How It Works

Private Knowledge Semantic Search

Postiller uses on-device AI to understand what your content means, finding connections across everything you've saved. No cloud. No servers. Just your ideas, indexed privately on your device.

5 min read

The Problem with Keyword Search

Traditional search fails creative work.

You've saved an article about leading teams through uncertainty. Three months later, you're writing about crisis management and can't find it. Why? Because you search for "crisis" but the article never used that word—it talked about "navigating ambiguity" and "organizational resilience."

Keyword search only finds exact matches. Your brain doesn't work that way. Neither should your tools.


Search by Meaning, Not Memory

Semantic search understands concepts, not just characters. When you search for "leadership during tough times," Postiller surfaces:

  • That article about decision fatigue in executives
  • Your note about a podcast on team morale
  • A bookmark about military leadership principles

None of these contain your exact search terms. All of them are exactly what you need.

How It Works

  1. Embedding creation — When you save content, Postiller converts it into a mathematical representation of its meaning (an "embedding")
  2. Similarity matching — Your search query gets the same treatment, then we find content with similar meaning
  3. Ranked results — Content is scored by conceptual similarity, with your own notes weighted higher than extracted text

The magic is in step one. An embedding captures the meaning of text as a high-dimensional vector—a list of numbers that represents concepts, relationships, and context. When two pieces of content have similar embeddings, they're about similar things, even if they use completely different words.


On-Device = Private by Design

Most AI-powered search sends your data to the cloud. Every query. Every result. Logged, processed, and stored on someone else's servers.

Postiller is different. We use Apple's NLEmbedding framework to generate embeddings entirely on your iPhone or iPad. The AI model runs locally. Your content stays local. There's no server to breach because there's no server at all.

The Privacy Comparison

Traditional AI SearchPostiller
Content uploaded to cloudContent stays on device
Search queries loggedNo query logging
Requires internetWorks offline
Company has accessOnly you have access

This isn't privacy theater. We literally cannot see your content because it never leaves your device. There's no opt-out checkbox because there's nothing to opt out of.


Built for Speed

On-device doesn't mean slow. Postiller searches 10,000+ items in under 50 milliseconds—faster than you can blink. The app feels instant because it is instant.

Behind the Scenes

Smart chunking — Long articles are split into focused ~400 character segments, each with its own embedding. This means semantic search can match specific sections of an article rather than requiring the whole document to be relevant.

Topic extraction — On-device AI identifies 3-5 topic tags per chunk for better matching. Raw embeddings tend to cluster by writing style; topic tags cluster by actual subject matter.

User content prioritization — Your notes and learnings rank higher than raw article text. When you write "this is important because..." that insight matters more than what an author happened to say.

Similarity scoring — Results are ranked by how closely their meaning matches your query, with the most relevant content surfaced first.


From Search to Creation

Finding related content is just the beginning. When you generate a post, Postiller's semantic search automatically pulls relevant context:

  • Bookmarks that match your topic
  • Ideas you've captured
  • Notes explaining why something matters to you

This is called Retrieval-Augmented Generation (RAG). Instead of the AI making things up, it draws from your specific knowledge—the articles you've read, the insights you've noted, the connections you've made.

Your posts sound like you because they're built from what you've actually read, saved, and thought about.


The Technical Foundation

Apple NLEmbedding

We use Apple's Natural Language framework, the same technology powering search in Notes, Mail, and Photos. It's optimized for Apple Silicon, runs without network access, and requires no API keys or subscriptions.

The embeddings are 512-dimensional vectors that capture semantic meaning. Similar concepts cluster together in this high-dimensional space, making similarity search a matter of finding nearby vectors.

Vector Similarity Search

Finding similar content means finding vectors that point in similar directions. This is a well-understood mathematical operation—cosine similarity—that's fast and reliable. No machine learning happens at search time; all the intelligence was baked in when the embeddings were created.

Why On-Device Works

Modern iPhones have Neural Engine processors specifically designed for AI workloads. What once required cloud servers can now run on your phone in milliseconds. The models are smaller than cloud models, but they're optimized for exactly this use case: understanding the meaning of text quickly and privately.


Your Second Brain, Actually Private

Unlike other "AI-powered" tools, Postiller doesn't ask you to trust us with your data. There's nothing to trust—we never see it.

Your bookmarks. Your ideas. Your notes. Your search queries. All of it stays on your device, processed by AI that runs locally, searchable without an internet connection.

This is what privacy-first actually means.

Experience private AI for yourself

Your data stays on your device. Your ideas stay yours. See what privacy-first content creation feels like.