app.py

# import gradio as gr
# from transformers import AutoProcessor, AutoModelForImageTextToText
# from PIL import Image
# import re

# # Load SmolDocling model & processor once
# processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
# model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")

# def extract_fcel_values_from_image(image, prompt_text):
#     """Run SmolDocling on an image and return numeric values inside <fcel> tags."""
#     # Prepare prompt for the model
#     messages = [
#         {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
#     ]
#     prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
#     inputs = processor(text=prompt, images=[image], return_tensors="pt")

#     # Generate output
#     outputs = model.generate(**inputs, max_new_tokens=2048)
#     prompt_length = inputs.input_ids.shape[1]
#     generated = outputs[:, prompt_length:]
#     result = processor.batch_decode(generated, skip_special_tokens=False)[0]
#     clean_text = result.replace("<end_of_utterance>", "").strip()

#     # Extract only <fcel> values
#     values = re.findall(r"<fcel>([\d.]+)", clean_text)
#     values = [float(v) for v in values]  # convert to floats

#     return values, clean_text

# def compare_images(image1, image2, prompt_text):
#     # Extract fcel values from both images
#     values1, raw1 = extract_fcel_values_from_image(image1, prompt_text)
#     values2, raw2 = extract_fcel_values_from_image(image2, prompt_text)

#     # Calculate accuracy
#     if len(values1) == len(values2) and values1 == values2:
#         accuracy = 100.0
#     else:
#         matches = sum(1 for a, b in zip(values1, values2) if a == b)
#         total = max(len(values1), len(values2))
#         accuracy = (matches / total) * 100 if total > 0 else 0

#     return {
#         # "Extracted Values 1": values1,
#         # "Extracted Values 2": values2,
#         "Accuracy (%)": accuracy
#     }

# # Gradio UI
# demo = gr.Interface(
#     fn=compare_images,
#     inputs=[
#         gr.Image(type="pil", label="Upload First Table Image"),
#         gr.Image(type="pil", label="Upload Second Table Image"),
#         gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Extract table as OTSL)", label="Prompt")
#     ],
#     outputs="json",
#     title="Table Data Accuracy Checker (SmolDocling)",
#     description="Uploads two table images, extracts only <fcel> values from OTSL output, and compares them for accuracy."
# )

# demo.launch()

import gradio as gr
from transformers import AutoProcessor, AutoModelForImageTextToText
from PIL import Image

# Load model & processor once at startup
processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
model = AutoModelForImageTextToText.from_pretrained("ds4sd/SmolDocling-256M-preview")

def smoldocling_readimage(image, prompt_text):
    messages = [
        {"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt_text}]}
    ]
    prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(text=prompt, images=[image], return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=1024)
    prompt_length = inputs.input_ids.shape[1]
    generated = outputs[:, prompt_length:]
    result = processor.batch_decode(generated, skip_special_tokens=False)[0]
    return result.replace("<end_of_utterance>", "").strip()

# Gradio UI
demo = gr.Interface(
    fn=smoldocling_readimage,
    inputs=[
        gr.Image(type="pil", label="Upload Image"),
        gr.Textbox(lines=1, placeholder="Enter prompt (e.g. Convert to docling)", label="Prompt"),
    ],
    outputs="ostl",
    title="SmolDocling Web App",
    description="Upload a document image and convert it to structured docling format."
)

demo.launch()