Fetching metadata from the HF Docker repository...
Jul 28 14:40 Fix OOM crashes on Render proxy by parsing PDFs using mmap from disk instead of massive byte duplication in memory 14:14 Add per-page rendering try-except resilience to prevent PDF parsing crashes during upload 14:06 Fix 502 upload crash: add proactive garbage collection (gc.collect) to keep RAM under 80MB on cloud containers (Render) 13:57 Optimize render speed to 1200px (0.02s/page) + remove redundant PDF text re-extraction in auto_classify to permanently fix 502 gateway timeouts 13:53 Fix 502 crash root cause: replace PIL Image.frombytes with native C++ PyMuPDF pix.tobytes('jpg', jpg_quality=95) 13:47 Fix classification syntax bug + preserve 4000 char page text + enrich UAE drawing keywords + fall back to PDF text re-extraction 12:59 Eliminate PyMuPDF thread-unsafe access causing process SIGSEGV crash and HTTP 502 error 01:46 Optimize PDF page extraction and classification speed with 0% impact on extraction accuracy Jun 4 00:17 fix: force white background in fast_save_pdf_pages for React backend 00:00 fix: force white background in PDF loader to prevent black squares in Gemini Jun 3 16:20 perf: eliminate Step 3 hangs by shrinking max PNG DPI to 50 for all blueprints, preventing massive network timeouts when sending images to Gemini. Increased concurrent extraction workers to 8 to slash processing time. 14:59 fix: reduce max_workers to 1 in fast_save_pdf_pages to prevent RAM thrashing on large A0 blueprints 14:42 perf: Massive speed boost for Step 1 (Upload) and Step 2 (Classification) by bypassing Numpy/PIL and writing native C++ compressed images directly to disk 00:42 Fix UnboundLocalError in CV extractor and revert hardcoded heuristics Jun 2 18:29 THE QS HUB v2.0 - Complete React + FastAPI Platform (clean deploy)