auditqa/doc_process.py

import glob
import os
from langchain.text_splitter import RecursiveCharacterTextSplitter, SentenceTransformersTokenTextSplitter
from transformers import AutoTokenizer
from torch import cuda
from langchain_community.document_loaders import PyMuPDFLoader
from langchain_community.embeddings import HuggingFaceEmbeddings, HuggingFaceInferenceAPIEmbeddings
from langchain_community.vectorstores import Qdrant
device = 'cuda' if cuda.is_available() else 'cpu'
#from dotenv import load_dotenv
#load_dotenv()

#HF_token = os.environ["HF_TOKEN"]
path_to_data = "./data/"


def process_pdf():
    files = {'MWTS2021':'./data/MWTS2021.pdf',
            'MWTS2022':'./data/MWTS2022.pdf',
            'Consolidated2021':'./data/Consolidated2021.pdf'}
    docs = {}
    for file,value in files.items():
        try:
            docs[file] = PyMuPDFLoader(value).load()
        except Exception as e:
            print("Exception: ", e)

    
    # text splitter based on the tokenizer of a model of your choosing
    # to make texts fit exactly a transformer's context window size
    # langchain text splitters: https://python.langchain.com/docs/modules/data_connection/document_transformers/
    chunk_size = 256
    text_splitter = RecursiveCharacterTextSplitter.from_huggingface_tokenizer(
            AutoTokenizer.from_pretrained("BAAI/bge-small-en-v1.5"),
            chunk_size=chunk_size,
            chunk_overlap=10,
            add_start_index=True,
            strip_whitespace=True,
            separators=["\n\n", "\n"],
    )
    
    all_documents = {'Consolidated':[], 'MWTS':[]}
    
    for file,value in docs.items():
        doc_processed = text_splitter.split_documents(value)
        for doc in doc_processed:
            doc.metadata["source"] = file
            doc.metadata["year"] = file[-4:]
        for key in all_documents:
            if key in file:
                print(key)
                all_documents[key].append(doc_processed)
  
    for key, docs_processed in all_documents.items():
        docs_processed = [item for sublist in docs_processed for item in sublist]
        all_documents[key] = docs_processed


    embeddings = HuggingFaceEmbeddings(
        model_kwargs = {'device': device},
        encode_kwargs = {'normalize_embeddings': True},
        model_name="BAAI/bge-small-en-v1.5"
    )

    qdrant_collections = {}

    for file,value in all_documents.items():
        print("emebddings for:",file)
        qdrant_collections[file] = Qdrant.from_documents(
            value,
            embeddings,
            location=":memory:", 
            collection_name=file,
        )
    print("done")
    return qdrant_collections