LlamaIndex

LlamaIndex builds tools for connecting private and enterprise data to large language models and AI agents. Its open-source framework handles ingestion, indexing, retrieval, and response synthesis; its commercial products include the document parser LlamaParse and the knowledge management platform LlamaCloud. Jerry Liu and Simon Suo developed the organization from GPT Index, an exploratory project later renamed LlamaIndex. In March 2025, they were identified as its CEO and CTO co-founders, respectively.[2][5]

The organization's later work on document processing grew out of an earlier problem: giving a model access to more information than its prompt could hold. The project began in November 2022, before LlamaIndex's incorporation in April 2023.[1][3]

Liu began experimenting with GPT-3 after an internal LLM hackathon at Robust Intelligence in October 2022. GPT Tree Index organized text into a hierarchy rather than starting with embedding search. It built the tree bottom-up, using model prompts to summarize child nodes into parent nodes. To answer a question, GPT-3 traversed the tree top-down, selecting a relevant node and repeating the choice among its children until reaching a leaf. The initial project tested how a language model could organize and look up information.[1]

In December 2022, the project added embeddings, vector-store support, and initial loaders for Notion, Slack, and Google Drive. Liu recalled deliberately making the code more modular and usable as a library while considering latency, cost, and performance. In his June 2023 founder account, he said former colleague Simon Suo joined him two months into the project; together they expanded it into a framework for connecting private data with LLMs.[1][3][2]

LlamaHub launched with Jesse Zhang in February 2023 as a repository of data loaders. A sweepstakes with OctoML brought more than 50 loader submissions, according to the company's first-anniversary retrospective. Liu later said most LlamaHub contributions came from the community, while developers could write custom loaders when existing connectors omitted needed API features or metadata.[3][1]

Liu recalled deciding to pursue the company around December 2022 and leaving Robust Intelligence in early March 2023. LlamaIndex incorporated that April. At the end of April, version 0.6.0 rewrote the framework for greater modularity and composability. Data Agents, agent tools on LlamaHub, and a TypeScript package followed in July; LLM and embedding fine-tuning abstractions followed in August.[1][3]

The framework separated data management—ingestion, parsing, slicing, storage, and indexing—from data querying—retrieval, response synthesis, and multi-step interactions. A Document represented a source such as a PDF or database result; Nodes represented chunks and their relationships. Retrievers selected relevant context, postprocessors filtered or reordered it, and response synthesizers passed it to the language model. Query engines supported individual questions, chat engines supported conversations, and agents could use tools to read or write data.[2][1]

In his October 2023 interview, Liu called RAG “a very good hack”: it kept the model fixed while software selected information for its prompt. He argued that larger context windows would still leave tradeoffs involving cost, latency, and enterprise data volumes. He also valued visibility into retrieved documents and the possibility of gating access through external storage, while acknowledging that the LlamaIndex framework then lacked its own access-control support.[1]

Retrieval optimization extended beyond choosing an embedding model. Liu discussed chunking, metadata, hybrid search, and query decomposition as adjustable parts of the system. Crediting Joe from Vespa, he described keeping document embeddings frozen while training a query-side transformation, avoiding the need to reindex a corpus; he reported only limited gains. LlamaIndex also introduced retrieval evaluation modules. Liu recommended first checking the final answer, then using component benchmarks to diagnose problems, noting that good retrieval alone did not guarantee good generation.[1]

SEC Insights, launched in September 2023, put the library inside a full-stack chatbot for analyzing SEC 10-K and 10-Q filings. Liu said building it exposed bugs and missing functionality around callbacks, streaming, reliable intermediate responses, and source citations—including highlighted passages in a PDF viewer. The team open-sourced the application as a reusable template rather than turning it into a financial product.[3][1]

On February 20, 2024, LlamaParse entered public preview for PDFs, while a managed ingestion and retrieval API entered private preview for selected enterprise design partners. The proprietary parser converted complex PDFs and tables into structured Markdown that fed the open-source library's Markdown parsing and recursive retrieval algorithms. The managed API handled processing, chunking, embedding, synchronization, and retrieval over connected storage.[4]

Early access included a February hackathon co-hosted with Futureproof Labs and DataStax, where applications parsed building codes for accessory dwelling unit planning and real-estate disclosures. Mendable AI incorporated LlamaParse into its data stack. The launch also included collaborations with DataStax, MongoDB, Qdrant, and NVIDIA across ingestion, storage, and compute.[4]

LlamaCloud became generally available on March 4, 2025. Integrated with the open-source framework, it connected, parsed, extracted, and indexed unstructured data for downstream agents. It offered SaaS or virtual private cloud deployment, shared index management, role-based access control, and single sign-on. In a company-distributed testimonial, Rakuten's Yusuke Kaji said it reduced pipeline maintenance; Salesforce's Phil Mui credited the open-source framework's asynchronous workflow abstractions with helping build concurrent agents.[5]

On June 6, 2023, LlamaIndex announced $8.5 million in seed funding led by Greylock. On March 4, 2025, it closed a $19 million Series A led by Norwest Venture Partners with Greylock participating, bringing disclosed funding to $27.5 million at that date. On May 1, 2025, it announced minority equity investments through Databricks Ventures' AI Fund and KPMG Ventures; financial terms were not disclosed.[2][5][6]

LlamaIndex reported 200,000 monthly downloads and 16,000 GitHub stars in June 2023, then nearly 900,000 monthly downloads and more than 450 contributors that November. By March 2025, it reported more than 3 million monthly downloads across multiple open-source packages, customers including Rakuten, Carlyle, and Salesforce, and a separate waitlist of more than 10,000 organizations, including 90 Fortune 500 companies. Downloads measure package activity rather than unique developers, and waitlist organizations are distinct from customers.[2][3][5]

In the May 2025 investment announcement, Cemex Senior Data Scientist Daniel G Zapata said its small data-science team had produced 10 production-grade use cases in a couple of months. He attributed the faster delivery to LlamaCloud handling ingestion and retrieval work, allowing the team to focus on products.[6]

On March 19, 2026, Logan Markewich announced LiteParse, open-sourcing core LlamaParse processing as a local CLI and TypeScript-native library. It preserved layout by projecting text onto a spatial grid and supplied text, screenshots, and bounding boxes. Agents could read text quickly, then request screenshots for deeper visual reasoning. Office documents and images used conversion tools before entering the PDF pipeline. The proprietary cloud LlamaParse remained available for structured outputs and more difficult document processing.[7]

The LiteParse announcement described evaluating parsers with LLM-generated question-answer pairs from screenshots, manual dataset auditing, and an LLM judge. On September 2, 2026, LlamaIndex announced ExtractBench with Kaggle, shifting the task from page parsing to structured extraction across entire documents. The company reported a corpus of 370 enterprise documents totaling 4,869 pages, spanning eight business domains and 67 document types.[7][8]

According to the ExtractBench announcement, the full evaluation covered 14 systems using frozen schemas and deterministic, rule-based scoring without an LLM judge. It included frontier models, coding agents, and specialized extraction APIs, while the Kaggle leaderboard covered frontier models. The benchmark examined challenges such as long record lists, merged table cells, scanned pages, and handwritten fields—cases where extraction errors can propagate into an agent's decisions.[8]