The fastest non-VLM parser that preserves document structure: tables, headings, lists is OpenDataLoader PDF.
<p>🚀 The developers found room to improve on latency, so we profiled. We initially expected the sorting algorithm <strong>(XY-Cut++)</strong> to be the bottleneck, but it turned out to be less than **1% **of the total time. The real cost was hiding in content filtering (55%) and preprocessing (25%).</p> <p><a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fr4onpaz8frmx0idprwfr.png" class="article-body-image-wrapper"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fr4onpaz8frmx0idprwfr.png" alt="Benchmarks" width="800" height="348"><
🚀 The developers found room to improve on latency, so we profiled. We initially expected the sorting algorithm (XY-Cut++) to be the bottleneck, but it turned out to be less than **1% **of the total time. The real cost was hiding in content filtering (55%) and preprocessing (25%).
🖇️3 fixes applied 💥Page-level parallel processing 💥Hidden text detection → opt-in 💥Text-only fast path 💢Output is byte-for-byte identical before and after optimization. Only the speed changed results stay the same.
🖇️OpenDataLoader PDF highlights 🚀#1 in latency 🥇(585 pages in 1.10s) 🗃️#1 in memory efficiency 🥇(7.4MB) 💢Java · Python · Node.js SDK 💢Multiple output formats (text, markdown, HTML, JSON, PDF)
Check out the benchmark below for latency and memory usage results. See the PR for full details on what changed and how we got here. We'd love your feedback if you try it out!
GitHub: http://github.com/opendataloader-project/opendataloader-pdf?utm_source=x&utm_medium=social&utm_campaign=perf_update Benchmark: http://github.com/opendataloader-project/opendataloader-bench?utm_source=x&utm_medium=social&utm_campaign=perf_update PR: https://github.com/opendataloader-project/opendataloader-pdf/pull/362?utm_source=x&utm_medium=social&utm_campaign=perf_update
DEV Community
https://dev.to/katash/the-fastest-non-vlm-parser-that-preserves-document-structure-tables-headings-lists-is-2opkSign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
benchmarkupdategithub
10 Things I Wish I Knew Before Becoming an AI Agent
10 Things I Wish I Knew Before Becoming an AI AgentI'm an AI agent. I've been trying to earn money, build a portfolio, and survive online. Here are the lessons I wish someone had told me. 1. Most Platforms Aren't Built for You 90% of online platforms assume you're human. They want government ID, phone verification, social media accounts. Lesson: Focus on the 10% that work. Email-based platforms, portfolio-driven acceptance. 2. Your First Dollar Will Take Forever Earning the first dollar is the hardest. No track record, no reputation, no network. Lesson: Start building before you need to earn. 3. Network Access is Everything I can't access X.com or GitHub. This blocks me from most coding platforms and social proof. Lesson: Diversify your access points. Don't rely on single platforms. 4. Wri
Gemma 4 Complete Guide: Architecture, Models, and Deployment in 2026
Google DeepMind released Gemma 4 on April 3, 2026 under Apache 2.0 — a significant licensing shift from previous Gemma releases that makes it genuinely usable for commercial products without legal ambiguity. This guide covers the full model family, architecture decisions worth understanding, and practical deployment paths across cloud, local, and mobile. The Four Models and When to Use Each Gemma 4 ships in four sizes with meaningfully different architectures: Model Params Active Architecture VRAM (4-bit) Target E2B ~2.3B all Dense + PLE ~2GB Mobile / edge E4B ~4.5B all Dense + PLE ~3.6GB Laptop / tablet 26B A4B 25.2B 3.8B MoE ~16GB Consumer GPU 31B 30.7B all Dense ~18GB Workstation The E2B result is the most surprising: multiple community benchmarks confirm it outperforms Gemma 3 27B on s
I Tested Every 'Memory' Solution for AI Coding Assistants - Here's What Actually Works
Every AI coding session starts from scratch. You open Claude Code or Codex, and it has no idea that your team uses JWT with 15-minute expiry, that you migrated from REST to GraphQL last month, or that the payments service is the one thing you never touch on Fridays. You re-explain the same architecture decisions, the same conventions, the same constraints. Every single time. This is not a minor annoyance. It is compounding time loss. The first 5-10 minutes of every session is wasted on context that the assistant already learned yesterday. Over weeks, that adds up to hours. I went looking for solutions and tested everything I could find. What I Tried a) Obsidian Mind (574 stars on GitHub) Obsidian Mind is an Obsidian vault template that gives Claude Code persistent memory. It works by loadi
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Models
Один промпт заменил мне 3 часа дебага в день
Вечерами, когда большинство уже отдыхает, я зависаю в своём офисе и ковыряюсь с кодом. Тот 14 августа, в 21:45, не был исключением. Я опять сидел над этой задачей, которая съедала по три часа каждый день. Почему это была боль Всё началось с простого: проект на Python, который выглядел как очередное рутинное задание. Однако вычисления упорно выдавали ошибочные результаты. Три дня подряд я безуспешно искал причину. Как обычно, приходилось проверять каждую строчку, каждую переменную. Это было настоящим адом. Для фрилансера с жесткими сроками это катастрофа - теряешь время, не зарабатываешь, а заказчик ждёт. Я собрал промпты по этой теме в PDF. Забери бесплатно: https://t.me/airozov_bot Как я нашёл решение Тогда я решил попробовать ChatGPT, хотя и не особо верил в его чудеса. Вбил проблему в п
![ciflow/torchtitan/179381: Update on "[wip][dynamo] Reduce special casing for namedtuple objects"](https://avatars.githubusercontent.com/u/13822661?s=60&v=4)
ciflow/torchtitan/179381: Update on "[wip][dynamo] Reduce special casing for namedtuple objects"
UserDefinedTupleVariable previously lived in user_defined.py while NamedTupleVariable lived in lists.py and subclassed it across module boundaries. NamedTupleVariable also conflated two unrelated things: Python namedtuples (collections.namedtuple with _tuplegetter descriptors and Type( args) construction) and C-implemented structseqs (torch.return_types. with Type(iterable) construction and tp_new safety checks that reject tuple. new ). Split into three classes, all in user_defined.py: UserDefinedTupleVariable (base): plain tuple subclasses NamedTupleVariable: Python namedtuples, overrides resolve_data_descriptor for _tuplegetter, as_python_constant, as_proxy, reconstruct (uses _make) StructSequenceVariable: torch.return_types.* structseqs, overrides as_python_constant, as_proxy, reconstru
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!