appStore/rag.py

import os
import numpy as np
import pandas as pd
import openai
from haystack.schema import Document
import streamlit as st
from tenacity import retry, stop_after_attempt, wait_random_exponential
from huggingface_hub import InferenceClient


# Get openai API key
hf_token = os.environ["HF_API_KEY"]

# define a special function for putting the prompt together (as we can't use haystack)
def get_prompt(context, label):
  base_prompt="Summarize the following context efficiently in bullet points, the less the better - but keep concrete goals. \
  Summarize only elements of the context that address vulnerability of "+label+" to climate change. \
  If there is no mention of "+label+" in the context, return nothing. \
  Do not include an introduction sentence, just the bullet points as per below. \
  Formatting example: \
    - Bullet point 1 \
    - Bullet point 2 \
  "

  prompt = base_prompt+"; Context: "+context+"; Answer:"

  return prompt


# # exception handling for issuing multiple API calls to openai (exponential backoff)
# @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6))
# def completion_with_backoff(**kwargs):
#     return openai.ChatCompletion.create(**kwargs)

# construct query, send to HF API and process response
def run_query(context, label):
    '''
    For non-streamed completion, enable the following 2 lines and comment out the code below
    '''
    chatbot_role = """You are an analyst specializing in climate change impact assessments and producing insights from policy documents."""
    messages = [{"role": "system", "content": chatbot_role},{"role": "user", "content": get_prompt(context, label)}]
    
    # Initialize the client, pointing it to one of the available models
    client = InferenceClient("meta-llama/Meta-Llama-3.1-405B-Instruct", token = hf_token)

    chat_completion = client.chat.completions.create(
        messages=messages,
        stream=True
    )

    # iterate through the streamed output
    report = []
    res_box = st.empty()
    for chunk in chat_completion:
        # extract the object containing the text (totally different structure when streaming)
        chunk_message = chunk.choices[0].delta
        # test to make sure there is text in the object (some don't have)
        if 'content' in chunk_message:
            report.append(chunk_message['content']) # extract the message
            # add the latest text and merge it with all previous
            result = "".join(report).strip()
            # res_box.success(result) # output to response text box
            res_box.success(result)