LangChain vs LlamaIndex vs Raw SDK

The one-line verdict

• Learning? Raw SDK — 20 lines, zero magic, you own every failure mode.
• RAG getting serious? LlamaIndex — loaders, indices, and evals purpose-built for data.
• Agents / fast prototypes? LangChain — biggest ecosystem of chains, tools, and integrations.

Official homes: LangChain · LlamaIndex · OpenAI API.

Same task, three ways

# Raw SDK: explicit and debuggable
resp = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Summarize: ..."}],
    temperature=0.2,
)

# LlamaIndex: data in, answers out
from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
docs = SimpleDirectoryReader("docs/").load_data()
index = VectorStoreIndex.from_documents(docs)
print(index.as_query_engine().query("Refund policy?"))

# LangChain: compose chains and tools fast
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
chain = ChatPromptTemplate.from_template("Summarize: {text}") | ChatOpenAI(model="gpt-4o-mini")
print(chain.invoke({"text": "..."}))

Honest limits

• Frameworks move fast — pin versions, expect API churn between tutorials.
• Abstractions hide prompts, retries, and costs; when answers go wrong, you debug one layer down anyway.
• Rule of thumb: prototype raw, adopt a framework the second time you rebuild the same plumbing.