src/RAG.py

import clip
import faiss
from PIL import Image
from pypdf import PdfReader
import pandas as pd
import re
import os
import fitz
import torch
import numpy as np
from tqdm import tqdm
import base64
import json


class RAG:
    def __init__(
            self,
            fais_index_path,
            clip_model="ViT-B/32",
            reranker=None,
            device="cpu",
            image_invoice_index_path=None,
            path_to_invoices=None,
            path_to_images=None,
            path_to_invoice_json=None
    ):
        self.index = faiss.read_index(fais_index_path)
        self.model, self.preprocess = clip.load(clip_model, device=device)
        self.device = device
        if image_invoice_index_path:
            self.image_invoice_index = pd.read_csv(image_invoice_index_path)
        self.path_to_invoices = path_to_invoices
        self.path_to_images = path_to_images
        self.reranker = reranker
        self.invoice_json = None
        self.invoice_json_granular = None
        if path_to_invoice_json:
            if type(path_to_invoice_json) == str:
                with open(path_to_invoice_json, "r") as f:
                    self.invoice_json = json.load(f)
            elif type(path_to_invoice_json) == dict and set(list(path_to_invoice_json.keys())) == {"invoices", "invoices_granular"}:
                with open(path_to_invoice_json["invoices"], "r") as f:
                    self.invoice_json = json.load(f)
                with open(path_to_invoice_json["invoices_granular"], "r") as f:
                    self.invoice_json_granular = json.load(f)
            else:
                raise ValueError("Invalid format for invoice json.")


    @staticmethod
    def image_to_base64(image_path):
        with open(image_path, "rb") as image_file:
            return base64.b64encode(image_file.read())

    def search_text(self, text, k=1):
        text_features = self.model.encode_text(clip.tokenize([text]).to(self.device))
        text_features /= text_features.norm(dim=-1, keepdim=True)
        text_features = text_features.detach().numpy()
        distances, indices = self.index.search(text_features, k)
        return distances, indices

    def search_image(self, image=None, image_path=None, k=1):
        if image is None and image_path is None:
            raise ValueError("Either image or image_path must be provided.")
        if image is None:
            image = Image.open(image_path)
        image_input = self.preprocess(image).unsqueeze(0).to(self.device)
        image_features = self.model.encode_image(image_input)
        image_features /= image_features.norm(dim=-1, keepdim=True)
        image_features = image_features.detach().numpy()
        distances, indices = self.index.search(image_features, k)
        return distances, indices

    def return_invoice_table(self, path=None, invoice_is_table=True, use_granular_invoice=False):
        if path is None and not invoice_is_table:
            raise ValueError("Path to invoice must be provided.")
        if self.invoice_json is None and invoice_is_table:
            raise ValueError("Path to invoice json must be provided.")
        if self.invoice_json_granular is None and use_granular_invoice:
            raise ValueError("Path to granular invoice json must be provided.")

        if invoice_is_table and not use_granular_invoice:
            return self.invoice_json[path]
        elif invoice_is_table and use_granular_invoice:
            return self.invoice_json_granular[path]

        pdf_path = f"{self.path_to_invoices}/{path}"
        reader = PdfReader(pdf_path)
        page = reader.pages[0]
        text = page.extract_text()

        table_text = re.search(r"Beschädigtes Teil.*?Gesamtsumme:.*?EUR", text, re.DOTALL).group()

        lines = table_text.splitlines()
        header = lines[0]
        other_text = "\n".join(lines[1:])
        cleaned_text = re.sub(r"(?<!\d)\n", " ", other_text)

        table = header + "\n" + cleaned_text

        inv = table.split("\n")
        reformatted_inv = "Beschädigtes Teil | Teilkosten (EUR) | Arbeitsstunden | Arbeitskosten (EUR/Stunde) | Gesamtkosten (EUR)\n" + "\n".join(
            " ".join(inv[i].split(" ")[:-4]) + " | " + ' | '.join(inv[i].split(" ")[-4:]) for i in
            range(1, len(inv) - 1)) + "\n" + inv[-1]

        return reformatted_inv

    def find_invoice(
            self,
            image=None,
            image_path=None,
            return_only_path=True,
            k=1,
            damage_description=None,
            invoice_is_table=True,
            use_granular_invoice=False
    ):
        if self.image_invoice_index is None:
            raise ValueError("No index for invoices found.")
        _, indices = self.search_image(image=image, image_path=image_path, k=k)
        img_ids = self.image_invoice_index.iloc[indices[0]]['img_id'].values
        invoices = self.image_invoice_index[self.image_invoice_index['img_id'].isin(img_ids)]['invoice'].values.tolist()
        images_paths = self.image_invoice_index[self.image_invoice_index['img_id'].isin(img_ids)]['image'].values.tolist()

        if self.reranker:
            if damage_description is None:
                raise ValueError("Damage description must be provided.")
            images = [f"{self.path_to_images}/{img_path}" for img_path in images_paths]
            results = self.reranker.rank(damage_description, images, doc_ids=invoices)
            invoices = [doc.doc_id for doc in results]
            print(invoices)

        if return_only_path:
            return invoices, images_paths

        if not self.path_to_invoices:
            raise ValueError("Path to data must be provided.")

        invoices_tables = []

        for invoice in invoices:
            reformatted_inv = self.return_invoice_table(
                invoice, invoice_is_table, use_granular_invoice=use_granular_invoice
            )
            invoices_tables.append(reformatted_inv)

        return invoices_tables, invoices


def build_rag(directory):
    invoices = os.listdir(f"{directory}/invoices_validated")
    invoices = [i for i in invoices if i.endswith(".pdf")]

    image_invoice = []
    os.makedirs(f"{directory}/images", exist_ok=True)
    os.makedirs(f"{directory}/invoices", exist_ok=True)

    for invoice in invoices:
        doc = fitz.open(f"{directory}/invoices_validated/{invoice}")

        page = doc[1]
        image_list = page.get_images(full=True)
        text = page.get_text()

        xref = image_list[0][0]
        base_image = doc.extract_image(xref)
        image_bytes = base_image["image"]
        image_name = invoice.replace(".pdf", ".png")
        with open(f"{directory}/images/{image_name}", "wb") as img_file:
            img_file.write(image_bytes)

        doc.delete_pages(range(1, doc.page_count))
        doc.save(f"{directory}/invoices/{invoice}")
        doc.close()

        image_invoice.append({
            "invoice": invoice,
            "image": image_name,
            "description": text
        })

    image_invoice = pd.DataFrame(image_invoice)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model, preprocess = clip.load("ViT-B/32", device=device)
    images = image_invoice["image"].tolist()

    embeddings = []
    image_indices = []
    img_ids = []

    for idx, img_path in enumerate(tqdm(images)):
        image = Image.open(f"{directory}/images/{img_path}")
        img_ids.append(idx)
        inputs = preprocess(image).unsqueeze(0).to(device)

        with torch.no_grad():
            image_embedding = model.encode_image(inputs)

        image_embedding = image_embedding / image_embedding.norm(dim=-1, keepdim=True)
        embeddings.append(image_embedding.cpu().numpy().astype("float32"))
        image_indices.append(img_path)

    image_invoice["img_id"] = img_ids
    image_invoice.to_csv(f"{directory}/image_invoice.csv", index=False)

    embeddings_np = np.vstack(embeddings)

    dimension = embeddings_np.shape[1]
    index = faiss.IndexFlatIP(dimension)
    index.add(embeddings_np)

    faiss.write_index(index, f"{directory}/invoice_index.faiss")