Skip to main content

LlamaIndex: Data Framework for LLMs

While LangChain and LangGraph focus on general agent execution and state, LlamaIndex is specialized for connecting private, unstructured enterprise data to LLMs.

LlamaIndex Focus

LlamaIndex provides high-performance data ingestion connectors (LlamaHub), document chunking, node relationships, vector/summary indexing, and advanced query engines.

Core LlamaIndex Abstractions​

  1. Documents & Nodes: Unstructured text parsed into chunked Node objects with rich metadata relationships (parent, child, next/prev chunk).
  2. Data Connectors: Over 100+ connectors to pull data from PDFs, Notion, SQL databases, Slack, Google Drive, and API endpoints.
  3. Indexes: Data structures (VectorStoreIndex, SummaryIndex, KnowledgeGraphIndex, TreeIndex) built over Node objects.
  4. Query Engines: High-level retrieval pipelines that take user queries, fetch relevant nodes, and generate synthesized answers.
from llamaindex.core import SimpleDirectoryReader, VectorStoreIndex

# 1. Load Documents
documents = SimpleDirectoryReader("./data_folder").load_data()

# 2. Build Vector Index
index = VectorStoreIndex.from_documents(documents)

# 3. Create Query Engine
query_engine = index.as_query_engine(similarity_top_k=5)

# 4. Query
response = query_engine.query("What are the key requirements for Project Panama?")
print(response)

Router & Sub-Question Query Engines​

LlamaIndex excels at complex query decomposition:

  • Router Query Engine: Dynamically routes user queries to the best matching index (e.g. sending summarize requests to a SummaryIndex and specific facts to a VectorStoreIndex).
  • SubQuestion Query Engine: Breaks complex questions into smaller sub-queries, executes them against relevant data sources in parallel, and synthesizes a unified answer.