OpenAI Assistants: JSON and Tools

Assistants are a new (beta) part of OpenAI’s API. They are powerful but also have some interesting quirks. In this talk, I’ll start with the basics of configuring an Assistant in the Playground. Next I will show a few different ways of getting assistants to respond with JSON output, including workarounds for current limitations when using function calling or file search/vector store. Then we’ll get into working via the API - understanding concepts like threads, runs and the various flows needed to programmatically engage with Assistants. I’ll conclude by discussing our observations about accuracy, consistency and cost.

Code that I will demo includes a NextJS app + API that uses the node OpenAI API library to lookup company information via an assistant and return it in JSON format. It then uses this information to query another assistant that has access to a vector store database (single json file) + file search to recommend projects that would be appropriate for the company based on the files in the database. Then it uses the same assistant to allow for creating a custom project brief.

If there’s time, I will also briefly show PHP code that uses a more complex tool calling flow.