app.py

# inspiration from Ekimetrics climate qa

import streamlit as st
import os
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
from haystack.nodes import EmbeddingRetriever
from haystack.schema import Document


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

# Select model
model_name = "gpt-3.5-turbo"

# Define the template
template = PromptTemplate(
    prompt="""
Answer the given question using the provided documents. The answer should be in the style of an academic report and should provide example quotes and references. Always start the response by stating the name of the country relevant to the documents. When relevant, use bullet points and lists to structure your answers. When relevant, use facts and numbers from the following documents in your answer. Whenever you use information from a document, reference it at the end of the sentence (ex: [doc 2]). You don't have to use all documents, only if it makes sense in the conversation. 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(documents)}\n\nQuestion: {query}\n\nAnswer:""",
)

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

# 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)
    if country:
        query = "For the country of "+country+", "+input_query
    else:
        query = input_query
    # Get top 150 because we want to make sure we have 10 pertaining to the selected country 
    # TEMP SOLUTION: not ideal, but FAISS document store doesnt allow metadata filtering. Needs to be tested with the full dataset
    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:
      df_docs = df_docs.query('country in @country')
    # Take the top 10
    df_docs = df_docs.head(10)
    # Convert back to Document format
    ls_dict = []
    for doc, _ in df_docs.iterrows():
       x = Document(df_docs['content'][doc])
       ls_dict.append(x)
    return(ls_dict)

def run_query(input_text):
    docs = get_docs(input_text)
    res = pipe.run(query=input_text, documents=docs)
    output = res["results"][0]
    st.write('Response')
    st.success(output)


# Setup retriever, pulling from local faiss datastore
retriever = EmbeddingRetriever(
    document_store=FAISSDocumentStore.load(
        index_path="./cpv_test_2.faiss",
        config_path="./cpv_test_2.json",
    ),
    embedding_model="sentence-transformers/multi-qa-mpnet-base-dot-v1",
    model_format="sentence_transformers",
    progress_bar=False,
)

# Initialize the PromptNode
pn = PromptNode(model_name_or_path=model_name, default_prompt_template=template, api_key=openai_key, max_length=700)

# Initialize the pipeline
pipe = Pipeline()
pipe.add_node(component=pn, name="prompt_node", inputs=["Query"])


# Guiding text             
st.title('Climate Policy Documents: Vulnerabilities Analysis Q&A (test)')
st.markdown('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) and limited to 9 Southern African countries (Angola, Botswana, Eswatini, Lesotho, Malawi, Mozambique, Namibia, South Africa, Zambia, Zimbabwe). 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.')

    

# Dropdown selectbox
country = st.selectbox('Select a country:', country_options) 

# Display the text passages as radio buttons
selected_example = st.radio("Example questions", examples)

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(text)