- Pre-requisites
- OpenAI API
- Typescript (Async, Await, basic syntax)
- GPT Chat completion "Roles"
- Training cutoff date
- JSON Schema: https://bjdash.github.io/JSON-Schema-Builder/
- Don't know JSON Schema? ChatGPT does.
- Demonstartion: Spam detection without functions.
- The problem
- Bridging natural language and computer language.
- Cleanly interface software with LLM results.
- Avoid string manipulation and hacks.
- The solution
- New GPT "function_calls"
- What is it?
- You give GPT a list of functions your application has.
- GPT intelligently selects the correct function based on context.
- Real-world applications
- Robotics (call robot function based on incoming data)
- Text-based interfaces (run application feature based on English requests)
- Giving GPT the ability to retrieve data after the training cutoff date.
- Caveats
- Only works with some models:
gpt-4-0613,gpt-3.5-turbo-0613 - non-helpful 400 errors are common starting out.
- Only works with some models:
- Looking forward
- SEE ALSO: Type Chat: https://microsoft.github.io/TypeChat/blog/introducing-typechat/