app.py

import streamlit as st
import os
import openai
from pymongo import MongoClient
from datetime import datetime
import random

# Schema Versions
# 1. First version, using text-davinci-003 model
# 2. Switched to gpt-3.5-turbo model
# 3. Logging the model as well


# you need to set your OpenAI API key as environment variable
openai.api_key = st.secrets["API_KEY"]

MOVIES_EXAMPLE_DOC = """{
  _id: ObjectId("573a1390f29313caabcd4135"),
  genres: [ 'Short' ],
  runtime: 1,
  cast: [ 'Charles Kayser', 'John Ott' ],
  num_mflix_comments: 0,
  title: 'Blacksmith Scene',
  countries: [ 'USA' ],
  released: ISODate("1893-05-09T00:00:00.000Z"),
  directors: [ 'William K.L. Dickson' ],
  rated: 'UNRATED',
  awards: { wins: 1, nominations: 0, text: '1 win.' },
  lastupdated: '2015-08-26 00:03:50.133000000',
  year: 1893,
  imdb: { rating: 6.2, votes: 1189, id: 5 },
  type: 'movie',
  tomatoes: {
    viewer: { rating: 3, numReviews: 184, meter: 32 },
    lastUpdated: ISODate("2015-06-28T18:34:09.000Z")
  }
}"""

MOVIES_EXAMPLE_QUESTIONS = [
    (
        "How many fantasy or horror movies from the USA with an imdb rating "
        "greater than 6.0 are there in this dataset?"
    ),
    (
        "Which movies were released on a Monday and have a higher tomato rating "
        "than IMDB rating? Keep in mind that IMDB goes from 1-10 and tomatoes "
        "only from 1-5, so you need to normalise the ratings to do a fair comparison."
    ),
    "What movies should I watch to learn more about Japanse culture?",
    (
        "How many movies were released in each decade? Write decade as a string, e.g. "
        "'1920-1929'. Sort ascending by decade."
    ),
    (
        "Find movies that are suitable to watch with my kids, both by genre and their "
        "parental guidance rating. Just recommend good movies."
    ),
]

BASE_CHAT_MESSAGES = [
    {
        "role": "system",
        "content": "You are an expert English to MongoDB aggregation pipeline translation system."
        "You will accept an example document from a collection and an English question, and return an aggregation "
        "pipeline that can answer the question. Do not explain the query or add any additional comments, only "
        "return a single code block with the aggregation pipeline without the aggregate command.",
    }
]

MODEL_NAME = "gpt-3.5-turbo"


@st.cache
def ask_model(doc, question):
    """This is the call to the OpenAI API. It creates a prompt from the document
    and question and returns the endpoint's response."""

    messages = BASE_CHAT_MESSAGES + [
        {
            "role": "user",
            "content": f"Example document: {doc.strip()}\n\nQuestion: {question.strip()}\n\n",
        }
    ]

    return openai.ChatCompletion.create(
        model=MODEL_NAME,
        messages=messages,
        temperature=0,
        max_tokens=1000,
        top_p=1.0,
    )


def extract_pipeline(response):
    content = response["choices"][0]["message"]["content"].strip("\n `")
    return content


st.set_page_config(layout="wide")

# initialise session state
if not "response" in st.session_state:
    st.session_state.response = None
if not "_id" in st.session_state:
    st.session_state._id = None
if not "feedback" in st.session_state:
    st.session_state.feedback = False
if not "default_question" in st.session_state:
    st.session_state.default_question = random.choice(MOVIES_EXAMPLE_QUESTIONS)

# DB access


st.markdown(
    """# English to MQL Demo

This demo app uses OpenAI's GPT-4 (gpt-4) model to generate a MongoDB
aggregation pipeline from an English question and example document.

🚧 The app is experimental and may return incorrect results. Do not enter any sensitive information! 🚧
"""
)


# two-column layout
col_left, col_right = st.columns(2, gap="large")

with col_left:
    st.markdown("### Example Document and Question")
    # wrap textareas in form
    with st.form("text_inputs"):
        doc = st.text_area(
            "Enter example document from collection, e.g. db.collection.findOne()",
            value=MOVIES_EXAMPLE_DOC,
            height=300,
        )

        # question textarea
        question = st.text_area(
            label="Ask question in English",
            value=st.session_state.default_question,
        )

        # submit button
        submitted = st.form_submit_button("Translate", type="primary")
        if submitted:
            st.session_state._id = None
            st.session_state.feedback = False
            st.session_state.response = ask_model(doc, question)


with col_right:
    st.markdown("### Generated MQL")

    # show response
    response = st.session_state.response
    if response:
        pipeline = extract_pipeline(response)
        # print result as code block
        st.code(
            pipeline,
            language="javascript",
        )

        # feedback form
        with st.empty():
            if st.session_state.feedback:
                st.write("✅ Thank you for your feedback.")

            elif st.session_state._id:
                with st.form("feedback_inputs"):
                    radio = st.radio("Is the result correct?", ("Yes", "No"))
                    feedback = st.text_area(
                        "If not, please tell us what the issue is:",
                    )

                    # submit button
                    feedback_submit = st.form_submit_button(
                        "Submit Feedback", type="secondary"
                    )
                    if feedback_submit:
                        st.session_state.feedback = {
                            "correct": radio == "Yes",
                            "comment": feedback,
                        }

            else:
                doc = {
                    "ts": datetime.now(),
                    "doc": doc,
                    "question": question,
                    "generated_mql": pipeline,
                    "response": response,
                    "version": 3,
                    "model": MODEL_NAME,
                }