How to Use LlamaIndex

LlamaIndex is the data framework for LLMs, purpose-built for ingesting documents and powering RAG over your private knowledge.

Where LangChain tries to be everything, LlamaIndex stays narrower and deeper: loaders for 150+ data sources, chunking and metadata extraction pipelines, a VectorStoreIndex abstraction over every vector DB that matters, and query engines that combine retrieval with re-ranking and response synthesis. A newer Workflows API adds event-driven orchestration for when you outgrow simple query pipelines.

What it is

LlamaIndex is MIT-licensed and maintained by LlamaIndex Inc. (Jerry Liu and team). The Python package llama-index-core is the base; integrations live in separate packages like llama-index-vector-stores-qdrant. A TypeScript port (llamaindex on npm) covers the essentials. LlamaParse, a paid managed service, handles complex PDFs and tables the OSS parser struggles with.

Install

pip install llama-index
# or the modular install
pip install llama-index-core llama-index-llms-openai llama-index-embeddings-openai

First run

Index a folder of documents and ask a question about them:

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader

docs = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(docs)
engine = index.as_query_engine()

print(engine.query("What is our refund policy?"))

Everyday workflows

• Build an IngestionPipeline with SentenceSplitter + TitleExtractor + embeddings and cache it to disk.
• Swap VectorStoreIndex’s backend for Qdrant, Pinecone, or pgvector with a few lines of config.
• Use Workflows and AgentWorkflow to combine RAG with tool-using agents for multi-step answers.

Gotchas and tips

RAG quality lives and dies by chunking. Default settings are generic; tune chunk_size and chunk_overlap to your content — contracts, forum posts, and code all want different values. Measure recall with a small labeled set before declaring victory.

Persistence is a common foot-gun. Calling from_documents every run re-embeds everything and bills you twice. Use StorageContext.persist() and load_index_from_storage, or push to a real vector DB that keeps state for you.

Who it’s for

Teams whose product is basically “chat with our documents” — legal, support, internal search, research. Tip: reach for LlamaParse the first time a client-provided PDF has merged cells or scanned tables — hand-rolling a parser for those is a month you will not get back.