LangChain

What is LangChain?

LangChain is an open-source development framework designed for large language model applications and AI Agents, offering support for Python as well as JavaScript/TypeScript. Developers can use a unified interface to connect different models, tools, retrieval systems, structured output mechanisms, and middleware in order to create chat assistants, RAG-based document query systems, research agents, data analysis tools, and applications for business automation.

The current LangChain product stack should be understood as having three layers:LangChain It is a high-level agent and application development framework.LangGraph It is an underlying orchestration runtime designed for stateful, long-running, and recoverable agents.LangSmith It is a commercial platform used for tracking, evaluating, monitoring, and deploying. The open-source libraries are available free of charge, but this does not mean that LangSmith and the use of models are also free.

Core features of LangChain

Unified model interface

LangChain reduces the need for repetitive code when switching between providers such as OpenAI, Anthropic, Google, AWS, Azure, and local models, by standardizing messages, tool calls, streaming outputs, and structured responses. The capabilities, parameters, and output formats of different models vary; therefore, it is necessary to test tool calls, images, context, and structured outputs before making any transitions.

AI Agents and Tools

The agent uses a model to determine when to invoke search, database, file, code, or business API tools, and then proceeds with further reasoning based on the results provided by those tools. Developers can define the parameters for these tools, as well as the error handling mechanisms and the format of the output.

When a tool has actual execution permissions, minimum privileges, timeout settings, retry mechanisms, idempotency, and manual approval should be implemented.

Middleware

Middleware allows logic to be inserted before and after model calls, tool calls, and state changes throughout the Agent’s lifecycle; it is used for dynamic prompts, model routing, summarization, filtering of sensitive information, retry mechanisms, human intervention, and call restrictions. This prevents cross-cutting capabilities from being scattered across individual tools, making it suitable for creating a unified governance layer.

RAG and document retrieval

The LangChain ecosystem provides components for document loading, text segmentation, embedding generation, vector storage, retrieval, and generation, which enable the delivery of results from private data to models. The quality of RAG depends on parsing, segmentation, indexing, filtering, and evaluation; it does not become reliable automatically just by connecting to a vector database.

The formal system should preserve source metadata and verify references.

LangGraph state orchestration

LangGraph represents nodes and state transitions using a graph structure, and it supports persistence, stream-based execution, a human-in-the-loop approach, memory handling, branching, loops, and recovery from interruptions. It is suitable for long-running tasks, agents that require clear tracking of their state, as well as multi-agent coordination.

LangChain’s capability for creating advanced agents operates on top of LangGraph; beginners can start by using LangChain, and switch to LangGraph only when they need more advanced control mechanisms.

LangSmith Observability and Evaluation

LangSmith is capable of recording a complete trace, including model requests, tool calls, delays, tokens, and errors; it also supports datasets, online/offline evaluation, manual annotation, Prompt Playground, monitoring, and insights.

It does not require that applications use LangChain; by integrating the tracking SDK, it is possible to monitor Agents written in other frameworks as well.

LangSmith Deployment

Deployment provides hosted infrastructure, streaming intermediate steps, status and memory APIs, scheduled tasks, authentication, as well as MCP Server capabilities for Agents that need to run on a long-term basis. The Plus version includes one free Serverless Small deployment, while additional Serverless or Dedicated instances are charged based on resource usage.

Enterprise offers the option of self-hosting or hybrid deployment.

Open source and licenses

The core Python repository of LangChain is licensed under the MIT license; core libraries such as LangGraph are also available on the official GitHub repository. The LangSmith cloud platform, deployment management tools, and certain enterprise services constitute commercial products.

Each model, vector database, and integration package also comes with its own license and terms of service.

Comparison of LangChain product stacks

LangSmith prices and packages

LangSmith uses a model of \"base fee per seat + pricing based on usage.\" LCU refers to computing resources, while LSU refers to tracking or storage resources.

The current basic price on the official website is $1.50 per LCU and $1 per LSU.

After the free amount of Trace is used up, a pay-as-you-go pricing model applies. The resources deployed are calculated in terms of vCPUs, memory, and database usage, expressed as LCU/LSU, with billing done on a per-second basis.

The fees for model tokens are usually still charged by the model provider.

There are many factors that affect pricing; the production budget should be calculated using the calculator on the official website, with regular monitoring of the actual progress.

LangChain Getting Started Tutorial

1. Select language:Use Python or TypeScript per project, create separate virtual environments, and lock down dependencies.
2. Install the core package:Install LangChain along with the integration packages provided by the relevant model providers, and store the API Key in an environment variable.
3. Create a model:Initialize the Chat Model by first testing ordinary messages, streaming responses, and structured outputs.
4. Definition tool:Provide a clear name, description, and parameter schema for the function, and restrict permissions for dangerous operations.
5. Create Agent:Configure system commands, tools, middleware, and termination conditions, and use test tasks to evaluate the selection of tools.
6. Add RAG as needed:Process documents, create indexes and retrievers, and establish an evaluation set with standard answers.
7. Complex states are represented using LangGraph:When iteration, recovery, memory, and approval are required, structure the process as an explicit state diagram.
8. Connect to LangSmith:Enable Tracing, compare the versions of the Prompt, the model, and the tools, verify the quality and cost, and then proceed with deployment.

LangChain usage tutorial

Create reusable professional workflows

1. Select a low-risk, real-world project as a template;
2. Standardize the interface of the unified model, as well as the sequence in which AI Agents, tools, and middleware are used;
3. Record the environment, model, prompts, and failure conditions;
4. Set up manual approval for write, deploy, and delete actions;
5. Compare speed, cost, test pass rate, and the amount of rework;
6. Expand to a team or production environment only after stability has been verified;

Which users are it suitable for

• Developers who need to rapidly create models, tools, and retrieval-based AI applications;
• An engineering team responsible for developing long-running, recoverable agents that allow for human intervention;
• Organizations that need to systematically track, evaluate, and monitor LLM applications;
• Product teams that wish to maintain a relatively consistent interface across various model suppliers;
• Companies that need to host, self-host, or deploy production agents in a hybrid manner.

Advantages and precautions

• The advantages of LangChain lie in its mature ecosystem, extensive integration of models and data, abundant documentation and community resources, as well as a complete product stack that covers development, orchestration, evaluation, and deployment.
• High-level APIs are suitable for a quick start, while LangGraph offers low-level control for advanced users;
• Abstraction at the framework layer also increases dependencies and complicates debugging.
• The package version should be fixed, the migration documentation should be read, and it is necessary to avoid over-engineering simple models just in order to use the framework.
• The output from the agent is unreliable, and RAG may also fail to detect things.
• Production systems must establish evaluation sets, tracking mechanisms, access controls, and manual backup solutions.
• The traces collected by LangSmith may contain user data; masking and data governance measures should be implemented before they are accessed.

Frequently Asked Questions

Is LangChain free?

The core libraries of LangChain and LangGraph are available freely and open-source. LangSmith offers a free developer package, while fees apply for team use, deployment, and excessive usage.

The model API is charged separately.

What is the difference between LangChain and LangGraph?

LangChain provides higher-level interfaces for Agents, models, tools, and Middleware; LangGraph is responsible for the underlying orchestration that ensures statefulness and recoverability.

The two can be used together.

Must LangSmith be used together with LangChain?

It’s not necessary. LangSmith can track and evaluate LLM applications built using other frameworks or custom code.

Does LangChain support JavaScript?

It supports the Python as well as JavaScript/TypeScript ecosystems. The degree of integration across different languages and the pace of releases may vary slightly.

Is LangChain open source?

The core framework is open-source and licensed under the MIT license. The LangSmith cloud platform and enterprise hosting services are considered commercial offerings.