LangChain-Conceptual-Guide
by Carson Kempf
High Level
Why LangChain?
- Helps developers build applications that reason.
- LangChain is modular so any useful components can be easily integrated
Architecture
| Architecture | Uses | Uses |
|---|---|---|
| Deployment: | LangGraph Platform | LangGraph Platform |
| Components: | Integrations | Integrations |
| Architecture | LangChain | LangGraph |
| LangSmith |
|---|
| Playground |
| Debugging |
| Prompt Management |
| Annotation |
| Testing |
| Monitoring |
langchain-core
- Basic packages
- Chat models
- Vector stores
- Tools
langchain
- Contains chains and retrieval strategies
- Not third party integrations
- Not specific to one integration
Integration Packages
- Popular integrations have their own packages
- langchain-openai
- langchain-anthropic
- langchain-deepseek
langgraph
- An extension of langchain
- Aimed at building robust and complex applications with LLMs
- Used when LLMs need multiple actions
Modeling
- Models steps as edges and nodes in a graph
langserve
- Deploys langchains as REST APIs
- Get a production ready API up and running
- Meant for very primitive langchain functionality
LangSmith
- Developer platform
- Work with LLM applications
Concepts
Chat Models
- LLMs exposed via a chat API
Messages
- The unit of communication in chat models (not including the atomic token)
- Used to represent model input and output
Chat History
- A conversation represented as a sequence of messages
- Alternates between user messages and model responses
Tools
- A function associated with a model’s input schema
- Requires specific arguments
Tool Calling
- A type of chat model API
- Accepts tool schemas as inputs
- Returns a specific JSON invocation to one of our tools
Structured Output
- A technique to make a chat model respond in a structured format
- Matches any schema we might want to fit
Memory
- Information about a conversation that persists to future conversations
Multimodality
- Working with multiple modes of information
- Could be text, audio, video, and images
Runnable Interface
- Basic class that LangChain components are built on top of
Streaming
- LangChain API for outputting results as they are generated
- As opposed to after the response is completed
Document Loaders
- Load a source as a list of documents
Retrieval
- Can retrieve information from structured or unstructured data
- Comes from a datasource after a query requests retrieval
Text Splitters
- Splits model outputs
Embedding Models
- Models can represent data such as text or images
- Does this in a vector space
Vector Stores
- Stores vectors
- Allows search over vectors and metadata
Retriever
- Returns documents from a knowledge base
- Does this after a query requtests
Retrieval Augmented Generation (RAG)
- Combines LLMs with external knowledge bases
Agents
- Using a LLM to choose a sequence of actions to take
- A model can interact with external resources (tools)
Prompt Templates
- Factoring static parts of a prompt
Output Parsers
- Taking the output of a model and transforming it
Few-shot Prompting
- Providing a few examples of the task in the prompt
Example Selectors
- Selects the most relevant examples from a dataset based on a given input
- Used in few-shot prompting
Testing
- Verify a component works