"""Handwritten form reader: paste or upload a form screenshot, get copyable fields. Runs on a Hugging Face ZeroGPU Space (Gradio SDK). Model: datalab-to/chandra-ocr-2 (override with the MODEL_ID variable). Access: one shared password, stored in the Space secret APP_PASSWORD. Privacy: the app writes no images or extracted values to disk or logs. """ import os # Keep usage analytics off. This must be set before gradio is imported. os.environ.setdefault("GRADIO_ANALYTICS_ENABLED", "False") import hmac import html import re import time try: import spaces # import before torch, as ZeroGPU expects except ImportError: # local runs without the spaces package class spaces: # noqa: N801 @staticmethod def GPU(*args, **kwargs): if len(args) == 1 and callable(args[0]) and not kwargs: return args[0] return lambda fn: fn import gradio as gr import torch from PIL import Image from transformers import AutoModelForImageTextToText, AutoProcessor from chandra.model.hf import generate_hf from chandra.model.schema import BatchInputItem from chandra.output import parse_markdown # ---------------------------------------------------------------- settings MODEL_ID = os.environ.get("MODEL_ID", "datalab-to/chandra-ocr-2") APP_PASSWORD = os.environ.get("APP_PASSWORD", "") MAX_FIELDS = 30 # field boxes shown; extra fields appear in "All fields" MAX_IMAGE_SIDE = 2048 # larger screenshots are shrunk to save GPU time GPU_SECONDS = 60 # GPU time requested per form (ZeroGPU default) # "ocr" writes plain text/tables and is faster. "ocr_layout" also writes a # position box and label for every block, which costs extra generation time. PROMPT_TYPE = os.environ.get("PROMPT_TYPE", "ocr") if PROMPT_TYPE not in ("ocr", "ocr_layout"): PROMPT_TYPE = "ocr" # Cap on generated tokens. Stops a runaway read from burning the whole GPU # slot. Raise it (Space variable MAX_OUTPUT_TOKENS) for very long forms. MAX_OUTPUT_TOKENS = int( os.environ.get("MAX_OUTPUT_TOKENS", 2048 if PROMPT_TYPE == "ocr" else 6000) ) if not APP_PASSWORD: # Fail closed: never start an open app by accident. raise RuntimeError( "APP_PASSWORD is not set. Add it in Space Settings > " "Variables and secrets, then restart the Space." ) # ------------------------------------------------------------------- model # ZeroGPU wants the model moved to "cuda" at module level; a real GPU is # attached only while a @spaces.GPU function runs. DEVICE = "cuda" if torch.cuda.is_available() else "cpu" model = AutoModelForImageTextToText.from_pretrained( MODEL_ID, dtype=torch.bfloat16 ).to(DEVICE) model.eval() model.processor = AutoProcessor.from_pretrained(MODEL_ID) model.processor.tokenizer.padding_side = "left" def to_markdown(raw: str) -> str: """Convert the model's HTML output to markdown text. chandra's parse_markdown only reads top-level
blocks, so plain "ocr" output (no divs) is wrapped in one first. It also backslash-escapes characters such as _ and *, which would corrupt emails, so undo that. """ if not re.search(r"{raw}
" markdown = parse_markdown(raw) return re.sub(r"\\([\\`*_{}\[\]()#+\-.!$|<>~])", r"\1", markdown) @spaces.GPU(duration=GPU_SECONDS) def run_ocr(image: Image.Image) -> tuple[str, bool]: """Read the form image. Returns (markdown text, hit_token_limit).""" batch = [BatchInputItem(image=image, prompt_type=PROMPT_TYPE)] result = generate_hf(batch, model, max_output_tokens=MAX_OUTPUT_TOKENS)[0] return to_markdown(result.raw), result.token_count >= MAX_OUTPUT_TOKENS # --- PARSING START --------------------------------------------------------- MAX_LABEL_CHARS = 160 # some form labels list all their options _HEADER_PAIRS = {("field", "value"), ("label", "value"), ("item", "value")} def _clean(text: str) -> str: """Strip markdown/HTML decoration from one piece of text.""" text = re.sub(r"", " ", text, flags=re.I) text = re.sub(r"<[^>]+>", "", text) text = re.sub(r"!\[[^\]]*\]\([^)]*\)", "", text) text = text.replace("**", "").replace("__", "").replace("`", "") text = html.unescape(text) text = text.strip(" *") return re.sub(r"\s+", " ", text).strip() def parse_fields(markdown: str) -> list[tuple[str, str]]: """Pull (label, value) pairs out of OCR markdown. Handles 'Label: value' lines, markdown tables, and HTML tables. Anything it cannot pair stays available in the full-text box. """ fields: list[tuple[str, str]] = [] seen: set[tuple[str, str]] = set() def add(label: str, value: str) -> None: label = _clean(label).rstrip("::").strip() value = _clean(value) if not label or len(label) > MAX_LABEL_CHARS: return if not re.search(r"[^\W\d_]", label): # label needs at least one letter return if value.startswith("//"): # skip URLs such as https://... return if (label.lower(), value.lower()) in _HEADER_PAIRS: return key = (label.lower(), value.lower()) if key in seen: return seen.add(key) fields.append((label, value)) def add_cells(cells: list[str]) -> None: cells = [c for c in cells] if len(cells) < 2: return for i in range(0, len(cells) - 1, 2): add(cells[i], cells[i + 1]) # HTML tables for row in re.findall(r"]*>(.*?)", markdown, flags=re.I | re.S): cells = re.findall(r"]*>(.*?)", row, flags=re.I | re.S) add_cells([_clean(c) for c in cells]) remaining = re.sub(r"", "", markdown, flags=re.I | re.S) # Markdown tables and 'Label: value' lines for line in remaining.splitlines(): s = line.strip() if not s: continue if s.startswith("|"): cells = [c.strip() for c in s.strip("|").split("|")] if all(re.fullmatch(r":?-{2,}:?", c) or not c for c in cells): continue # separator row add_cells([_clean(c) for c in cells]) continue m = re.match(r"^[#>\-\*\s]*([^::|]{1,160}?)\s*[::]\s*(.*)$", s) if m: add(m.group(1), m.group(2)) return fields # --- PARSING END ----------------------------------------------------------- def prepare_image(image: Image.Image) -> Image.Image: image = image.convert("RGB") image.thumbnail((MAX_IMAGE_SIDE, MAX_IMAGE_SIDE)) return image # --------------------------------------------------------------------- UI def copy_box(**kwargs): """Textbox with a copy button on Gradio 5 and Gradio 6.""" try: return gr.Textbox(buttons=["copy"], **kwargs) # Gradio 6 except TypeError: return gr.Textbox(show_copy_button=True, **kwargs) # Gradio 5 def extract(image): if image is None: raise gr.Error("Upload or paste a form screenshot first.") started = time.perf_counter() try: markdown, truncated = run_ocr(prepare_image(image)) except gr.Error: raise # GPU quota messages already explain themselves except Exception as exc: # never echo form content into errors or logs raise gr.Error( f"Reading the form failed ({type(exc).__name__}). Try again in a minute." ) seconds = time.perf_counter() - started # includes any GPU queue wait fields = parse_fields(markdown) all_fields = "\n".join(f"{label}: {value}" for label, value in fields) if fields: status = ( f"Found {len(fields)} fields. Check the values against the " "screenshot before copying." ) else: status = "No label and value pairs found. Copy from the full text below." if len(fields) > MAX_FIELDS: status += f" The first {MAX_FIELDS} have their own boxes; all are in All fields." if truncated: status += ( " The form was long and the text may be cut off. Set the Space " "variable MAX_OUTPUT_TOKENS higher to read all of it." ) status += f" Read in {seconds:.1f} s." slot_updates = [] for i in range(MAX_FIELDS): if i < len(fields): label, value = fields[i] slot_updates.append(gr.Textbox(label=label, value=value, visible=True)) else: slot_updates.append(gr.Textbox(value="", visible=False)) return [status, all_fields, markdown, *slot_updates] def clear(): hidden = [gr.Textbox(value="", visible=False) for _ in range(MAX_FIELDS)] return [None, "", "", "", *hidden] with gr.Blocks( title="Form reader", analytics_enabled=False, delete_cache=(600, 600), # remove cached uploads after 10 minutes ) as demo: gr.Markdown( "## Form reader\n" "Paste or upload a screenshot of a filled form, then select **Read form**. " "Nothing is saved by this app." ) image_in = gr.Image( label="Form screenshot", type="pil", sources=["upload", "clipboard"], height=380, ) with gr.Row(): read_btn = gr.Button("Read form", variant="primary") clear_btn = gr.Button("Clear") status_md = gr.Markdown("") slots = [ copy_box(label=f"Field {i + 1}", value="", interactive=True, visible=False, lines=1, max_lines=4) for i in range(MAX_FIELDS) ] all_fields_box = copy_box(label="All fields", lines=6, interactive=True) raw_box = copy_box(label="Full text", lines=10, interactive=True) outputs = [status_md, all_fields_box, raw_box, *slots] read_btn.click(extract, inputs=image_in, outputs=outputs) clear_btn.click(clear, inputs=None, outputs=[image_in, status_md, all_fields_box, raw_box, *slots]) def check_login(username: str, password: str) -> bool: """Shared password; the username can be anything.""" return hmac.compare_digest(password.encode(), APP_PASSWORD.encode()) demo.queue(max_size=10) if __name__ == "__main__": demo.launch( auth=check_login, auth_message="Team access. Enter any username and the shared password.", )