app.py

import streamlit as st
import os
import pkg_resources

# Using this wacky hack to get around the massively ridicolous managed env loading order
def is_installed(package_name, version):
    try:
        pkg = pkg_resources.get_distribution(package_name)
        return pkg.version == version
    except pkg_resources.DistributionNotFound:
        return False

@st.cache_resource 
def install_packages():
    install_commands = []

    if not is_installed("spaces", "0.12.0"):
        install_commands.append("pip install spaces==0.12.0")
    
    if not is_installed("pydantic", "1.8.2"):
        install_commands.append("pip install pydantic==1.8.2")

    if install_commands:
        os.system(" && ".join(install_commands))

# install packages if necessary
install_packages()


import re
import json
from dotenv import load_dotenv
import numpy as np
import pandas as pd
from haystack.schema import Document
from haystack.document_stores import PineconeDocumentStore
from haystack.nodes import EmbeddingRetriever
import openai

# for local st testing, may need to run source ~/.zshrc to point to env vars

# Get openai API key
openai.api_key = os.environ["OPENAI_API_KEY"]

# Get openai API key
pinecone_key = os.environ["PINECONE_API_KEY"]

#___________________________________________________________________________________________________________

# @st.cache_resource 
# def get_document_store():
#     doc_file_name="cpv_full_southern_africa"
#     document_store = PineconeDocumentStore(api_key=pinecone_key,
#                                        environment="asia-southeast1-gcp-free",
#                                        index=doc_file_name)
#     return document_store


# # Get (or initialize and get) the document store
# document_store = get_document_store()


@st.cache_resource
def get_retriever():
    doc_file_name="cpv_full_southern_africa"
    document_store = PineconeDocumentStore(api_key=pinecone_key,
                                       environment="asia-southeast1-gcp-free",
                                       index=doc_file_name)
    retriever = EmbeddingRetriever(
        document_store=document_store,
        embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
        model_format="sentence_transformers",
        progress_bar=False,
    )
    return retriever

retriever = get_retriever()


# # Instantiate retriever
# retriever = EmbeddingRetriever(
#   document_store=document_store,
#   embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
#   model_format="sentence_transformers",
#   progress_bar=False,
# )

prompt_template="Answer the given question using the following documents. \
Formulate your answer in the style of an academic report. \
Provide example quotes and citations using extracted text from the documents. \
Use facts and numbers from the documents in your answer. \
ALWAYS include references for information used from documents at the end of each applicable sentence using the format: '[ref. #]', where '[ref. #]' is included in the text provided at the start of each document (demarcated by the pattern '- &&& [ref. #] document_name &&&:')'. \
Do not include page numbers in the references. \
If no relevant information to answer the question is present in the documents, just say you don't have enough information to answer."


# Create a list of options for the dropdown
# model_options = ['chatGPT','Llama2']

# Create a list of options for the dropdown
country_options = ['All Countries','Angola','Botswana','Lesotho','Kenya','Malawi','Mozambique','Namibia','Rwanda','South Africa','Zambia','Zimbabwe']

# Create a list of options for the dropdown
vulnerability_options = ['All Categories','Agricultural communities', 'Children', 'Coastal communities', 'Ethnic, racial or other minorities', 'Fishery communities', 'Informal sector workers', 'Members of indigenous and local communities', 'Migrants and displaced persons', 'Older persons', 'Persons living in poverty', 'Persons with disabilities', 'Persons with pre-existing health conditions', 'Residents of drought-prone regions', 'Rural populations', 'Sexual minorities (LGBTQI+)', 'Urban populations', 'Women and other genders','Other']

# List of examples
examples = [
    "-",
    "What specific initiatives are presented in the context to address the needs of groups such as women and children to the effects climate change?",
    "In addition to gender, children, and youth, is there any mention of other groups facing disproportional impacts from climate change due to their geographic location, socio-economic status, age, gender, health, and occupation?"
]


def get_docs(input_query, country = [], vulnerability_cat = []):
  if not country:
    country = "All Countries"
  if not vulnerability_cat:
    if country == "All Countries":
      filters = None
    else:
      filters = {'country': {'$in': country}}
  else:
    if country == "All Countries":
          filters = {'vulnerability_cat': {'$in': vulnerability_cat}}
    else:
      filters = {'country': {'$in': country},'vulnerability_cat': {'$in': vulnerability_cat}}
  docs = retriever.retrieve(query=input_query, filters = filters, top_k = 10)
  # Break out the key fields and convert to pandas for filtering
  docs = [{**x.meta,"score":x.score,"content":x.content} for x in docs]
  df_docs = pd.DataFrame(docs)
  # Get ourselves an index setup from which to base the source reference number from (in the prompt and matching afterwards)
  df_docs = df_docs.reset_index()
  df_docs['ref_id'] = df_docs.index + 1 # start the index at 1
  # Convert back to Document format
  ls_dict = []
  # Iterate over df and add relevant fields to the dict object
  for index, row in df_docs.iterrows():
      # Create a Document object for each row
      doc = Document(
          row['content'],
          meta={'country': row['country'],'document': row['document'], 'page': row['page'], 'file_name': row['file_name'], 'ref_id': row['ref_id'], 'vulnerability_cat': row['vulnerability_cat'], 'score': row['score']}
      )

      # Append the Document object to the documents list
      ls_dict.append(doc)
  return ls_dict

def get_refs(docs, res):
  '''
  Parse response for engineered reference ids (refer to prompt template)
  Extract documents using reference ids
  '''
  res = res.lower()  # Convert to lowercase for matching
  # This pattern should be returned by gpt3.5
  # pattern = r'ref\. (\d+)\]\.'
  pattern = r'ref\. (\d+)'
  ref_ids = [int(match) for match in re.findall(pattern, res)]
  # extract
  result_str = ""  # Initialize an empty string to store the result
  for i in range(len(docs)):
      doc = docs[i].to_dict()
      ref_id = doc['meta']['ref_id']
      if ref_id in ref_ids:
        if doc['meta']['document'] == "Supplementary":
            result_str += "**Ref. " + str(ref_id) + " [" + doc['meta']['country'] + " " + doc['meta']['document'] + ':' + doc['meta']['file_name'] + ' p' + str(doc['meta']['page']) + '; vulnerabilities: ' + doc['meta']['vulnerability_cat'] + "]:** " + "*'" + doc['content'] + "'*<br> <br>" # Add <br> for a line break
        else:
            result_str += "**Ref. " + str(ref_id) + " [" + doc['meta']['country'] + " " + doc['meta']['document'] + ' p' + str(doc['meta']['page']) + '; vulnerabilities: ' + doc['meta']['vulnerability_cat'] + "]:** " + "*'" + doc['content'] + "'*<br> <br>" # Add <br> for a line break
  
  return result_str

# define a special function for putting the prompt together (as we can't use haystack)
def get_prompt(docs, input_query):
  base_prompt=prompt_template
  # Add the meta data for references
  context = ' - '.join(['&&& [ref. '+str(d.meta['ref_id'])+'] '+d.meta['document']+' &&&: '+d.content for d in docs])
  prompt = base_prompt+"; Context: "+context+"; Question: "+input_query+"; Answer:"
  return(prompt)

def run_query(input_text, country, model_sel):
    # first call the retriever function using selected filters
    docs = get_docs(input_text, country=country,vulnerability_cat=vulnerabilities_cat)
    # model selector (not currently being used)
    if model_sel == "chatGPT":
        # instantiate ChatCompletion as a generator object (stream is set to True)
        response = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": get_prompt(docs, input_text)}], stream=True)
        # iterate through the streamed output
        report = []
        for chunk in response:
            # 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

        references = get_refs(docs, result) # extract references from the generated text
    # Llama2 selection (was running on HF)
    # else:
    #     res = client.text_generation(get_prompt(docs, query=input_query), max_new_tokens=4000, temperature=0.01, model=model)
    #     output = res
    #     references = get_refs(docs, res)
    
    st.markdown("----")
    st.markdown('**REFERENCES:**')
    st.markdown('References are based on text automatically extracted from climate policy documents. These extracts may contain non-legible characters or disjointed text as an artifact of the extraction procedure')
    st.markdown(references, unsafe_allow_html=True)


#___________________________________________________________________________________________________________

# Sidebar (filters)
with st.sidebar:
    country = st.sidebar.multiselect('Filter by country:', country_options)
    vulnerabilities_cat = st.sidebar.multiselect('Filter by vulnerabilities category:', vulnerability_options)
    with st.expander("ℹ️ - About filters", expanded=False):
        st.markdown(
            """
            * *These selections will filter the data matched against your query*
            * *For a comparative analysis of multiple countries or vulnerability categories, select the items you require or select **'All Countries'** or **'All Categories'***
            * *Be careful in using the vulnerabilities category filter, as many of the categories are not well represented in the documents. Therefore, this will severly limit the data available for analysis*
            """
        )

# Main window title
with st.container():
        st.markdown("<h2 style='text-align: center;'> Climate Policy Documents: Vulnerabilities Analysis Q&A </h2>", unsafe_allow_html=True)
        st.write(' ')

# Main window instructions
with st.expander("ℹ️ - About this app", expanded=False):
    st.write(
        """
        This tool seeks to provide an interface for quering national climate policy documents (NDCs, LTS etc.). The current version is powered by chatGPT (3.5). The document store is limited to 10 Southern African countries (Angola, Botswana, Eswatini, Lesotho, Malawi, Mozambique, Namibia, South Africa, Zambia, Zimbabwe), as well as Kenya and Rwanda. The intended use case is to allow users to interact with the documents and obtain valuable insights on various vulnerable groups affected by climate change.
        
        **DISCLAIMER:** *This prototype tool based on LLMs (Language Models) is provided "as is" for experimental and exploratory purposes only, and should not be used for critical or production applications. Users are advised that the tool may contain errors, bugs, or limitations and should be used with caution and awareness of potential risks, and the developers make no warranties or guarantees regarding its performance, reliability, or suitability for any specific purpose.*
        """)
    # Display the text passages as radio buttons
    selected_example = st.radio("Example questions", examples)
    st.write(
    """
    You can request comparative analyses between countries by filtering by country, and using more advanced prompts. For example:
    
    *Provide a comparative analysis between Angola and Kenya with regard to specific initiatives presented in the context to address the needs of groups such women and children to the effects climate change.*

    Make sure your filters match the countries you have specified for the analysis!
    """)


# Dropdown selectbox: model (currently not used)
# model_sel = st.selectbox('Select an LLM:', model_options)
model_sel = "chatGPT"

#----Model Select logic-------
if model_sel == "chatGPT":
    model_name = "gpt-3.5-turbo"

# else:
#     model = "meta-llama/Llama-2-70b-chat-hf"
#     # Instantiate the inference client
#     client = InferenceClient()

# get prompt from user or example prompt
if selected_example == "-": #hyphen used as a work around (st won't allow null selection)
    text = st.text_area('Enter your question in the text box below using natural language or select an example from above:')
else:
    text = st.text_area('Enter your question in the text box below using natural language or select an example from above:', value=selected_example)

if st.button('Submit'):
    st.markdown("----")
    st.markdown('**RESPONSE:**')
    res_box = st.empty()
    run_query(text, country=country, model_sel=model_sel)