app.py

import streamlit as st
from annotated_text import annotated_text, annotation
import fitz
import os
import chromadb
import uuid
from pathlib import Path
import os
os.environ['OPENAI_API_KEY'] = os.environ['OPEN_API_KEY']
st.title("Contracts Multiple File Search ")
import pandas as pd
import time

from langchain.retrievers import BM25Retriever, EnsembleRetriever
from langchain.schema import Document
from langchain.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
embedding = HuggingFaceEmbeddings(model_name='BAAI/bge-base-en-v1.5')
from FlagEmbedding import FlagReranker
reranker = FlagReranker('BAAI/bge-reranker-base')
import spacy
# Load the English model from SpaCy
from spacy_download import load_spacy
nlp = load_spacy("en_core_web_sm")

def util_upload_file_and_return_list_docs(uploaded_files):
    #util_del_cwd()
    list_docs = []
    list_save_path = []
    for uploaded_file in uploaded_files:
        save_path = Path(os.getcwd(), uploaded_file.name)
        with open(save_path, mode='wb') as w:
            w.write(uploaded_file.getvalue())
        #print('save_path:', save_path)
        docs = fitz.open(save_path) 
        list_docs.append(docs)
        list_save_path.append(save_path)
    return(list_docs, list_save_path)
#### Helper Functions to Split using Rolling Window (recomm : use smaller rolling window )
def split_txt_file_synthetic_sentence_rolling(ctxt, sentence_size_in_chars, sliding_size_in_chars,debug=False):
    sliding_size_in_chars = sentence_size_in_chars - sliding_size_in_chars
    pos_start = 0
    pos_end = len(ctxt)
    final_return = []
    if(debug):
        print('pos_start : ',pos_start)
        print('pos_end : ',pos_end)
    if(pos_end<sentence_size_in_chars):
        return([{'section_org_text':ctxt[pos_start:pos_end],'section_char_start':pos_start,'section_char_end':pos_end}])
    if(sentence_size_in_chars<sliding_size_in_chars):
        return(None)
    stop_condition = False
    start = pos_start
    end = start + sentence_size_in_chars
    mydict = {}
    mydict['section_org_text'] = ctxt[start:end]
    mydict['section_char_start'] = start
    mydict['section_char_end'] = end
    final_return.append(mydict)
    #### First Time ENDS
    while(stop_condition==False):
        start = end - sliding_size_in_chars
        end = start + sentence_size_in_chars
        if(end>pos_end):
            if(start<pos_end):
                end = pos_end
                mydict = {}
                mydict['section_org_text'] = ctxt[start:end]
                mydict['section_char_start'] = start
                mydict['section_char_end'] = end
                final_return.append(mydict)
                stop_condition=True
            else:
                stop_condition=True
        else:
            mydict = {}
            mydict['section_org_text'] = ctxt[start:end]
            mydict['section_char_start'] = start
            mydict['section_char_end'] = end
            final_return.append(mydict)
        if(debug):
            print('start : ', start)
            print('end : ', end)
    return(final_return)
### helper to make string out of iw_status
# def util_get_list_page_and_passage(docs):
#     page_documents = []
#     passage_documents = []
#     for txt_index, txt_page in enumerate(docs):
#         page_document = txt_page.get_text()##.encode("utf8") # get plain text (is in UTF-8)
#         page_documents.append(page_document)
#         sections = split_txt_file_synthetic_sentence_rolling(page_document,700,200) 
#         for sub_sub_index, sub_sub_item in enumerate(sections):
#             sub_text=sub_sub_item['section_org_text']
#             passage_document = Document(page_content=sub_text, metadata={"page_index": txt_index}) 
#             passage_documents.append(passage_document) 
#     return(page_documents,passage_documents)

def split_into_sentences_with_offsets(text):
    """
    Splits a paragraph into sentences and returns them along with their start and end offsets.
    :param text: The input text to be split into sentences.
    :return: A list of tuples, each containing a sentence and its start and end offsets.
    """
    doc = nlp(text)
    return [(sent.text, sent.start_char, sent.end_char) for sent in doc.sents]
    
def util_get_list_page_and_passage(list_docs, list_save_path):
    #page_documents = []
    passage_documents = []
    for ind_doc, docs in enumerate(list_docs):
        for txt_index, txt_page in enumerate(docs):
            page_document = txt_page.get_text()##.encode("utf8") # get plain text (is in UTF-8)
            #page_documents.append(page_document)
            sections = split_into_sentences_with_offsets(page_document) 
            for sub_sub_index, sub_sub_item in enumerate(sections):
                sub_text=sub_sub_item[0]
                passage_document = Document(page_content=sub_text, metadata={"page_content": page_document,"page_index": txt_index, "file_name" : str(list_save_path[ind_doc])}) 
                passage_documents.append(passage_document) 
    return(passage_documents)
    
# def util_index_chromadb_passages():
#     ##### PROCESSING
#     # create client and a new collection
#     collection_name = str(uuid.uuid4().hex)
#     chroma_client = chromadb.EphemeralClient()
#     chroma_collection = chroma_client.get_or_create_collection(collection_name)
#     # define embedding function
#     embed_model = LangchainEmbedding(HuggingFaceEmbeddings(model_name="BAAI/bge-small-en"))
#     vector_store = ChromaVectorStore(chroma_collection=chroma_collection)   
#     return(chroma_client,chroma_collection,collection_name,vector_store,embed_model)

def util_get_only_content_inside_loop(page_no,page_documents):
    for index, item in enumerate(page_documents):
        if(page_documents[index].metadata['txt_page_index']==page_no):
            return(page_documents[index].get_content())  
    return(None)
# def util_get_list_pageno_and_contents(page_documents,passage_documents,passage_nodes):
#     ''' page no starts with index 1 '''
#     return_value = []
#     for index, item in enumerate(passage_nodes):
#         page_no = passage_nodes[index].metadata['txt_page_index']
#         page_content = util_get_only_content_inside_loop(page_no,page_documents)
#         return_value.append((page_no+1,page_content))
#     return(return_value)

def util_get_list_pageno_and_contents(some_query_passage,passage_documents,passage_nodes):
    ''' page no starts with index 1 '''

    return_value = []
    rescore = reranker.compute_score([[some_query_passage , x.page_content] for x in passage_nodes])
    print('rescore :: ',rescore)
    tmp_array = []
    for i, x in enumerate(passage_nodes):
        tmp_dict = {"passage_content":x.page_content, 
                    "page_no":int(x.metadata['page_index'])+1, 
                    "page_content":str(x.metadata['page_content']), 
                    "file_name": str(x.metadata['file_name']), 
                    "score" : float(rescore[i])}
        tmp_array.append(tmp_dict)
    df = pd.DataFrame(tmp_array)
    df = df.sort_values(by='score', ascending=False)
    df = df.drop_duplicates(subset=['file_name'], keep='first')
    df = df[["passage_content","file_name","page_no","page_content"]]
    return(df)
    
# # def util_openai_extract_entity(example_passage, example_entity, page_content):
# #     import openai
# #     openai.api_key = os.environ['OPENAI_API_KEY']
    
# #     content = """Find the Entity based on Text . Return empty string if Entity does not exists. Here is one example below
# #     Text: """ + example_passage + """
# #     Entity: """ + example_entity + """
    
# #     Text: """ + page_content + """
# #     Entity: """
    
# #     return_value = openai.ChatCompletion.create(model="gpt-4",temperature=0.0001,messages=[{"role": "user", "content": content},])
# #     return(str(return_value['choices'][0]['message']['content']))
def util_openai_extract_clause(example_prompt, page_content):
    import openai
    openai.api_key = os.environ['OPENAI_API_KEY']
    content = example_prompt
    content = content + "\n Answer precisely; do not add anything extra, and try to locate the answer in the below context   \n context: "
    return_value = openai.ChatCompletion.create(model="gpt-4o-mini",temperature=0.0001,messages=[{"role": "user", "content": content + "\n" + page_content},])
    return(str(return_value['choices'][0]['message']['content']))


def util_openai_hyde(example_prompt):
    import openai
    openai.api_key = os.environ['OPENAI_API_KEY']
    content = example_prompt
    return_value = openai.ChatCompletion.create(model="gpt-4o-mini",temperature=0.0001,messages=[
        {"role": "system", "content": "You are a legal contract lawyer. generate a summary from below text " + "\n"},
        {"role": "user", "content": example_prompt + "\n"},
    
    ]
                                               )
    return(str(return_value['choices'][0]['message']['content']))

    
def util_openai_format (example_passage, page_content):
    '''
    annotated_text("  ",annotation("ENTITY : ", str(page_no)),)
    '''
    if(True):
        found_value = util_openai_extract_clause(example_passage, page_content)
        if(len(found_value)>0):
            found_value = found_value.strip()
            first_index = page_content.find(found_value)
            if(first_index!=-1):
                print('first_index : ',first_index)
                print('found_value : ',found_value)
                return(annotated_text(page_content[0:first_index-1],annotation(found_value, " FOUND ENTITY "),page_content[first_index+len(found_value):]))        
    return(annotated_text(page_content))
def util_openai_modify_prompt(example_prompt, page_content):
    import openai
    openai.api_key = os.environ['OPENAI_API_KEY']
    my_prompt = """Expand the original Query to show exact resuls for extraction\n
    Query: """ + example_prompt # + """\nDocument: """ + page_content + """ """
    return_value = openai.ChatCompletion.create(model="gpt-4o-mini",temperature=0.0001,messages=[{"role": "user", "content": my_prompt},])
    return(str(return_value['choices'][0]['message']['content']))    

# def create_bm25_page_rank(page_list_retrieve, page_query):
#     """ page_corpus : array of page text , page_query is user query """
#     from operator import itemgetter
#     from rank_bm25 import BM25Okapi
#     tokenized_corpus = [doc.split(" ") for x, doc in page_list_retrieve]
#     tokenized_query = page_query.split(" ")
#     bm25 = BM25Okapi(tokenized_corpus)
#     doc_scores = bm25.get_scores(tokenized_query).tolist()
#     tmp_list = []
#     for index, item in enumerate(page_list_retrieve):
#         tmp_list.append((item[0], item[1],doc_scores[index]))
#     tmp_list = sorted(tmp_list, key=itemgetter(2), reverse=True)
#     return(tmp_list)
    

passage_documents = []

with st.form("my_form"):
    multi = '''1. Download and Upload Multiple contracts
    
    e.g. https://www.barc.gov.in/tenders/GCC-LPS.pdf
    
    e.g. https://www.montrosecounty.net/DocumentCenter/View/823/Sample-Construction-Contract
    '''
    st.markdown(multi)
    multi = '''2. Insert Query to search or find similar language '''
    st.markdown(multi)
    multi = '''3. Press Index.'''
    st.markdown(multi)
    multi = '''
    ** Attempt is made for appropriate page and passage retrieval ** \n
    '''
    st.markdown(multi)
    #uploaded_file = st.file_uploader("Choose a file")  

    list_docs = []
    list_save_path = []
    uploaded_files = st.file_uploader("Choose file(s)", accept_multiple_files=True)
    print('uploaded_files ', uploaded_files)
    single_example_passage = st.text_area('Enter Query Here',"What is Governing Law ")
    submitted = st.form_submit_button("Index and Calculate")
    
    if submitted and (uploaded_files is not None):
        list_docs, list_save_path = util_upload_file_and_return_list_docs(uploaded_files)
        # print('list_docs ' ,list_docs)
        # print('list_save_path ' , list_save_path)
        passage_documents = util_get_list_page_and_passage(list_docs, list_save_path)
        bm25_retriever = BM25Retriever.from_documents(passage_documents)
        bm25_retriever.k = 2
        chroma_vectorstore = Chroma.from_documents(passage_documents, embedding)
        chroma_retriever = chroma_vectorstore.as_retriever(search_kwargs={"k": 2})
        ensemble_retriever = EnsembleRetriever(retrievers=[bm25_retriever, chroma_retriever],weights=[0.25, 0.75])
        passage_nodes = ensemble_retriever.get_relevant_documents(single_example_passage)
        print('len(passage_nodes):', len(passage_nodes))
        time.sleep(5)
        df = util_get_list_pageno_and_contents(single_example_passage,passage_documents,passage_nodes)
        st.write(df)
        # print('len(page_list_retrieve):', len(page_list_retrieve))
        # if(len(page_list_retrieve)>0):
        #     page_list_retrieve = list(set(page_list_retrieve))
        #     for iindex in page_list_retrieve:
        #         page_no = iindex[0]
        #         page_content = iindex[1]
        #         annotated_text("  ",annotation("RELEVANT PAGENO : ", str(page_no), font_family="Comic Sans MS", border="2px dashed red"),)
        #         util_openai_format(single_example_passage, page_content)
        #         annotated_text("  ",annotation("RELEVANT PASSAGE : ", "", font_family="Comic Sans MS", border="2px dashed red"),)
        #         st.write(found_passage)
        # pchroma_client = chromadb.Client()
        # for citem in pchroma_client.list_collections():
        #     print(citem.name)