Read Less Understand More
The AI-powered workspace where you chat with, mark up, and edit PDFs to drive business results together. Natively on your Mac ✦ Zero cloud uploads required.
No credit card needed ✦ Unlimited time on Free plan
Attention Is All You Need
Page 4 of 15AUTHORS & AFFILIATIONS
Ashish Vaswani*, Noam Shazeer*, Niki Parmar*, Jakob Uszkoreit*, Llion Jones*, Aidan N. Gomez*,Łukasz Kaiser* — Google Brain & Google Research*Equal contribution. Correspondence: avaswani@google.com
1 Introduction
Recurrent networks, long short-term memory [18] and gated recurrent architectures [7] have long been the standard approach to sequence modeling. Their step-by-step structure, however, blocks parallelization inside a single training example — a limitation that bites hardest on long sequences. To address this, we propose the Transformer: a model architecture eschewing recurrence and relying entirely on self-attention to draw global dependencies.
3.2 Attention
Attention maps a query and a set of key-value pairs onto an output. That output is a weighted sum over the values, and each value's weight comes from a compatibility function between the query and the matching key. Self-attention, sometimes called intra-attention, is an attention mechanism relating different positions of a single sequence in order to compute a representation of the sequence.
METADATA
arXiv:1706.03762v7 [cs.CL] 12 Jun 2017Journal: Advances in Neural Information Processing Systems 30 (NIPS 2017)
3.2.1 Scaled Dot-Product Attention
Queries and keys carry dimension dk; values carry dimension dv. Each query is dotted against every key, each result is divided by √dk, and a softmax converts those numbers into weights over the values.Queries are processed as a batch, packed into a matrix Q, with keys and values packed into K and V. The full output matrix is then: Attention(Q, K, V) = softmax(Q Kᵀ / √dk) V
Figure 1: Attention Mechanism Complexity
Table 1: Maximum path lengths, sequential operations and maximum complexity. Self-Attention layer complexity per layer is O(n^2 · d). Under this paradigm, connection path lengths are minimized, leading to more stable context retention across sentences.
How Velox PDF stacks up
| Feature Matrix | Velox PDF | Adobe Acrobat | PDF Expert |
|---|---|---|---|
| Local AI Assistant Chat | ✓ | — | — |
| Inline Page Citations | ✓ | — | — |
| Native macOS UI App | ✓ | ✓ | ✓ |
| Permanent Free AI Tier | ✓ | — | — |
| Markup, Highlighting, Signature | ✓ | ✓ | ✓ |
| Zero Cloud Server Logs | ✓ | — | — |
AI Agent in action
Supercharge your reading
Yes. Since everything runs on local CPU/Apple Silicon resources, no third-party servers compile or store your PDF files.
📄 Page 1 · Sec. 2 Data Security🛡️ 100% HIPAA compliant ✦ SOC2 level local sandbox ✦ Absolute privacy
Why chose Velox PDF
An agent for every document task
High-performance local AI reading paired with rich document markup and encryption.
Local AI Chat agent
Ask your PDFs anything in plain English and receive cited answers with exact page highlights.
Precise Page Citations
Every answer references the exact page location and source paragraph, so you avoid hallucinations.
100% Privacy-First Architecture
Everything remains computed locally on your Mac hardware. Your files never touch external clouds.
Multiple AI Speeds
Toggle between Fast (quick summaries) and Expert (deep technical analysis) models on the fly.
Native Markup & Signatures
Highlight, circle, draw, and drop electronic signatures directly on contract lines.
Unrivaled Value Matrix
Includes a permanent free tier with 10 questions/week, or Pro version access at only $9.99/mo.
pricing
Pay monthly, or once. Never both.
Starts with a 14-day free trial.
- Every markup & signing tool
- AI assistant included
- All future updates
- Cancel anytime, no questions
One payment. Yours for good.
- Every markup & signing tool
- AI assistant included
- Every future update, free
- No subscription, ever
Prices in AUD. Sales tax or VAT may apply at checkout, handled by Stripe.
Have a promo code?
Enter your code to unlock free access or extended trial.
questions
Before you ask.
What AI models does Velox PDF support?
Velox PDF's built-in AI works free with no setup — pick Fast, Balanced, or High Perf in Settings depending on how much you want to wait for an answer. If you'd rather use your own account, you can add your own API key instead.
Does Velox PDF work offline?
Yes for everything except the AI assistant. Highlighting, drawing, signing, page management, and search all run fully offline. Asking a question needs a connection, since it talks to a model in the cloud.
Is my PDF private?
Your PDFs stay on your Mac for viewing, highlighting, drawing, signing, and local search — that never uploads anywhere. Asking the AI assistant a question sends only the relevant text to generate an answer: to Velox PDF's own servers on the built-in AI, or directly to your own provider if you've added your own key.
Does it work on Apple Silicon?
Yes — Velox PDF is a native Swift app compiled to run on M-series Macs, as well as Intel Macs, on macOS 14 (Sonoma) or later.
Can I chat with scanned PDFs?
Yes — run OCR on a scanned page (one click) and the text becomes selectable, searchable, and answerable by the AI assistant just like any other PDF.
What happens after my trial?
On day 15, your card is charged $9.99 and your subscription continues monthly. Cancel any time before then from the link in your receipt email and you won't be charged.
Can I switch from monthly to a lifetime license later?
Yes — just purchase the $39.99 lifetime license whenever you're ready and cancel your monthly subscription.