Jerry Liu

Jerry Liu created LlamaIndex, a data framework for building applications with large language models, and cofounded its company with Simon Suo. He helped turn an experiment in model-guided information retrieval into a modular open-source library; his later work centers on document context for agents.[2][3][1][7]

By 2026, Liu was framing document processing as an enduring bottleneck for agents, shifting LlamaIndex’s focus from a broad retrieval-augmented generation (RAG) framework toward document OCR infrastructure, according to a March 10 Latent Space roundup. The June 3 Chain of Thought synopsis described LlamaCloud and LlamaParse as the managed infrastructure behind this direction, and characterized his bet as one on context quality as agent loops absorb more scaffolding.[4][7]

He also emphasized the user’s role in making models useful. Latent Space’s March 5, 2026 commentary quoted him saying: “The Model Harness is Everything — the biggest barrier to getting value from AI is your own ability to context and workflow engineer the models.” His statement particularly emphasized broad, horizontal tools.[5]

Liu gave a concrete example in a self-identified LlamaIndex community comment about LiteParse. He described an agent first parsing text to represent an entire document, then requesting screenshots of selected pages. These tools, he explained, spare the agent from writing and executing all the supporting code from scratch. The accompanying announcement, authored by tuanacelik, explained that LiteParse preserves spatial layout rather than reconstructing document structure, passing that representation to the language model.[6]

These later positions followed Liu’s earlier work on recommendation systems, computer vision and the practical limits of supplying data to language models. His account of that work explains the technical problems and relationships behind LlamaIndex’s beginnings in 2022.[1]

In his October 2023 Latent Space interview, Liu recalled interning at Quora in 2015 and joining full time in 2017. He worked on recommendation systems for about a year, using user and item embeddings to rank content by user preferences. He also read deep-learning papers and wrote Quora answers to explain new techniques, describing his later educational posts on Twitter as a continuation of that practice. Quora’s mission of spreading knowledge, he said, had helped draw him to the company.[1]

At Uber, Liu worked on deep-learning training for self-driving and computer vision. He subsequently became a machine learning engineering manager at Robust Intelligence. There, he worked on the same team as Harrison, the creator of LangChain, for roughly two years. Liu recalled an October 2022 GPT-3 hackathon, followed by a vacation; when he returned, Harrison had started LangChain. Liu then began his own experiments with supplying GPT-3 with information that exceeded its context window.[1][3]

Liu initially conceived GPT Tree Index as an experiment in whether language models could organize and traverse information themselves, rather than as a developer toolkit centered on embeddings. It built a tree from text by summarizing child nodes into parent nodes, working upward. To answer a question, GPT-3 selected promising nodes and descended through their children until reaching a leaf. The experiment tested model-guided organization and lookup; it was not presented as an attempt to outperform existing embedding-based retrieval.[1]

Around December 2022, Liu and his collaborators deliberately moved toward a modular open-source library. They added embeddings, considered latency, cost and performance, and expanded the scope to data ingestion and querying. LlamaHub supplied data loaders, with much of that work contributed by the community. Liu supported custom loaders when existing integrations did not expose the APIs or metadata a developer needed.[1]

Liu recalled deciding around December to pursue LlamaIndex over an earlier startup idea involving video data, then officially leaving Robust Intelligence in early March 2023. Before committing, he considered whether larger context windows, fine-tuning or foundation-model providers’ own orchestration tools would undermine the opportunity. His reasoning was that enterprise datasets could still exceed practical prompt sizes, while transferring huge amounts of data for each answer would impose costs and latency. Retrieval offered control over how much context to supply, even as models’ context windows grew.[1]

Liu knew Simon Suo from Uber ATG in Toronto and sought a cofounder with strong technical judgment for the CTO role. He described Suo as someone whose understanding of machine learning, engineering approach and willingness to undertake startup work fit those criteria. Discussing company values, Liu emphasized passion and integrity, including respecting code ownership and giving attribution.[1][3]

In the 2023 interview, Liu called RAG a “very good hack”: it keeps the model fixed while software retrieves information and assembles prompt context. He recommended it as an accessible starting point, while arguing that hand-selected components leave room for learned retrieval and more integrated optimization. His expectation that fine-tuning would become more important was accompanied by a practical qualification: his own informal attempt to teach a model text knowledge through generated question-and-answer examples had not worked particularly well.[1]

Liu identified visibility into retrieved documents and access control through external storage as architectural advantages of RAG. He distinguished those possibilities from LlamaIndex’s capabilities at the time: he described the library as an orchestration framework without built-in access-control support.[1]

His optimization advice included the operational consequences of changing embeddings. Fine-tuning an embedding model can change its embedding space and require reindexing every document. Liu credited Joe from Vespa with suggesting an alternative: keep document embeddings frozen and train a transformation on query embeddings. Liu reported modest improvements from this approach, rather than large performance gains.[1]

The team’s SEC Insights application, which analyzed SEC 10-K and 10-Q filings, served as a production test for the library. Liu said building it exposed bugs and requirements around streaming, callbacks, reliable responses and visible citations. The team chose to open-source the application as a reusable full-stack template rather than sell a financial application.[1]

Liu recommended that developers build RAG from scratch to understand its parameters and learn how to improve applications. He contrasted that understanding with the convenience—and opacity—of LlamaIndex’s three-line quick start, while acknowledging that the quick start could serve some users. For evaluation, he suggested beginning with an end-to-end sanity check, then diagnosing components and building retrieval benchmarks. Good retrieval, he cautioned, does not guarantee good generation; for agents assigned concrete tasks, he emphasized whether they actually completed the task.[1]