app.py

import streamlit as st
import os
import re
import json
from dotenv import load_dotenv
from haystack.nodes.prompt import PromptNode, PromptTemplate
from haystack.nodes import EmbeddingRetriever
from haystack import Pipeline
import numpy as np
import pandas as pd
from haystack.document_stores import FAISSDocumentStore, PineconeDocumentStore
from haystack.nodes import EmbeddingRetriever
from haystack.schema import Document
from huggingface_hub import login, HfApi, hf_hub_download, InferenceClient
import openai

# Get HF token
hf_token = os.environ["HF_TOKEN"]
login(token=hf_token, add_to_git_credential=True)

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

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


# template = PromptTemplate(
#     prompt="""
# 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. \
# Reference information used from documents at the end of each applicable sentence (ex: [source: document_name]), where 'document_name' is the text provided at the start of each document (demarcated by '- &&&' and '&&&:')'. \
# If no relevant information to answer the question is present in the documents, just say you don't have enough information to answer. \
# Context: {' - '.join(['&&& '+d.meta['document']+' ref. '+str(d.meta['ref_id'])+' &&&: '+d.content for d in documents])}; Question: {query}; Answer:""",
# )

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 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 = None):
#     '''
#     Construct a hacky query to focus the retriever on the target country (see notes below)
#     We take the top 150 k because we want to make sure we have 10 pertaining to the selected country
#     '''
#     if country == 'All Countries':
#         query = input_query
#     else:
#         query = "For the country of "+country+", "+input_query
#     # Retrieve top k documents
#     docs = retriever.retrieve(query=query,top_k = 150)
#     # 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)
#     if country != 'All Countries':
#       df_docs = df_docs.query('country in @country')
#     # Take the top 10
#     df_docs = df_docs.head(10)
#     # 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'], 'score': row['score']}
#         )

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

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']) + "]:** " + "*'" + 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']) + "]:** " + "*'" + 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):
    # docs = get_docs(input_text, country)
    docs = get_docs(input_text, country=country,vulnerability_cat=vulnerability_options)
    # st.write('Selected country: ', country) # Debugging country
    if model_sel == "chatGPT":
        # res = pipe.run(query=input_text, documents=docs)
        res = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": get_prompt(docs, input_text)}])
        output = res.choices[0].message.content
        references = get_refs(docs, output)
    # else:
    #     res = client.text_generation(get_prompt_llama2(docs, query=input_query), max_new_tokens=4000, temperature=0.01, model=model)
    #     output = res
    #     references = get_refs(docs, res)
    st.write('Response')
    st.success(output)
    st.write('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)

# # Setup retriever, pulling from local faiss datastore
# retriever = EmbeddingRetriever(
#     document_store=FAISSDocumentStore.load(
#         index_path="./cpv_full_southern_africa_kenya.faiss",
#         config_path="./cpv_full_southern_africa_kenya.json",
#     ),
#     embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
#     model_format="sentence_transformers",
#     progress_bar=False,
# )

# Setup retriever, pulling from pinecone
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"
)


with st.sidebar:
    # Dropdown selectbox
    country = st.sidebar.multiselect('Select a country:', country_options)
    st.markdown(
        """
        * *For a comparative analysis of multiple countries, select **'All Countries'***
        * *Then be sure to mention the country names of interest in your query*
        """
    )
    vulnerabilities_cat = st.sidebar.multiselect('Select a vulnerabilities category:', vulnerability_options)
    # choice = st.sidebar.radio(label = 'Select the Document',
    #                         help = 'You can upload the document \
    #                         or else you can try a example document', 
    #                         options = ('Upload Document', 'Try Example'), 
    #                         horizontal = True)

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

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)




# Dropdown selectbox: model
# 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"
    # # Initialize the PromptNode
    # pn = PromptNode(model_name_or_path=model_name, default_prompt_template=template, api_key=openai_key, max_length=2000, model_kwargs={"generation_kwargs": {"do_sample": False, "temperature": 0}})
    
    # # Initialize the pipeline
    # pipe = Pipeline()
    # pipe.add_node(component=pn, name="prompt_node", inputs=["Query"])
else:
    # Currently disabled
    model = "meta-llama/Llama-2-70b-chat-hf"
    # Instantiate the inference client
    client = InferenceClient()




if selected_example == "-":
    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'):
    run_query(input_text = text, country=country, model_sel=model_sel)