app.py

import streamlit as st
import pandas as pd
from langchain_text_splitters import TokenTextSplitter
from langchain.docstore.document import Document
from torch import cuda
from langchain_community.embeddings import HuggingFaceEmbeddings, HuggingFaceInferenceAPIEmbeddings
device = 'cuda' if cuda.is_available() else 'cpu'


st.set_page_config(page_title="SEARCH IATI",layout='wide')
st.title("SEARCH IATI Database")
var=st.text_input("enter keyword")
title = var.replace(' ','+')

def create_chunks(text):
    text_splitter = TokenTextSplitter(chunk_size=500, chunk_overlap=0)
    texts = text_splitter.split_text(text)
    return texts 

def get_chunks():
    orgas_df = pd.read_csv("iati_files/project_orgas.csv")
    region_df = pd.read_csv("iati_files/project_region.csv")
    sector_df = pd.read_csv("iati_files/project_sector.csv")
    status_df = pd.read_csv("iati_files/project_status.csv")
    texts_df = pd.read_csv("iati_files/project_texts.csv")

    projects_df = pd.merge(orgas_df, region_df, on='iati_id', how='inner')
    projects_df = pd.merge(projects_df, sector_df, on='iati_id', how='inner')
    projects_df = pd.merge(projects_df, status_df, on='iati_id', how='inner')
    projects_df = pd.merge(projects_df, texts_df, on='iati_id', how='inner')
    giz_df = projects_df[projects_df.client.str.contains('bmz')].reset_index(drop=True)

    giz_df.drop(columns= ['orga_abbreviation', 'client',
        'orga_full_name', 'country', 
        'country_flag', 'crs_5_code', 'crs_3_code',
        'sgd_pred_code'], inplace=True)

    giz_df['text_size'] = giz_df.apply(lambda x: len((x['title_main'] + x['description_main']).split()), axis=1)
    giz_df['chunks'] = giz_df.apply(lambda x:create_chunks(x['title_main'] + x['description_main']),axis=1)
    giz_df = giz_df.explode(column=['chunks'], ignore_index=True)

        
    placeholder= []
    for i in range(len(giz_df)):
        placeholder.append(Document(page_content= giz_df.loc[i,'chunks'], 
                                metadata={"iati_id": giz_df.loc[i,'iati_id'],
                                        "iati_orga_id":giz_df.loc[i,'iati_orga_id'],
                                        "country_name":str(giz_df.loc[i,'country_name']),
                                        "crs_5_name": giz_df.loc[i,'crs_5_name'],
                                        "crs_3_name": giz_df.loc[i,'crs_3_name'],
                                        "sgd_pred_str":giz_df.loc[i,'sgd_pred_str'],
                                        "status":giz_df.loc[i,'status'],
                                        "title_main":giz_df.loc[i,'title_main'],}))
    return placeholder

def embed_chunks(chunks):
    embeddings = HuggingFaceEmbeddings(
        model_kwargs = {'device': device},
        encode_kwargs = {'normalize_embeddings': True},
        model_name='BAAI/bge-m3'
    )
    # placeholder for collection
    qdrant_collections = {}
    qdrant_collections['all'] = Qdrant.from_documents(
                chunks,
                embeddings,
                path="/data/local_qdrant",
                collection_name='all',
            )
            
    print(qdrant_collections)
    print("vector embeddings done")
    return qdrant_collections
    
chunks = get_chunks()
qdrant_col = embed_chunks(chunks)

button=st.button("search")

if button :
    st.write(chunks[0])