Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/langchain-ai/langchain/llms.txt

Use this file to discover all available pages before exploring further.

LangChain is built as a modular Python monorepo with a clear separation of concerns across multiple layers. This architecture enables both flexibility and stability while supporting a rich ecosystem of integrations.

​
Monorepo Structure

The LangChain project is organized into independently versioned packages: 
langchain/
├── libs/
│   ├── core/             # langchain-core: Base abstractions
│   ├── langchain/        # langchain-classic (legacy)
│   ├── langchain_v1/     # langchain: Main package
│   ├── partners/         # Third-party integrations
│   │   ├── openai/
│   │   ├── anthropic/
│   │   ├── ollama/
│   │   └── ...
│   ├── text-splitters/   # Document chunking
│   ├── standard-tests/   # Integration test suite
│   └── model-profiles/   # Model configurations

​
Package Dependencies

Each package maintains its own pyproject.toml and dependency specifications. The dependency flow is:
  • Partners depend on Core
  • LangChain depends on Core
  • Core has minimal third-party dependencies
The langchain-core package intentionally keeps dependencies lightweight to serve as a stable foundation. No third-party integrations are defined in core.

​
Layered Architecture

LangChain uses a three-layer architecture that separates abstractions from implementations:

​
Core Layer (langchain-core)

Defines base abstractions, interfaces, and protocols that all components implement: 
from langchain_core.runnables import Runnable
from langchain_core.language_models import BaseLanguageModel
from langchain_core.messages import BaseMessage
from langchain_core.tools import BaseTool
Key abstractions defined in core:
  • Runnable: Universal invocation protocol
  • BaseLanguageModel: Language model interface
  • BaseMessage: Message types for chat interactions
  • BaseTool: Tool interface for agent actions
  • BasePromptTemplate: Prompt template system
  • BaseOutputParser: Output parsing interface
  • Serializable: Serialization support
Location: /libs/core/langchain_core/

​
Implementation Layer (langchain)

Provides concrete implementations and high-level utilities: 
from langchain.chains import LLMChain
from langchain.agents import AgentExecutor
from langchain.memory import ConversationBufferMemory
This layer builds on core abstractions to provide:
  • Pre-built chains and agent patterns
  • Memory systems for stateful applications
  • Document loaders and transformers
  • Utilities for common workflows
Location: /libs/langchain_v1/langchain/

​
Integration Layer (Partners)

Third-party service integrations as separate packages: 
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from langchain_ollama import ChatOllama
Each integration:
  • Lives in its own versioned package
  • Depends only on langchain-core
  • Implements core abstractions
  • May be maintained in this monorepo or external repositories
Location: /libs/partners/*/
Some integrations (like langchain-google and langchain-aws) are maintained in separate repositories. Check the integrations documentation for the complete list.

​
Extension Points

LangChain provides multiple extension mechanisms to customize behavior:

​
1. Custom Runnables

Extend the Runnable protocol to create composable components: 
from langchain_core.runnables import Runnable, RunnableConfig
from typing import Any

class CustomProcessor(Runnable[str, str]):
    """Custom runnable with full LCEL support."""
    
    def invoke(self, input: str, config: RunnableConfig | None = None) -> str:
        # Your processing logic
        return input.upper()
    
    # async, batch, and stream are auto-generated
Custom runnables automatically get:
  • Async support (ainvoke, astream, abatch)
  • Streaming capabilities
  • Batch processing
  • Composition with | operator
See: /libs/core/langchain_core/runnables/base.py:124

​
2. Tool Development

Create custom tools by extending BaseTool: 
from langchain_core.tools import BaseTool
from pydantic import BaseModel, Field

class CalculatorInput(BaseModel):
    expression: str = Field(description="Mathematical expression to evaluate")

class CalculatorTool(BaseTool):
    name: str = "calculator"
    description: str = "Evaluates mathematical expressions"
    args_schema: type[BaseModel] = CalculatorInput
    
    def _run(self, expression: str) -> str:
        return str(eval(expression))
See: /libs/core/langchain_core/tools/base.py:289

​
3. Custom Language Models

Implement custom LLM wrappers: 
from langchain_core.language_models import BaseChatModel
from langchain_core.messages import BaseMessage, AIMessage
from langchain_core.outputs import ChatResult, ChatGeneration

class CustomChatModel(BaseChatModel):
    """Custom chat model implementation."""
    
    def _generate(
        self,
        messages: list[BaseMessage],
        stop: list[str] | None = None,
        **kwargs: Any,
    ) -> ChatResult:
        # Call your model API
        response = "Model response"
        message = AIMessage(content=response)
        generation = ChatGeneration(message=message)
        return ChatResult(generations=[generation])
    
    @property
    def _llm_type(self) -> str:
        return "custom"
See: /libs/core/langchain_core/language_models/chat_models.py

​
4. Callbacks and Tracing

Hook into the execution lifecycle: 
from langchain_core.callbacks.base import BaseCallbackHandler
from langchain_core.messages import BaseMessage
from langchain_core.outputs import LLMResult
from uuid import UUID

class LoggingCallback(BaseCallbackHandler):
    """Custom callback for observability."""
    
    def on_llm_start(
        self,
        serialized: dict,
        prompts: list[str],
        **kwargs,
    ) -> None:
        print(f"LLM started with prompts: {prompts}")
    
    def on_llm_end(
        self,
        response: LLMResult,
        **kwargs,
    ) -> None:
        print(f"LLM finished: {response}")
See: /libs/core/langchain_core/callbacks/base.py

​
Plugin Lifecycle

Components in LangChain follow a consistent lifecycle:

​
1. Initialization

Components are initialized with configuration: 
from langchain_openai import ChatOpenAI

model = ChatOpenAI(
    model="gpt-4",
    temperature=0.7,
    api_key="your-key"
)

​
2. Configuration

Runtime configuration through RunnableConfig: 
from langchain_core.runnables import RunnableConfig

config = RunnableConfig(
    tags=["production", "important"],
    metadata={"user_id": "123"},
    callbacks=[logging_callback],
    max_concurrency=5
)

result = model.invoke("Hello", config=config)
Configuration fields defined in /libs/core/langchain_core/runnables/config.py:49:
  • tags: For filtering and categorization
  • metadata: JSON-serializable context
  • callbacks: Lifecycle hooks
  • run_name: Custom tracer name
  • max_concurrency: Parallel execution limit

​
3. Invocation

Components expose standard invocation methods: 
# Synchronous
result = runnable.invoke(input, config=config)

# Asynchronous
result = await runnable.ainvoke(input, config=config)

# Streaming
for chunk in runnable.stream(input, config=config):
    print(chunk)

# Batch processing
results = runnable.batch([input1, input2], config=config)

​
4. Composition

Components compose using the pipe operator: 
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

chain = prompt | model | StrOutputParser()
result = chain.invoke({"topic": "AI"})

​
5. Serialization

Components can be serialized for storage and versioning: 
from langchain_core.load import dumps, loads

# Serialize
serialized = dumps(chain)

# Deserialize
reconstructed = loads(serialized)
See: /libs/core/langchain_core/load/serializable.py:88

​
Development Tools

The monorepo uses modern Python tooling:
  • uv: Fast package installer and resolver
  • ruff: Linting and formatting
  • mypy: Static type checking
  • pytest: Testing framework
  • make: Task automation
# Install all dependencies
uv sync --all-groups

# Run tests
make test

# Lint code
make lint

# Format code
make format

# Type checking
uv run --group lint mypy .

​
Package Management

Each package in /libs/ has:
  • pyproject.toml: Package metadata and dependencies
  • uv.lock: Locked dependency versions
  • Independent versioning and release cycle
Local development uses editable installs via [tool.uv.sources] configuration.

​
Testing Strategy

LangChain uses a layered testing approach:
  • Unit tests (tests/unit_tests/): No network calls, fast feedback
  • Integration tests (tests/integration_tests/): Real API calls
  • Standard tests (libs/standard-tests/): Shared test suite for integrations

​
Next Steps

Runnables

Learn about the universal invocation protocol

Messages

Understand message types and structure

Tools

Build custom tools for agents

Agents

Create autonomous agent systems

Build docs developers (and LLMs) love

Get started for freeTalk to us