jinja2_openai_gpt_example.py

# This code snippet shows one way we could use gpt inside a Jinja2 template
# to hack-around the context limitations. Have fun! And remember, the ART is the prompt...
# MAKE SURE you modify the 'set titulo' (titulo is title in spanish)
# Author: Arturo "Buanzo" Busleiman <buanzo@buanzo.com.ar>

import os
import openai
import json
from jinja2 import Template

# Set up OpenAI API authentication
openai.api_key = os.environ["OPENAI_API_KEY"]

# Define your template using Jinja2 syntax
# In this example, the template includes a title, an abstract/scoop,
# Table of Contents and then content. You may create markdown, html, reStructuredText, etc.
template = Template("""
{% set titulo = "Super useful Python Libraries" %}
{% set toc = gpt3(titulo, type="Array", len=5, random=True, lang='Spanish') %}{{ titulo }}
{% set abstract = gpt3("Article scoop about {}: {}".format(titulo,','.join(toc)), lang='Spanish', type='Content') %}

TITULO: {{ titulo }}

ABSTRACT: {{ abstract }}

Table of contents
-----------------

{% for item in toc %}
* {{ item }}
{% endfor %}

{% for item in toc %}
{{ item }}
  {% set content = gpt3('{}: {}'.format(article_object,item), type='Content', lang='Spanish') %}
  {{ content }}
{% endfor %}
""")

# Define a function to call the OpenAI API
def gpt3(prompt, **kwargs):
    prompt = prompt.strip()

    # Adapt response to whatever content-type user requires
    if kwargs['type'] == 'Array':
        newprompt = f"Provide a python-compatible one-dimensional json response for: {prompt}."
    elif kwargs['type'] == 'Content':
        newprompt = f"Write content for an article section about {prompt}."
        
    lang = None
    if 'lang' in kwargs:
        lang = kwargs['lang']
        newprompt += f"Respond in {lang}."
    
    cant = None
    if 'len' in kwargs:
        cant = kwargs['len']
        newprompt += f"Provide only {cant} items."

    response = openai.Completion.create(
        engine="text-davinci-003",
        prompt=newprompt,
        max_tokens=2048,
        n=1,
        stop=None,
    )

    raw_response = response.choices[0].text.strip()

    respuesta = None

    # Now adapt response to whatever content-type user requested:
    if kwargs['type'] == 'Array':
        respuesta = json.loads(response.choices[0].text.strip())
    elif kwargs['type'] == 'Content':
        respuesta = response.choices[0].text.strip()

    return respuesta

# Render the completed template and output it to the command prompt
completed_template = template.render(gpt3=gpt3)
print(completed_template)