Pipit: Building a 100% On-Device AI Universal Search Engine for Android (<50ms, Open Source)

Описание и применение

Pipit — это открытый поиск на Android, работающий полностью офлайн, использующий OCR и мультимодальные модели для быстрого поиска по фотографиям, PDF, сообщениям и другим данным.

Задача

Finding content on smartphones is difficult because stock search only checks filenames and cloud search uploads private files.

Результат

Pipit delivers a unified, offline search across seven domains with sub‑50 ms latency, eliminating the need for cloud uploads.

Рабочий процесс

• Index local media using Google ML Kit OCR for PDFs and receipts
• Extract visual features for semantic image search
• Store embeddings in a local database
• Query unified search bar with natural language
• Return results from all domains within 50 ms

Доказательства

Smartphones hold our entire digital lives: thousands of photos, receipts, work PDFs, voice notes, and gate codes buried in messages.

Источник

Yet finding anything is painful because stock search only checks cryptic filenames like IMG_20240914.jpg or Scan_Doc_09.pdf.

Источник

Pipit 🐦 — an open-source, on-device AI universal search engine for Android.

Источник

Deep Content OCR: Searches inside PDFs, receipts, and screenshots using on-device Google ML Kit.

Источник

Semantic Visual Search: Natural language search across your gallery ("yellow vintage car", "handwritten grocery list").

Источник

Unified Search Bar: 7 domains in one place (Photos, PDFs, Messages, Contacts, Video, Audio, Apps).

Источник

100% Offline & Air-Gapped: Runs quantized Google LiteRT multimodal models directly on your phone's silicon. Zero cloud, zero tracking.

Источник

Sub-50ms Latency: Accelerated via local Vulkan GPU/NPU delegates.

Источник

Source code and release APKs are available on GitHub: Repository: https://github.com/sanjay434343/Pipit

Источник