ArcLibrary
L1入门L2进阶L3生态

LlamaIndex

A framework that specialises in 'data ingestion + RAG' — the fastest way to feed your private content to an LLM.

LlamaIndexRAGFramework
核心 · Key Idea

In one line: LlamaIndex is the specialist for "how to feed your company's data to an LLM" — from reading files, chunking, indexing, retrieving, to feeding the retrieved snippets into the model. The full RAG pipeline is abstracted out. If your core need is "make the AI understand my docs", it's often faster than LangChain to start.

What it is#

The classic three lines:

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
 
# Read all files in a directory
docs = SimpleDirectoryReader("./my_docs").load_data()
 
# Build an index in one line (embedding + vector store)
index = VectorStoreIndex.from_documents(docs)
 
# Q&A in one line
query_engine = index.as_query_engine()
print(query_engine.query("What is our Q3 refund policy?"))

Its core abstractions are Index and QueryEngine — less general than LangChain, but the RAG path is smoother.

Analogy#

打个比方 · Analogy
  • LangChain = the Lego flagship store — every brick imaginable; you can build anything.
  • LlamaIndex = a "furniture kit" — purpose-built for the RAG chair: shorter manual, sturdier result.

Key concepts#

Reader / LoaderData loader
Hundreds of connectors: PDF / Notion / Slack / GitHub / SQL / web…
IndexIndex
Vector index / summary index / knowledge-graph index — not just vectors.
QueryEngineQuery engine
Encapsulates 'retrieve + rerank + generate' as one entrypoint.
NodeNode (chunk)
The smallest post-chunk unit, with metadata for filtering / citation.
WorkflowWorkflow
Recent versions add LangGraph-style agent / state-machine capability.

How it works#

It packages the full RAG flow as plug-and-play takeover — you pick a Reader / Index / Engine and write less glue.

Practical notes#

  • If your core is RAG, use this. Customer-support KBs, private Q&A, doc QA — a working demo in hours.
  • Readers are a treasure trove. Notion / Confluence / Jira / GitHub each have first-class loaders — much faster than self-scraping.
  • Beyond vectors. Long docs → Summary Index first, drill down to nodes — far more accurate than naive vector recall.
  • In production use IngestionPipeline. Turn "read → chunk → embed → store" into an incremental pipeline; don't rebuild on each new doc.
  • Complex agents still prefer LangGraph. LlamaIndex's Workflow is catching up.

Easy confusions#

LlamaIndex
Focused on **RAG / data**.
Faster to start; tighter constraints.
LangChain
General orchestration — flows, agents, tools.
LlamaIndex
A few lines and you have a demo.
Great for medium-scale KBs.
Roll your own RAG
Full control; all the fancy optimisations.
Megascale deployments usually go this route.

Further reading#