- Get an OpenAI API key and add it to the environment variable
OPENAI_KEY pip install -r requirements.txtpython3 app.py
Firstly, understand that GPT is conversational. That means even if you say Please only respond with JSON, it will probably still say something like:
OK! I understand, here's your JSON:
{
"key": "Value"
}
Because while it's one of the world's most advanced AIs, it's still dumb as a brick. Here is a Python method that might help with extracting the JSON from a GPT response:
def getJSONfromResponse(json_text):
"""
Returns a JSON object from the openai api response. You'll need to get the openAIResponse
['choices'][0]["message"]['content'] if you're using the ChatCompletion endpoint
or openAIResponse['choices'][0]["text"] if you're just using the Completion endpoint.
"""
# Now try and load it as a JSON object
try:
# Remove or escape control characters
json_bytes = json_text.encode('utf-8')
json_text = json_bytes.decode('unicode_escape')
# Remove trailing comma if it exists
if json_text[-2] == ",":
json_text = json_text[:-2] + "}"
# If it's not just json, try and find the json in the string
json_start = json_text.find("{")
json_end = json_text.rfind("}") + 1
json_str = json_text[json_start:json_end]
# strict=false because we're not sure if it's valid json
return json.loads(json_str, strict=False)
except Exception as e:
print(f"Unable to find any JSON: {e}.")
return NoneIf you want to use priming, you have to send them to the openai.ChatCompletion.create endpoint (the openai.Completion.create doesn't support priming like the below, although you can prime with a standard prompt). Priming is an important part of working with LLMs and can save you a bunch of time writing prompts, here's an example:
priming_sequence = [
{"role": "system", "content": "You are an AI writer. You write business documents that are relevant to a certain standard I will provide later."},
{"role": "user", "content": "Mike: Hi, I'm a programmer setting up your environment."},
{"role": "assistant", "content": "It's nice to meet you."},
{"role": "user", "content": f"Mike: I will provide you a [title] of a blogpost. I want you to provide at least 500 words of [content] for that blogpost. To help you, you will receive the title of the blog and the top of the blog"},
{"role": "assistant", "content": "OK, I understand."},
{"role": "user", "content": 'A computer will read your response, so I need you to respond to me with ONLY a JSON object and NOTHING ELSE. The next thing I send is an example of what I want you to return. '},
{"role": "user", "content": '{"[title]": "[content]"}'},
{"role": "assistant", "content": "OK, I understand."},
]
priming_sequence.append(
{"role": "user", "content": f"The [title] is 'How I like to clean my shoes'. The title of the blog is 'Mikes Housework Blog', the topic of the blog is 'Housework, Cleaning and Home Maintenance'"},
)
options = {
"messages": priming_sequence,
"temperature": 1,
"max_tokens": 2048,
"n": 1,
"stop": None,
}Note the use of [title] and [content]. While I'm not sure if GPT treats these as variables, I've had more success with getting actually usable JSON by using these.