04 Frameworks & Tools 3 min read 562 words

LlamaIndex (Dec 2025)

While LangChain focuses on "Orchestration," LlamaIndex is the master of Data-Centric AI. In late 2025, it has evolved from a RAG library into a framework for Workflows and Agentic Data Manipulation.

frameworkragretrievalcore
01framework philosophy

The Data Framework Philosophy

LlamaIndex is built on the belief that the data is more important than the model.

  • The Node: Every chunk of data is a "Node" with rich metadata (relationships, summaries, and parent-child links).
  • The Retriever: LlamaIndex provides the most diverse set of retrievers (Summary, Knowledge Graph, Tree, and Keyword).
02dec tech

LlamaIndex Workflows (Dec 2025 Tech)

In late 2024, LlamaIndex introduced Workflows, its answer to LangGraph.

  • Event-Driven Architecture: Nodes communicate by emitting Events.
  • Concurrency: Workflows are natively async and handle large-scale parallel data processing better than linear chains.
Pythonpython · 6 lines
123456
# Conceptual 2025 Workflow
class RAGWorkflow(Workflow):
    @step
    async def ingest(self, ev: StartEvent) -> RetrievalEvent:
        # Custom logic...
        return RetrievalEvent(results=nodes)
03indexing

Advanced Indexing

  1. Property Graphs: Linking vector chunks to graph nodes for RAG.
  2. Context-Aware Splitters: Grouping text by "Meaning" rather than "Token count" (using smaller LLMs to find optimal breakpoints).
  3. Dynamic Pathing: The retriever decides which index to query based on the complexity of the question.
04managed ingestion

LlamaCloud and Managed Ingestion

For enterprise scale, LlamaIndex focuses on LlamaCloud.

  • Managed Ingestion: Handling PDF parsing, OCR, and Table extraction as a service.
  • Parsing as a Model: Using Vision-LLMs (like Gemini 3) to "Understand" layouts instead of using rule-based parsers.
05tools

Agents as Tools

LlamaIndex treats agents as high-level retrievers.

  • You can "wrap" a complex LlamaIndex query engine as a tool and give it to a LangGraph agent.
  • Benefit: The agent gets "Smart Data Access" without needing to know the technical details of the vector DB or Graph schema.
06questions

Interview Questions

Q: LangChain and LlamaIndex now both have "Graph/Workflow" features. How do you choose?

Strong answer: I choose LlamaIndex Workflows for Data-Intensive tasks where the main complexity is ingestion, multimodal parsing, and complex retrieval. Its event-driven architecture is more performant for massive parallel data processing. I choose LangGraph for Logic-Intensive multi-agent systems where the complexity is in the "Reasoning" and "Human-in-the-loop" logic. In many senior architectures, we use Both: LlamaIndex for the RAG engine and LangGraph for the overall agentic supervisor.

Q: What is the "Property Graph" in LlamaIndex and why is it superior to basic Vector RAG?

Strong answer: A Property Graph combines the Semantic flexibility of vectors with the Structural precision of a database. In basic RAG, you might find a chunk about "Project Alpha," but you don't know who owns it. In a Property Graph, the vector chunk is a node linked to a User node and a Timeline node. This allows for Global Reasoning (e.g., "Find all documents written by Tom in the last month about Project Alpha"). Basic RAG would likely miss many related nodes because they don't contain the exact keyword "Alpha."

07references

References

  • LlamaIndex. "The Workflows Framework: Event-Driven Agents" (2025)
  • Jerry Liu. "Data-Centric AI in the LLM Era" (2024/2025)
  • LlamaHub. "The Repository of 1000+ Data Loaders" (2025)

Next: DSPy: Programming Language Models

• summary · added by this rebuild

Key takeaways

01

Data-centric, not orchestration-centric

The unit is the Node, a chunk carrying metadata, summaries and parent-child links, and the library's breadth shows in its retrievers: summary, knowledge graph, tree and keyword.

02

Workflows are event-driven, not linear chains

Steps emit typed Events and the runtime is natively async, which the page gives as the reason it handles large-scale parallel data processing better than a linear chain does.

03

Property graphs add structure to vectors

Linking a chunk node to User and Timeline nodes supports queries like documents written by Tom last month about Project Alpha that keyword-free vector search would simply miss.

04

Choose by where the complexity lives

LlamaIndex Workflows for data-intensive ingestion, parsing and retrieval; LangGraph for reasoning and human-in-the-loop logic. The page notes senior architectures commonly run both together.