Description

The aim of this project is assigning a specific job to GPT. And trying to understand user messages' intent whether is an appointment request with specific date/time or not, basically labelling There are some rules assigned to GPT, you can see them as systemContent in /routes/appointment.ts The tests are written in Turkish language but the system I tried to build is trying to support other languages that OpenAI supports (systemContent is written in English) Its using gpt-3.5-turbo

Ps. The main part of this project is prompt engineering: https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-openai-api (this is the main part of this project)

How it works?

• Its an API endpoint that extracts the user intent for a new appointment from a natural language message using GPT4.
• Handles both relative and absolute time and date. For example today, tomorrow, this Friday, 30 of July (or from other languages for example in Turkish bugün)
• No "parsing" natural language - GPT4 is doing that for us.
• Given a message like Are you free for tomorrow at 2pm?, the API endpoint output a JSON similar to the following (this example assumes the date is 2023-07-29): { intent: "new_appointment", datetimeStr: "2023-07-30T14:00" }
• Included a suite of tests that call the API endpoint with a few different example inputs. Tests are asserting that the correct intent and datetime is extracted, also assert that intents other than new appointments are classified as "other."

How to setup

1. Clone this project
2. Install dependencies using yarn (yarn install)
3. Create an env file called .env.local under root folder
4. Add PORT env variable to .env.local file
5. Create your own OpenAI API key and add it as env variable in .env.local
6. So resulting .env.local file should look like

PORT=<example_port>
OPENAI_API_KEY=<your OpenAI API key>

7. Run yarn start:dev for development
8. Run yarn test to run tests
9. Make a GET request to /appointment-intent to extracts the user intent for a new appointment
10. Make a GET request to /appointment-classify to start building your own classifier model (please read CLASSIFICATION_README)
11. Make a GET request to /ask to chat with OpenAI API (please read ASK_README)

Things can be improved

In the project I tried to keep time zones into consideration but in some scenerios its working flaky. Based on the run time For example if you are in Turkey and run this project at night there can be some +/- 1day. You can observe this with running the tests (please check expected & received difference). So I advise you to run it at noon :D

Also since this project using GPT, we shouldnt expect 100% correctness.

• Sinan Talha KOSAR (sinantalhakosar@gmail.com)