LlamaIndex is a Python/TS LLM framework whose worldview is data-shaped: it optimizes for retrieval-augmented and data-aware workflows. The through-line is the ingestion to query pipeline: a Document (raw file or string) is parsed into Nodes (chunks with metadata), Nodes are organized into an Index (vector, keyword, tree, knowledge-graph), and an Index exposes a QueryEngine that orchestrates retrieve → postprocess → synthesize. The mental model is 'how does data flow to the model,' which is why LlamaIndex slots into RAG more naturally than LangChain. LangChain's worldview is composition-shaped (Runnables, LCEL), not data-shaped.
**LlamaIndex** is a Python and TypeScript LLM framework built around data. Its core primitives line up on a single ingestion to query path.
First, a **Document** holds a raw input. A PDF, a Notion page, a chunk of text. Second, a NodeParser splits the Document into **Nodes**. Smaller pieces with metadata such as source, page number, and relationships to neighbouring nodes. Third, the Nodes are organised into an **Index**. Most commonly a VectorStoreIndex backed by a vector database, but tree, keyword, knowledge-graph, and summary indexes exist too. Fourth, the Index exposes a **QueryEngine** that runs retrieve → postprocess → synthesize when you ask a question.
The through-line is data flow. Every primitive answers one question on the path from raw input to final answer. That is why LlamaIndex feels natural for RAG: the whole framework is shaped like a RAG pipeline. LangChain, by contrast, gives you Runnable composition primitives and lets you build any flow, RAG or otherwise.
**LlamaIndex** is a Python and TypeScript LLM framework whose worldview is data-shaped. Its core primitives line up on an ingestion to query path:
- **Document**. Raw input (file, string, Notion page) with optional metadata.
- **Node**. Chunk produced by a NodeParser, carries metadata and relationships to neighbours.
- **Index**. Collection of Nodes organised for lookup (VectorStoreIndex, TreeIndex, KeywordTableIndex, KnowledgeGraphIndex, SummaryIndex).
- **QueryEngine**. Pulls retrieve, postprocess (rerank, filter), and synthesize (compact, refine, tree_summarize, accumulate) into one call.
The through-line is data flow. Every primitive has a single responsibility on the path from raw input to final answer.
This is why LlamaIndex slots into RAG more naturally than LangChain. LangChain's worldview is composition-shaped, Runnables and LCEL, so RAG is one pattern you assemble from generic pieces. LlamaIndex's worldview is data-shaped, so the RAG pipeline IS the framework. Many teams ship retrieval on LlamaIndex and reach for LangChain or LangGraph only when the workload becomes agent-shaped rather than data-shaped.
Quick reference
TL;DR
LlamaIndex is a data-shaped LLM framework whose primitives line up on an ingestion to query path: Document → Node → Index → QueryEngine.
Picture a research librarian's workflow. A messy stack of papers arrives. That is a Document. The librarian cuts each paper into labelled note cards. Those are Nodes. The cards go into a filing cabinet with a clever lookup system. That is the Index. When a reader asks a question, the librarian's lookup and summarise routine pulls the right cards, trims them, and reads back an answer. That is the QueryEngine. LlamaIndex builds this whole library for you. LangChain, by contrast, gives you a kit of generic pipes for connecting any step to any other step, and you decide how to play librarian yourself.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Pure conversational agent with no retrieval surface The data-shaped primitives do not earn their keep when there is no Document to Node to Index pipeline to build.
TypeScript-first streaming UI as the main product surface Vercel AI SDK's streaming hooks are a better match; LlamaIndex.TS still trails Python on feature parity.
Multi-agent orchestration where explicit routing and time-travel debugging matter LangGraph's StateGraph plus checkpointer is purpose-built for that shape; the AgentRunner split does not match it.
One-off prompt with a single fixed context block You are not building a retrieval pipeline; raw provider SDK is shorter and clearer.
Quick scan · 2–4 min
Everything important, quickly.
8 min: the four data-shaped primitives, the ingestion to query path, contrasts with LangChain's composition model, when to pair LlamaIndex retrieval with LangGraph control.
Code example
python
from llama_index.core import (
SimpleDirectoryReader,
VectorStoreIndex,
Settings,
)
from llama_index.core.node_parser import SentenceSplitter
Settings.node_parser = SentenceSplitter(chunk_size=512, chunk_overlap=64)
docs = SimpleDirectoryReader("./papers").load_data() # Documents
index = VectorStoreIndex.from_documents(docs) # Documents -> Nodes -> Index
query_engine = index.as_query_engine(similarity_top_k=5)
response = query_engine.query("Summarise the methodology section.")
print(response)
Side-by-side
Aspect
LlamaIndex (data-shaped)
LangChain (composition-shaped)
Core unit
Document → Node → Index → QueryEngine
Runnable composed via LCEL
RAG ergonomics
First-class. The framework IS a RAG pipeline
Assembled from generic Runnable pieces
Agent surface
AgentRunner / AgentWorker split
AgentExecutor (deprecated path) → LangGraph
Best fit workload
Retrieval-heavy, data-centric apps
Mixed composition, multi-step orchestration
Real-world examples
Real products, models, and research that use this idea.
1Replit's code-context retrieval uses LlamaIndex-style chunking with neighbour metadata to ground completions
2Notion AI's enterprise search ingests Documents through a parser to index pipeline that mirrors the LlamaIndex shape
3Harvey (legal AI) builds case-law retrieval on LlamaIndex's hierarchical and metadata-rich indexes to support citation
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QWhen would you prefer SummaryIndex or TreeIndex over VectorStoreIndex?
Sign in or go Pro to see the approach for this follow-up.
QHow does LlamaIndex's IngestionPipeline handle incremental updates?
Sign in or go Pro to see the approach for this follow-up.
QWhat is the AutoMergingRetriever doing under the hood?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Calling LlamaIndex 'a RAG library' and stopping there. The same primitives also power agent and workflow surfaces, and the data-shaped framing is the durable insight.