AntflyAntfly
Back to Builder Stories

SearchAF

A free macOS app for on-device semantic search across your own files. PDFs, screenshots, and documents, all indexed locally by Antfly.

0
Bytes sent off-device
1 binary
Engine, models, and index
PDF, image, Office
File types read
Free
For everyone

What it is#

SearchAF is the simplest thing you can build on Antfly: point it at a few folders and ask questions about what is in them. It reads PDFs, Office documents, screenshots, and images, and it answers by meaning rather than by filename. Nothing leaves your machine.

It is also the thing on this site you can download without writing a line of code. If you want to feel what the engine does before you run it yourself, start here.

How it works#

SearchAF is a thin macOS shell around a single local Antfly engine. The app watches the folders you pick and hands every new or changed file to Antfly. Extraction happens inside the engine: text from documents, OCR and captions for images, produced by Antfly Inference models that ship with the binary.

Each document is chunked, embedded, and indexed as it arrives. A query runs against the keyword index and the vector index together and the results are fused, which is why a vague memory of a file ("the screenshot of the error from last week") lands on the right one.

The same index is exposed over MCP, so an agent working in Claude Code or Cursor can search your files with the same tools you use.

Stack#

  • Engine: Antfly, single node, local data directory
  • Inference: Antfly Inference (embedders, OCR, image captions), all on-device
  • Search: Hybrid full-text + vector with reciprocal rank fusion
  • Agents: MCP server over the same index
  • Platform: macOS
Architecture
1
Folder watcher
Watches the folders you choose and feeds new and changed files into Antfly as they land
2
Extraction
Text from PDFs and Office files, OCR and captions for screenshots and images, all produced locally by Antfly Inference
3
Chunk, embed, index
Antfly chunks and embeds each document inside the engine, alongside a full-text index, with nothing to wire together
4
Hybrid search
Queries hit the keyword and vector indexes together and are fused, so "that invoice from the airline" finds the PDF
5
MCP for agents
The same index is exposed to Claude Code and other MCP clients, so your agents can search your files the way you do