What Is LlamaIndex?

The other major RAG/agent framework

This repo's lessons/langchain course already taught you the biggest name in this space. LlamaIndex is the other one, and it's worth learning on its own terms rather than as "LangChain but different," because it starts from a genuinely different center of gravity.

LangChain is chain-centric: its core abstraction is a Runnable, a step you can compose with other steps into arbitrary pipelines. RAG (retrieval-augmented generation, searching your own documents before asking the model to answer) is one thing you can build with LangChain, alongside agents, simple prompt chains, tool-calling loops, and more.

LlamaIndex is data-centric: its core abstraction is an Index, and the framework's whole shape is built around one path: ingest documents, build an index, query it. RAG isn't a feature you opt into, it's the default thing the framework does. Agents and tool-calling exist in LlamaIndex too (Lessons 14-17 of this course), but they're built as a layer on top of that indexing core, not the other way around.

When to reach for which

Neither is strictly better. If you're building something where retrieval is one step among several arbitrary ones (call a tool, then a model, then another tool, maybe loop), LangChain's composability is a more natural fit, especially paired with LangGraph. If the heart of what you're building really is "let an LLM answer questions about a pile of documents," LlamaIndex gets you there in fewer lines, with more of the RAG-specific machinery (chunking strategies, retrieval modes, response synthesis modes) built in and named for exactly that job.

In production, they're not even mutually exclusive: Lesson 23 of this course shows wrapping a LlamaIndex query engine as a tool inside a LangGraph or Pydantic AI agent, using each framework for what it's best at.

This course's prerequisite

This course assumes you've done lessons/langchain at least through its RAG lessons (27-29), so it moves faster through ideas you've already seen (what an embedding is, why chunking matters, what similarity search does) and spends its explanations on what's actually new here: LlamaIndex's own vocabulary, Document, Node, Index.

Setup

None beyond what you already have. This course reuses the same GOOGLE_API_KEY from your .env file and the same uv sync-installed .venv as every other course in this repo, llama-index, llama-index-llms-google-genai, and llama-index-embeddings-google-genai are already in pyproject.toml.

The code, piece by piece

LANGCHAIN_VS_LLAMAINDEX = { ... }

A plain Python dict used purely to print a comparison table, no LlamaIndex import yet. dict.items() is used to loop over both the key (aspect) and value (comparison) together.

print("This course's core loop, spelled out ...")

Names the four objects this course builds up over Lessons 2-5: Document (raw source text), Node (a chunk of a Document, the unit actually stored and searched), Index (Nodes plus their embeddings), and QueryEngine (wraps an Index to answer questions).