GPT Builder — Code Interpreter, Persistent Storage and Actions

If you want to build a useful GPT using GPT Builder, you are going to need a way to persist data between prompts. ChatGPT does this to an extent by itself but it has a bit of a dodgy memory. Sometimes it remembers what you are talking about but sometimes it doesn’t. To an extent this is the nature of generative AI but it would be nice to have finer control over what the Large Language Model (LLM) remembers. We want to maintain the generative benefits but supplement this with a dash of deterministic behaviour.

Figure 1. DALL-E’s Vision for Persistent Storage on ChatGPT

The code interpreter plugin is powered by generative Python and operates within a secure, sandboxed Python execution environment. It is the key to persistent storage in ChatGPT.

Code Interpreter Constraints

The easiest way to persist data within a session is to use the code interpreter, however, you need to be aware of its constraints. Some of these are not immediately obvious and may make debugging harder than it needs to be.

1. No Internet Access: The code environment doesn’t have access to the internet. This means it can’t fetch data from external sources or interact with online APIs. This is a bit weird, as the GPT itself can access the internet. To work around this, you have to scrape any data you need using a prompt and then feed this…