app.py

import os
import easyocr
import gradio as gr
from PIL import Image
from llama_index.core import Settings
from llama_index.llms.gemini import Gemini
from llama_index.core import Document, VectorStoreIndex
from llama_index.embeddings.gemini import GeminiEmbedding

reader = easyocr.Reader(['en'])

llm = Gemini(api_key=os.getenv('GEMINI_API_KEY'), model_name="models/gemini-2.0-flash")
gemini_embedding_model = GeminiEmbedding(api_key=os.getenv('GEMINI_API_KEY'), model_name="models/embedding-001")

# Set Global settings
Settings.llm = llm
Settings.embed_model = gemini_embedding_model

def inference(img_path, width_ths):
    output = reader.readtext(img_path, detail=0, slope_ths=0.7, ycenter_ths=0.9,
                          height_ths=0.8, width_ths=width_ths, add_margin=0.2)
    
    output = "\n".join(output)
    
    # create a Document object from the extracted text
    doc = Document(text = output)

    # Create an index from the documents and save it to the disk.
    index = VectorStoreIndex.from_documents([doc])

    # save the index
    index.storage_context.persist(persist_dir = "./receiptsembeddings")
    
    return output

title = "Receipt RAG"
description = "A simple Gradio interface to query receipts using RAG"
examples = [["data/receipt_00000.JPG", 7.7],
            ["data/receipt_00001.jpg", 7.7]]

demo = gr.Interface(inference, 
                    inputs = [gr.Image(width=320, height=320, label="Input Receipt"), 
                              gr.Slider(0, 10, 7.7, 0.1, label="Width Threshold to merge bounding boxes")],
                    outputs= [gr.Textbox(label="OCR Output", type="text")],
                    title=title,
                    description=description,
                    examples=examples)

demo.launch()