Jason Liu

Jason Liu created Instructor, a Python library that turns language-model responses into typed, validated data structures using Pydantic. He designed it as a small layer around model-provider SDKs, leaving developers control over their applications while handling response parsing and validation. His consulting and teaching extended that approach to retrieval-augmented generation (RAG): define the information a system needs, measure whether it retrieves it, and improve against actual user needs.[1][4][6]

Latent Space dates Instructor’s open sourcing to June 2023, under the name “OpenAI Function Call and Pydantic Integration Module.” In his April 2024 interview, Liu recalled writing it on a bullet train in Japan. Function calling let him separate the output schema from instructions and input data, then express complex schemas and validators in Pydantic. He wanted developers to retain access to chat messages rather than surrender their formatting to a framework.[1]

Liu took inspiration from Python’s requests library: a utility developers could use inside their own frameworks. As he put it, “People want to build their own frameworks, but people don't want to build like JSON parsing.” The Instructor repository demonstrates passing a Pydantic class through response_model, receiving a typed object, and retrying failed validations with the error message. It credits the Instructor community, Liu, and other contributors.[1][4]

In April 2024, Liu described function calling and JSON mode as different implementations of typed responses. His priority was a schema and usable object—with validation and IDE autocomplete—rather than valid JSON syntax alone. For extraction, he preferred a single schema because it could describe relationships among entities; he distinguished that preference from agent applications that need multiple tool calls.[1]

His examples included extracting receipt items and prices, identifying relationships between story characters, and translating questions into structured search requests. A question about what happened “this week,” he argued, needs a date interval rather than embedding similarity alone. He also advocated generating plans as directed acyclic graphs, with explicit nodes and edges, so developers could inspect the whole workflow and repair a failing component during execution.[1]

In April 2024, Liu explained his decision to pursue consulting while keeping Instructor a utility rather than expanding it into a venture-backed company. Consulting gave him access to varied applied problems and interesting data. He did not consider Instructor a billion-dollar business, and saw raising venture capital as introducing an expectation to build one.[1]

His May 2024 RAG improvement runbook, which credits a discussion with Hamel, recommends evaluating retrieval before concentrating on answer synthesis. Liu proposes generating questions for known text chunks, testing whether those chunks are retrieved, and establishing precision and recall baselines. He recommends searchable metadata such as dates, filenames, and ownership, alongside testing both full-text and vector search on the application’s own data.[6]

Liu illustrated the operational costs with an unnamed construction-data client whose separate project indices could fall out of synchronization, leaving an embedding retrievable while its text was missing. That experience informed his recommendation to consider text search, embeddings, and SQL over the same data object. He also recounted changing feedback wording because thumbs-down responses mixed answer correctness with complaints about tone and latency. Asking “Did we answer the question correctly? Yes or no.” made the intended evaluation question explicit.[6]

Liu argued that building a RAG application is not itself a business outcome: teams need to identify and measure goals such as reduced churn. In his September 2025 RAG series, he advocated report generation and structured analysis beyond isolated answers, arguing that reports support decisions, resource allocation, and standardized processes.[1][5]

On February 2, 2026, Liu announced that he was sunsetting 567 Labs and closing its courses that day. He also announced the open sourcing of written course content as transcripts edited into ebook format, preserving material on RAG improvement and independent consulting.[7]

Before Instructor and consulting, Liu worked on recommendation systems at Stitch Fix. His resume lists his employment there from 2018 to 2023, following a data-science role at Meta (Facebook) in 2017. It also lists a University of Waterloo Bachelor of Mathematics in computational mathematics and statistics, with studies spanning 2012–2017.[3][1]

At Stitch Fix, Liu built Flight, a shared recommendation framework. He recalled repeatedly refactoring bespoke systems: one team might use a small real-time Python application, while another used Spark for batch processing. Standardizing those systems consolidated maintenance and enabled pipeline-wide observability, making duplicated work, caching opportunities, and slow operations easier to identify. Liu named Bryan Bischof as one of the main users of his recommendation frameworks.[1][3]

Liu recalled a hand injury interrupting his ability to code for roughly a year or two, during which he took medical leave and worked as a technical lead keeping systems running. He later spent time in New York winding down technical work. After about five months experimenting with prompting, requests from startup friends for help with AI strategy developed into full-time consulting.[1]

He also described a reversal in his technical judgment. Initially skeptical of transformers and language models, he recalled writing an apology to Stitch Fix colleagues after seeing ChatGPT and shifting from computer vision and recommendations to LLM work. His subsequent advice was to document negative experiments with their inputs, outputs, and conditions—and specify what changes would justify trying them again.[1]