- Changing a model behind a shared interface can still change tool behaviour and output quality.
Put the idea to work
Try it yourself
Sketch the stock assistant as a direct API application and as a framework application. List which code the framework removes, which behaviour it adds and what you still need to test.
Common questions
Is LangChain an LLM?
No. It is application software that connects to models and tools. The selected model produces the model responses. You choose a model provider, such as OpenAI, Anthropic or an open model, and LangChain gives you components for prompts, tools, retrieval and chaining calls together.
Is LangChain required for RAG?
No. You can build retrieval-augmented generation with ordinary application code, a provider SDK or another framework. Frameworks such as LangChain or LlamaIndex can speed up common patterns, but a small RAG system is often easier to debug when written directly against the provider’s API.
Does using LangChain remove model API charges?
No. Model hosting, API usage and other services have their own terms and costs, independent of the framework. Track token usage, embedding calls and vector database charges separately. A framework can make repeated or hidden calls, so monitor actual usage during testing.
Your next step
Continue learning
Connected concepts
Sources and further reading
Examples on this page are teaching exercises using fictional information. They are not client case studies or measured outcomes.