AI/ML7 min read

LangChain

Compose LLM applications from chains, tools, memory, and retrievers - LangChain provides the building blocks for RAG pipelines, agents, and multi-step AI workflows.

LangChainLLMRAGAgentsPython

skill: LangChain category: AI/ML tags: [LangChain, LLM, RAG, Agents, Python] readingTime: 7

LangChain accelerates building LLM applications with pre-built retrievers, memory stores, and tool integrations. For simple one-shot LLM calls, use the OpenAI or Anthropic SDK directly. For production agents requiring reliability and observability, pair LangChain with LangSmith tracing.

Workflow

Install and set up LangChainLangChain is modular - install only the integrations you need to keep bundle size small.
Build chains with LCELLangChain Expression Language (LCEL) composes prompts, models, and output parsers using a pipe | syntax - streams by default and is easy to reason about.
Build a RAG chain with a retrieverLangChain's retriever interface abstracts over Pinecone, pgvector, Chroma, and more - swap vector stores without changing chain logic.
Add memory for multi-turn conversationsLLMs are stateless - without memory, each call has no context from prior turns. Memory stores and injects conversation history.
Trace and debug with LangSmithLangSmith captures every prompt, response, latency, and token count in a UI - essential for debugging multi-step chains where intermediate outputs are invisible.

Output Format

Composable LangChain application: LCEL chain → retriever → memory → LangSmith tracing.

from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langchain_core.output_parsers import StrOutputParser

chain = prompt | ChatOpenAI(model='gpt-4o') | StrOutputParser()
result = chain.invoke({'key': 'value'})

Examples

Input

Output

Edge Cases & Gotchas

Version fragmentation - LangChain frequently breaks APIs between versions (0.1, 0.2, 0.3). Pin exact versions in requirements.txt and read the migration guide before upgrading.
Context window overflow - Injecting too many retrieved chunks exceeds the model's context window. Limit with top_k=5, and trim chunks to max_tokens/k characters.
Parallel retrieval latency - Sequential retrieval is slow when fetching from multiple sources. Use RunnableParallel to fetch from multiple retrievers simultaneously.
Memory grows unbounded - InMemoryChatMessageHistory grows with every message. Trim with a ConversationBufferWindowMemory(k=5) to keep only the last k turns.

When to Break These Rules

Use the raw OpenAI/Anthropic SDK for simple one-shot generation - LangChain adds overhead and complexity for single calls.
Use LlamaIndex instead of LangChain when the application is purely a knowledge retrieval system - LlamaIndex's data connectors and index types are more mature for that use case.

Reference Files

https://python.langchain.com/docs/
https://docs.smith.langchain.com/