Jerry Liu is the co-founder and chief executive of LlamaIndex, an open-source framework and enterprise platform that helps AI systems understand documents and automate knowledge work. He has helped turn retrieval-augmented generation from a basic chatbot technique into infrastructure for parsing, evaluating, searching, and acting on organizational information.
From machine learning to LlamaIndex
Liu graduated summa cum laude from Princeton University in 2017 with a computer science degree and a certificate in finance. He worked on feed ranking at Quora, researched autonomous-driving systems at Uber ATG, and later led machine-learning monitoring efforts at Robust Intelligence. His research included reactive planning and LiDAR compression.
In November 2022, while experimenting with GPT-3, he created GPT Index to help language models use information outside their training data. The project became LlamaIndex, which he developed into a company with Simon Suo, a former Uber colleague. Liu became chief executive and Suo chief technology officer; the founders announced an $8.5 million seed round led by Greylock.
The company expanded from open-source retrieval tooling into managed ingestion, LlamaParse and LlamaCloud, agent orchestration, and document automation. Liu favors developer-controlled infrastructure that can adapt to changing models, enterprise permissions, heterogeneous data, and specialized workflows.
- Production-ready retrieval requires measurement. Liu separates ingestion, retrieval, and answer synthesis because each fails differently: missing evidence, irrelevant passages, stale documents, and flawed generation demand different remedies. His approach to production-ready RAG starts with task-specific benchmarks and independent retrieval evaluation, then improves chunking, metadata filters, hybrid search, and parsing before introducing more expensive agent architectures. With small-to-big retrieval, systems locate precise passages first and supply broader surrounding context for synthesis.
- Document structure carries meaning. Tables, headings, page layouts, and images encode relationships that disappear when PDFs are flattened into undifferentiated text. Liu illustrates the problem with a Caltrain timetable: collapse its rows and columns, and a model can confidently return the wrong train time. LlamaParse preserves meaningful structure, while more advanced pipelines combine conventional parsing, language and vision models, and agentic validation.
- Longer context windows still need retrieval. Larger prompts can reduce awkward chunking and support document-level reasoning, but they cannot make indiscriminate ingestion economical or transparent across large enterprise collections. Liu’s long-context retrieval work emphasizes selective access, source visibility, latency, and the appropriate amount of context for each task.
- Document agents need complete toolboxes. Useful systems require enterprise connectors, synchronized permissions, parsing, indexing, semantic search, structured queries, and document manipulation. Liu distinguishes interactive assistants from more constrained background automations, with human review before consequential downstream actions. His spreadsheet agent interprets irregular Excel layouts, normalizes tables, and equips agents with spreadsheet-specific tools. He has also developed architectures in which specialist agents operate as independently orchestrated services.
- Agent economics depend on the whole workflow. Liu’s LiteParse samples demonstrate fast, local, model-free document preparation. His two-pass document processing approach first processes large collections inexpensively, then applies costlier visual interpretation only to relevant pages. He also argues for model routing at the harness layer: optimizing accuracy and cost requires considering the complete combination of models, tools, orchestration, and task context.
On September 9, 2026, Liu announced a native LlamaParse connector for ChatGPT, following its Claude connector. His explanation separates specialized document parsing and extraction from the assistant’s higher-level reasoning: the connector exposes LlamaParse endpoints for preparing document context and extracting fields.
▶ Watch ↗
▶ Watch ↗
