Spaces:
Sleeping
Sleeping
Download app.py from Prashant-Mahto/vrag-pdf-reader: direct link, hf CLI and curl.
- Browser
- Download file 3.49 kB
-
https://huggingface.co/spaces/Prashant-Mahto/vrag-pdf-reader/resolve/main/app.py
- Command line
-
hf download hf://spaces/Prashant-Mahto/vrag-pdf-reader/app.py
-
curl -L -o app.py https://huggingface.co/spaces/Prashant-Mahto/vrag-pdf-reader/resolve/main/app.py
3.49 kB
| """V-RAG PDF reader: the live demo sends a visitor's upload here; this Space reads it with | |
| the pipeline's own ingestion code (layoutlmv3 tables, Qwen2.5 VL 3B figures) and returns the chunks, which the | |
| visitor's browser embeds and indexes. Files are deleted right after reading; the log keeps only a short hash of the | |
| name, the page count and the time.""" | |
| import os | |
| import time | |
| os.environ.setdefault("VRAG_CONFIG", os.path.join(os.path.dirname(os.path.abspath(__file__)), "config.yaml")) | |
| import gradio as gr # noqa: E402 | |
| import spaces # noqa: E402 | |
| import reader # noqa: E402 | |
| from src.ingestion import figures, tables # noqa: E402 | |
| from src.ingestion.upload_checks import UploadRefused # noqa: E402 | |
| # Models load once, at module scope, so ZeroGPU can move them to the GPU for each call | |
| tables.check_table_parser() | |
| tables.get_detector().load() | |
| tables.get_layout_model().load() | |
| figures.check_figure_parser() | |
| cap = reader.DailyCap() | |
| def gpu_seconds(folder: str, pages: int) -> int: | |
| """The GPU time to reserve (counted against the visitor's daily quota, which ZeroGPU charges at about 1.5 times | |
| this): 10 s plus 9 s a page, measured on this Space (2026-10-16: 3 pages 31 s, 10 pages 92 to 112 s).""" | |
| return min(110, 10 + 9 * max(1, pages)) | |
| def read_pdf(folder: str, pages: int): | |
| return reader.read(folder) | |
| def parse_document(file: str, request: gr.Request) -> dict: | |
| """One document in, its chunks out (the demo's /parse_document API).""" | |
| started = time.perf_counter() | |
| name = os.path.basename(file or "") | |
| ip, source = reader.client_hop(request.headers.get("x-forwarded-for"), request.client.host if request.client else None) | |
| folder = None | |
| try: | |
| if not cap.allowed(ip): | |
| raise UploadRefused("daily_cap", 429, "This address has read its uploads for today; try again tomorrow.") | |
| folder, pages = reader.prepare(file, name) | |
| chunks = read_pdf(folder, pages) if pages is not None else reader.read(folder) # text needs no GPU | |
| cap.record(ip) | |
| seconds = round(time.perf_counter() - started, 1) | |
| print(f"[reader] {reader.log_name(name)}: {pages or 0} pages, {len(chunks)} chunks, {seconds} s " | |
| f"(client from {source})") | |
| return {"document": name, "pages": pages, "chunks": chunks, "seconds": seconds, | |
| "readers": {"tables": tables.effective_parser(), "figures": figures.effective_parser()}} | |
| except UploadRefused as e: | |
| print(f"[reader] {reader.log_name(name)}: refused ({e.code})") | |
| raise gr.Error(f"{e.code}: {e.message}") | |
| finally: | |
| reader.discard(folder, file) | |
| with gr.Blocks(title="V-RAG PDF reader", delete_cache=(1800, 1800)) as demo: | |
| gr.Markdown("## V-RAG PDF reader\nThe document reader behind the " | |
| "[V-RAG live demo](https://prashantmaht0.github.io/Voice_Enabled_RAG_Pipeline/): tables with a fine " | |
| "tuned LayoutLMv3, charts with Qwen2.5 VL 3B. Upload one PDF (5 pages, 10 MB) or a Markdown or text " | |
| "file. Files are deleted right after reading.") | |
| upload = gr.File(label="Document", file_types=[".pdf", ".md", ".txt"], type="filepath") | |
| result = gr.JSON(label="Chunks") | |
| upload.upload(parse_document, upload, result, api_name="parse_document", concurrency_limit=1) | |
| demo.queue(default_concurrency_limit=1, max_size=8) | |
| if __name__ == "__main__": | |
| demo.launch(max_file_size="10mb") | |