LlamaIndex

A data framework for LLM apps, focused on ingesting, indexing, and querying documents for retrieval-augmented generation.

At a glance

Field	Value
Category	RAG / data framework for LLMs
Difficulty	Intermediate
When to use	Document-heavy RAG systems
When not to use	Non-retrieval LLM features
Alternatives	LangChain, Haystack, custom

What it is

LlamaIndex provides readers (PDF, Notion, Google Drive, SQL, and hundreds more), node parsers (chunking strategies), indexes (vector, tree, keyword, knowledge graph), and query engines that tie them together. The abstraction is explicitly data-first: “how do I ingest, index, and query my documents” rather than LangChain’s more general “how do I chain LLM calls”.

When we reach for it at Ephizen

Heterogeneous document sets (PDFs, Markdown, Confluence) where its loaders save us days of work.
Trying different indexing strategies (hierarchical summarization, auto-retriever) without rewriting the pipeline.
Quickly benchmarking retrieval approaches against the same evaluation set.

Getting started

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader

docs = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(docs)
engine = index.as_query_engine(similarity_top_k=5)
print(engine.query("What's the refund window?"))

Gotchas

Defaults use OpenAI for both embeddings and LLM. Configure a local embedding model and a provider of your choice before the first run, or your bill will surprise you.
The chunking defaults are generic — tune chunk size and overlap for your corpus.
LlamaIndex and LangChain overlap a lot. Pick one and stick with it in a single codebase.