LlamaIndex

§ 01 · Who they are nowRight now, somewhere, a contract is being read by a llama.

It is 11:47pm and a paralegal at a Big Four firm is asleep. Her work is not. A LlamaIndex agent is moving through a 412-page master services agreement, extracting indemnity clauses, flagging exposure, and writing structured JSON to a database. The next morning she will check its work. It will be mostly right. She will keep her job and lose about four hours of misery.

This is the unglamorous, quietly important business LlamaIndex is in. Not chatbots. Not vibes. Documents. The unstructured pile of them that every enterprise has been hauling around since the photocopier. LlamaIndex turns that pile into something an AI agent can actually use.

The company is three years old, runs out of San Francisco, and as of March 2025 has $19 million in fresh Series A capital from Norwest Venture Partners and Greylock. It also has minority strategic checks from Databricks and KPMG, the kind of investors you collect when you are no longer a curiosity and not yet a household name.

§ 02 · The problem they sawLanguage models are smart. They are also, on their own, oddly illiterate.

Ask GPT a question about Napoleon and it will hold forth. Ask it a question about your employee handbook, your loan agreements, your scanned 1998 insurance binder, and it will hallucinate something confident and useless. The model never read the documents. It cannot. They live in your S3 bucket; the model lives behind an API. The two have not been introduced.

This was the gap in 2022 - a gap large enough to drive an industry through. Retrieval-augmented generation, or RAG, was the academic answer: chunk the documents, embed them as vectors, fetch the relevant bits at query time, and stuff them into the prompt. Simple in theory. In practice, every team trying to ship it ran into the same wall: parsing a PDF is hell, table extraction is worse, chunking strategies matter more than anyone admitted, and good luck doing any of this at scale on a Tuesday.

§ 03 · The founders' betJerry Liu was not trying to start a company. He was trying to read a document.

Liu, previously a machine learning engineer at Uber and Quora, started hacking on a tool to feed his own documents to GPT-3. He called it GPT Index. He pushed it to GitHub. A few months later he had thousands of stars, a Discord channel that would not stop pinging, and the dawning realization that the side project was now a job.

He renamed it LlamaIndex - partly because the GPT prefix was getting legally awkward, partly because llamas are funnier - and co-founded the company with Simon Suo, formerly a senior research engineer at Cruise. Greylock led an $8.5 million seed in mid-2023. The bet was specific: the LLM stack would split, like every stack before it, into models on one side and data infrastructure on the other. Anthropic and OpenAI could fight over the brains. LlamaIndex would build the librarian.

§ 04 · The productTwo things, really. One is free. One pays the bills.

The open-source library, still called LlamaIndex, is Python and TypeScript. It ingests data from roughly anywhere - Google Drive, Notion, S3, Postgres, Salesforce - chops it sensibly, indexes it, and gives developers a small, opinionated set of primitives for query engines, retrievers and agents. It has around 40,000 GitHub stars and roughly three million downloads a month. By the rough metrics of the open-source LLM world, it is a giant.

Then there is LlamaCloud, the commercial layer. Same primitives, but hosted, governed, observable, and supported. Inside it sits LlamaParse, the document parser that went generally available alongside the Series A. LlamaParse is the part of the stack that does the genuinely hard work: turning a complex PDF - charts, nested tables, three languages, scanned in 2007 - into structured text an LLM can use. By GA, it had handled hundreds of millions of documents.

LlamaExtract takes the next step, turning those parsed documents into validated JSON with citations and confidence scores. LlamaHub, the open catalog of community-built connectors, fills in the long tail.

Figures, approximate. The kind of numbers that make a VC lean forward and a competitor lean back.

§ 05 · The proofYou can argue with hype. You cannot argue with downloads.

LlamaIndex's claim on the LLM data layer is not theoretical. It shows up in the metrics most developers actually look at: pip installs, npm pulls, GitHub stars, integrations shipped. The chart below sketches the rough shape of that growth across the open-source library, with LlamaParse coming online as a secondary engine.

Customers are the other tell. KPMG is not just an investor; it is a user, putting LlamaIndex inside compliance review and document workflows. Insurance carriers use LlamaParse to sift through claims files. Banks use it on financial reports. Healthcare networks use it on records that have technically been digitized since 2009 but might as well still be in a manila folder. Manufacturing teams use it on multi-language equipment manuals that no human has read end to end since the Reagan administration.

§ 06 · The missionMake the rest of the stack boring.

The official mission statement is something close to: make it easy to build LLM-powered applications and agents over your own data. The unofficial one, which the team will tell you if you press, is shorter. They want the document layer of AI to feel like Postgres. Reliable. Boring. There when you need it. Nobody throws a party for their database. LlamaIndex would like the same dignity.

There is also a quieter ambition. The team is unusually invested in citations, confidence scores and traceability - features that do not demo well but matter enormously when the AI is reading a loan agreement or a discharge summary. The bet is that the next phase of enterprise AI will be won by the systems that can show their work, not just produce an answer.

§ 07 · Why it matters tomorrowBack to the paralegal. It is now morning.

She opens the file. The contract is annotated. Indemnity clauses are tagged with section numbers and confidence scores. One flag - a low-confidence extraction on an oddly worded limitation of liability - is highlighted in orange. She reads that one carefully, agrees with the agent's hesitation, and corrects it. The rest she approves in twenty minutes.

Multiply that by every contract, claim, policy, manual, filing, transcript and medical record currently sitting in a corporate folder. That is the surface area LlamaIndex is quietly trying to own. Not the model. Not the chat interface. The layer in between, where documents become knowledge and knowledge becomes an agent that can act on it. The unglamorous middle of the stack.

It is the kind of business that gets less press than it deserves and more market than it advertises. Which, on balance, is probably how the people building it would prefer it.