LlamaIndex
Last reviewed
Apr 28, 2026
Sources
21 citations
Review status
Source-backed
Revision
v7 ยท 4,444 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 28, 2026
Sources
21 citations
Review status
Source-backed
Revision
v7 ยท 4,444 words
Add missing citations, update stale details, or suggest a clearer explanation.
LlamaIndex is an open-source data framework designed for building large language model (LLM) applications, with a particular focus on retrieval-augmented generation (RAG). Originally released as GPT Index in November 2022 by Jerry Liu, the framework provides the infrastructure for connecting LLMs with custom data sources, enabling developers to build production-ready AI applications that can query private documents, databases, and APIs. LlamaIndex has become one of the most widely adopted frameworks in the LLM application ecosystem, with over 47,000 GitHub stars and a growing commercial offering through LlamaCloud [1].
LlamaIndex began as a side project in November 2022, when Jerry Liu started experimenting with ways to help large language models reason over proprietary data outside their training sets. Liu, a Princeton graduate (B.S.E. in Computer Science, summa cum laude, 2017) and former research scientist at Uber, had been trying to build a sales bot but ran into difficulty feeding proprietary data to GPT-3. That frustration led him to create a simple tool called GPT Index, which he released with his first commit and tweet on November 8, 2022 [2].
What started as a basic tree index for organizing information quickly attracted developer attention as the broader AI community grappled with the challenge of grounding LLM outputs in real-world, private data. Early feature releases in December 2022 included support for indexing embeddings with vector stores, and initial data loaders for Notion, Slack, and Google Drive. By January 2023, GPT Index had hit the GitHub trending page for the first time [3].
After two months of working on the project, Liu teamed up with Simon Suo, his former colleague at Uber's Advanced Technologies Group (ATG), to build the product and community. Suo, who studied at the University of Toronto and had spent five years at Uber ATG and Waabi working on automated driving research under Raquel Urtasun, became co-founder and CTO [4].
The project was renamed from GPT Index to LlamaIndex in early 2023 to reflect its broader applicability beyond OpenAI models. The name references the llama, a nod to the growing ecosystem of open-source LLMs (including Meta's LLaMA family), while "Index" emphasizes the framework's core strength in data indexing and retrieval.
Jerry Liu is the co-founder and CEO of LlamaIndex. Before founding the company, Liu built a career spanning machine learning research and engineering across several notable organizations:
| Period | Role | Organization |
|---|---|---|
| 2014 | Software Engineering Intern | Apple |
| 2015 | Software Engineering Intern | Quora |
| 2016 | Software Engineering Intern | Two Sigma |
| 2017-2018 | Machine Learning Engineer | Quora |
| 2018-2019 | AI Resident | Uber |
| 2019-2021 | Research Scientist | Uber |
| 2021-2023 | ML Engineering Manager | Robust Intelligence |
| 2023-present | Co-Founder and CEO | LlamaIndex |
At Princeton, Liu earned the Shapiro Prize for Academic Excellence (2014) and graduated with a 3.97 GPA. His research experience at Uber focused on machine learning for autonomous driving, while at Robust Intelligence he led the ML monitoring team, gaining experience in building reliable AI systems. This combination of research depth and engineering leadership shaped the design philosophy behind LlamaIndex [5].
Simon Suo is the co-founder and CTO of LlamaIndex. Suo attended the University of Toronto and worked at Uber ATG as a Research Scientist (2018-2021) before joining Waabi as a Senior Research Scientist (2021-2023), where he contributed to automated driving research. His background in large-scale machine learning systems and autonomous driving infrastructure informs LlamaIndex's technical architecture [4].
LlamaIndex provides a layered architecture that handles the full pipeline from raw data to LLM-powered responses. The framework is organized around several core abstractions that work together to enable sophisticated data-augmented AI applications.
A Document is the foundational data unit in LlamaIndex. It serves as a generic container around any data source, whether that is a PDF file, an API response, a database row, or a web page. Documents can be constructed manually by the developer or created automatically through data loaders. Each Document stores the raw text content along with two important attributes [6]:
A Node represents a "chunk" of a source Document. Nodes are first-class citizens in LlamaIndex and form the basic unit that gets indexed and retrieved. When data is loaded into the framework, it is typically broken down into Node objects through a process called parsing or chunking. Each Node encapsulates [6]:
By default, every Node derived from a Document inherits the same metadata from that Document. Developers can customize the chunking strategy through NodeParser classes, choosing between fixed-size chunking, sentence-based splitting, semantic chunking, or custom approaches.
Once data is ingested and parsed into Nodes, LlamaIndex organizes them into index structures optimized for different retrieval patterns. The framework offers several indexing strategies:
| Index Type | Description | Best For |
|---|---|---|
| Vector Store Index | Generates semantic embeddings for each Node and stores them for similarity search | Semantic similarity search |
| Summary Index | Creates summaries of document chunks | Broad question answering |
| Tree Index | Organizes data in a hierarchical tree structure | Multi-level summarization |
| Keyword Table Index | Extracts keywords and maps them to Nodes | Exact keyword matching |
| Knowledge Graph Index | Builds a knowledge graph from extracted entities and relationships | Relationship-based queries |
The Vector Store Index is the most commonly used, as it powers the standard RAG workflow. It splits documents into Nodes, generates embeddings using a configurable embedding model, and stores them in a vector database for fast similarity search. LlamaIndex integrates with dozens of vector store backends including Pinecone, Weaviate, Chroma, Qdrant, Milvus, and FAISS.
Retrievers are responsible for fetching the most relevant Nodes from an index given a user query. LlamaIndex provides several retriever types to support different search strategies [7]:
| Retriever Type | Description |
|---|---|
| Vector Index Retriever | Finds Nodes by semantic similarity using embedding cosine distance |
| Keyword Retriever | Matches Nodes based on keyword frequency (BM25) |
| Hybrid Retriever | Combines vector and keyword retrieval with configurable alpha weighting |
| Router Retriever | Uses an LLM to decide which sub-retriever to route the query to |
| Auto-Merging Retriever | Merges small child Nodes into larger parent Nodes when enough children are retrieved |
| Recursive Retriever | Traverses node relationships recursively to gather broader context |
Hybrid retrieval, which blends semantic vector search with keyword-based BM25 search, is particularly effective in production settings. LlamaIndex exposes an alpha parameter that controls the balance between the two methods, allowing developers to tune retrieval behavior for their specific domain.
Query engines combine a Retriever and a Response Synthesizer into an end-to-end pipeline. When a user submits a question, the query engine retrieves relevant Nodes from the index and uses an LLM to synthesize a coherent answer. LlamaIndex provides several query engine types [8]:
| Query Engine Type | Description |
|---|---|
| Simple Query Engine | Retrieves top-k relevant Nodes and synthesizes a response in a single pass |
| Sub-Question Query Engine | Breaks complex questions into sub-questions, retrieves answers for each, and combines them |
| Router Query Engine | Routes queries to different indices or retrieval strategies based on query content |
| SQL Query Engine | Translates natural language questions into SQL queries for structured databases |
| Pandas Query Engine | Enables natural language queries over dataframes |
| Citation Query Engine | Provides source citations alongside generated answers |
Query engines can be composed and chained together. For instance, a router engine might send factual questions to a vector index while directing analytical questions to a SQL engine, combining the results into a unified response.
Response synthesizers take the retrieved Nodes and the original query, then produce a natural language answer using an LLM. LlamaIndex offers multiple synthesis strategies:
Retrieval-augmented generation is the core use case that LlamaIndex was built to support. A typical RAG pipeline in LlamaIndex follows five stages [8]:
This five-stage pipeline can be set up in just a few lines of code for simple use cases, but each stage is fully customizable for production deployments. Developers can swap embedding models, vector stores, chunking strategies, retrieval methods, and LLM providers without changing the overall pipeline structure.
Beyond the basic pipeline, LlamaIndex supports several advanced RAG patterns:
LlamaIndex agents extend the framework beyond simple retrieval into autonomous reasoning and action. An agent can use tools (including query engines, APIs, and custom functions) to accomplish complex tasks that require multiple steps of reasoning [9].
Agents in LlamaIndex support:
The framework provides built-in agent types compatible with various LLM providers, as well as a lower-level agent protocol for building custom agent architectures.
Agentic RAG represents an evolution of the basic RAG pipeline where autonomous agents handle the retrieval and synthesis process. Instead of a fixed retrieve-then-generate pipeline, an agentic RAG system can dynamically decide what to retrieve, when to retrieve, and how to synthesize the final answer [10].
LlamaIndex's agentic RAG architecture supports several patterns:
This architecture scales naturally as new documents are added, since each receives a dedicated sub-agent without requiring changes to the overall system design.
Introduced as a beta feature in August 2024 and released as a stable 1.0 version in June 2025, Workflows is an event-driven, async-first orchestration engine that allows developers to define multi-step pipelines where each step can involve retrieval, generation, tool calls, or custom logic [11].
Workflows support branching, looping, conditional execution, and parallel processing, making them suitable for complex enterprise automation scenarios. The architecture uses typed events that flow between steps, with each step decorated to specify which events it accepts and emits.
Key features of Workflows 1.0 include:
| Feature | Description |
|---|---|
| Typed Workflow State | Full type safety for state management in both Python and TypeScript |
| Resource Injection | Dynamic injection of database clients and external resources at runtime |
| Observability | Integration with OpenTelemetry, Arize Phoenix, and other monitoring tools |
| Stateful Execution | Workflows can start, pause, and resume across sessions |
| Standalone Package | Available as llama-index-workflows (Python) and @llamaindex/workflow-core (TypeScript) |
Workflows have grown significant enough to become their own dedicated package, separate from the core LlamaIndex library. They are the recommended abstraction for building complex agentic systems, replacing earlier chain-based patterns [11].
LlamaHub is the community-driven ecosystem of data connectors, tools, and integrations for LlamaIndex. Originally maintained as a separate GitHub repository, LlamaHub was merged into the core LlamaIndex Python repository starting with version 0.10 in early 2024. The integrations are browsable at llamahub.ai [12].
Data connectors pull information from various sources and transform raw data into a standardized Document representation. LlamaHub provides loaders for over 300 integration packages:
| Source Category | Examples |
|---|---|
| Cloud Storage | Amazon S3, Google Drive, Dropbox |
| Productivity Tools | Notion, Slack, Confluence, Google Docs |
| Databases | PostgreSQL, MySQL, MongoDB, Snowflake |
| File Formats | PDF, CSV, DOCX, EPUB, HTML, Markdown |
| APIs | Twitter, Wikipedia, YouTube, GitHub |
| Enterprise Systems | Salesforce, Hubspot, Jira, Zendesk |
| Vector Stores | Pinecone, Weaviate, Chroma, Qdrant, Milvus |
| LLM Providers | OpenAI, Anthropic, Cohere, Google Gemini, Mistral |
Each connector handles the specifics of authentication, pagination, and data extraction for its respective source, presenting a consistent interface to the rest of the framework. Developers can also write custom connectors for proprietary data sources.
LlamaIndex and LangChain are the two most prominent frameworks for building LLM applications, and they are frequently compared. While there is overlap in their capabilities, each framework has a distinct focus [13].
| Aspect | LlamaIndex | LangChain |
|---|---|---|
| Primary Focus | Data indexing, retrieval, and document processing (RAG) | General-purpose LLM application building and workflow orchestration |
| Core Strength | Deep data connection, indexing quality, and retrieval precision | Flexible chaining of models, tools, and complex workflows |
| Learning Curve | Gentler; high-level abstractions for common RAG patterns | Steeper; more configuration options and concepts |
| Data Connectors | 300+ via LlamaHub | Fewer built-in, but extensible |
| Agent Framework | Workflows (event-driven, async-first) | LangGraph (graph-based state machines) |
| Agent Capabilities | Growing; focused on document-centric and retrieval agents | Mature; broad tool and agent ecosystem |
| Retrieval Accuracy | 35% boost reported in 2025 through advanced retrieval techniques | Strong, with extensive retriever options |
| Commercial Product | LlamaCloud (managed RAG, parsing, agent deployment) | LangSmith (observability, evaluation, monitoring) |
| Best For | Document-heavy RAG applications, enterprise knowledge systems, document AI | Complex multi-step workflows, tool-heavy agents, chatbot applications |
| License | MIT | MIT |
| Community Size | ~47,800 GitHub stars | ~100,000+ GitHub stars |
LlamaIndex is generally considered the better choice for straightforward RAG applications, especially those involving large volumes of documents where indexing quality and retrieval precision are paramount. LangChain excels when the application requires chaining multiple models, tools, and APIs into complex workflows that go beyond data retrieval [13].
In practice, many developers use both frameworks together. LlamaIndex handles data ingestion and retrieval, while LangChain orchestrates the broader application logic. The two frameworks are interoperable, with LlamaIndex query engines usable as tools within LangChain agents.
LlamaCloud is the commercial managed service built on top of the LlamaIndex open-source framework. Launched with general availability in March 2025 alongside the Series A announcement, it provides a cloud-hosted platform for building and deploying data agents and RAG applications without managing infrastructure [14].
LlamaParse is LlamaCloud's flagship document parsing service, designed to handle complex document formats that simple text extraction cannot process accurately. It uses AI-powered parsing to extract structured content from PDFs, presentations, spreadsheets, and other document types, preserving tables, headers, and layout information [15].
LlamaParse v2, launched in 2025, introduced a simplified four-tier configuration system and up to 50% cost reduction compared to v1:
| Tier | Description | Use Case |
|---|---|---|
| Fast | Quick extraction with basic formatting | High-volume simple documents |
| Cost Effective | Balanced accuracy and cost | Standard business documents |
| Agentic | AI-powered layout understanding | Complex documents with tables and figures |
| Agentic Plus | Maximum accuracy with multi-model approach | Critical documents requiring highest fidelity |
The v2 redesign replaced complex configuration modes with automatic model routing, making it simpler for developers to get accurate parsing results without manual tuning.
LlamaAgents provides one-click deployment of document agents, allowing organizations to deploy AI agents that can process, analyze, and act on document content. Built on the open-source Workflows library, it supports persistent memory, filesystem tools, and integration with MCP (Model Context Protocol) servers [16]. Pre-built agent templates cover common use cases including invoice processing, contract review, claims handling, and document Q&A.
| Service | Description |
|---|---|
| LlamaSheets | Transforms disorganized spreadsheets into AI-ready data through intelligent region classification and multi-stage processing with 40+ cell features |
| LlamaSplit | Automatically segments bundled multi-document files into distinct sections by analyzing page content and grouping consecutive pages by category |
| LlamaExtract | Structured extraction service for pulling specific data fields from documents, with a Table Row Mode for handling repeating entities like catalogs |
| Managed Index | Managed ingestion and RAG service with hosted vector stores |
LlamaCloud can be deployed as a SaaS installation or within a virtual private cloud, and includes enterprise features such as role-based access control and single sign-on. Pricing follows a credit-based model: a free tier with 10,000 credits, a Starter plan at $50/month with 50,000 credits, and a Pro plan at $500/month with 500,000 credits [14].
In 2025, LlamaIndex introduced the concept of Agentic Document Workflows (ADW), an architecture combining document processing, retrieval, structured outputs, and agentic orchestration to enable end-to-end knowledge work automation [16]. An ADW system can:
This approach represents a shift from passive document retrieval toward active document processing, where AI agents can handle tasks like invoice processing, contract review, compliance checking, and report generation with minimal human intervention. LlamaIndex reported achieving 90%+ pass-through rates with agentic workflows, compared to 60-70% with legacy systems [17].
LlamaIndex is available in both Python and TypeScript, with the Python library being the more mature and feature-complete implementation. The framework follows a modular design that allows developers to swap components at every layer.
| Abstraction | Purpose |
|---|---|
| Document | Raw data unit from a connector |
| Node | Chunk of a Document with metadata and relationships |
| Index | Data structure for organizing and retrieving Nodes |
| Retriever | Fetches relevant Nodes from an Index given a query |
| Response Synthesizer | Generates a natural language response from retrieved Nodes |
| Query Engine | Combines Retriever and Synthesizer into an end-to-end pipeline |
| Agent | Autonomous entity that uses tools and reasoning to accomplish tasks |
| Workflow | Event-driven, multi-step orchestration pipeline with branching and state management |
LlamaIndex integrates with a wide range of LLM providers, embedding models, and vector stores:
LlamaIndex added native support for the Model Context Protocol (MCP) in 2025, allowing agents to connect to any MCP-compatible server and use the tools it exposes. The llama-index-tools-mcp package enables connecting to MCP servers in a single line of code. LlamaIndex also provides its own MCP server for documentation search, offering tools for searching, grepping, and reading docs directly from coding environments like Claude Code [18].
LlamaIndex can build and query knowledge graphs from unstructured text. The Knowledge Graph Index extracts entities and relationships from documents, constructing a graph that enables relationship-based queries. This is particularly useful for domains where connections between entities matter, such as biomedical research, legal analysis, and supply chain management.
LlamaIndex includes built-in evaluation tools for measuring retrieval quality and response accuracy. The framework integrates with observability platforms to provide tracing, logging, and performance monitoring for production RAG applications. Developers can track metrics like retrieval precision, response faithfulness, and latency. The Workflow Debugger, shipped in 2025, provides real-time event logs, workflow visualization, and run comparison capabilities [17].
LlamaIndex has raised over $27.5 million in total funding across multiple rounds.
| Round | Date | Amount | Lead Investor | Other Investors |
|---|---|---|---|---|
| Seed | June 2023 | $8.5M | Greylock Partners | Angel investors including Jack Altman, Lenny Rachitsky, Mathilde Collin (Front), Raquel Urtasun (Waabi) |
| Series A | March 2025 | $19M | Norwest Venture Partners | Greylock Partners (existing) |
| Strategic | May 2025 | Undisclosed | N/A | Databricks Ventures, KPMG Ventures |
The seed round in 2023 was used to build an enterprise offering on top of the open-source framework and to grow the team. The Series A round in 2025 was directed toward expanding the team (approximately 20 people at the time) and accelerating development of LlamaCloud and the agentic platform [14]. The strategic investments from Databricks and KPMG reflect growing adoption in enterprise environments. KPMG's investment signals uptake within professional services, while Databricks' backing aligns LlamaIndex with one of the largest data and AI platforms [19].
LlamaIndex powers a variety of LLM application patterns:
| Use Case | Description |
|---|---|
| Enterprise Knowledge Base | Employees query internal documents, wikis, and databases using natural language |
| Customer Support | AI agents retrieve product documentation and past tickets to answer customer questions |
| Research Assistant | Researchers query large corpora of academic papers, extracting findings and connections |
| Legal Document Review | Agents analyze contracts, extract clauses, and flag compliance issues |
| Financial Analysis | Natural language queries over financial reports, SEC filings, and market data |
| Code Documentation | Developers query codebases and documentation to understand systems and debug issues |
| Invoice Processing | Agentic workflows automatically extract, validate, and route invoice data |
| Compliance Checking | Document agents review policies and regulations against organizational documents |
As of early 2026, LlamaIndex has evolved from a simple RAG library into a comprehensive platform for building document-centric AI applications. Several trends define its current trajectory:
llama_cloud_services repository would be deprecated by May 2026, with users encouraged to migrate to new packages (llama-cloud>=1.0 for Python, @llamaindex/llama-cloud for TypeScript) offering improved performance and active development [21].Jerry Liu has articulated a vision for 2026 in which "agents go from workflows to employees," suggesting that the next phase of LlamaIndex development will focus on increasingly autonomous document processing agents that can handle complex knowledge work with minimal human oversight [16].