Anvit Launch

Overview

Anvit is an Android application that lets users chat with PDF and Word documents using an AI model that runs entirely on the device. According to the site, it "reads, reasons over, and answers questions about your files without sending a single byte to the cloud." After the model downloads once, no internet connection is required and the app works in airplane mode. Answers are produced through an agentic retrieval pipeline described as "Plan → Decompose → Search → Filter → Answer." The site states that complex questions are broken into sub-queries, retrieved across multiple passes, and refined before the final answer is shown, using hybrid dense semantic search plus BM25 keyword search, with corrective re-querying when retrieval quality is low. Every response cites the document and passages it was built from, and a Think Mode displays the model's chain-of-thought. Documents are organized into named collections and indexed locally. When a document is added, Anvit parses its headings, sections, tables and lists, then "splits it into section-aware chunks of up to 8,000 tokens each." The language model is Google's Gemma 4, available in E2B (1.3 GB) and E4B (2.5 GB) variants, running on-device via LiteRT-LM on CPU or GPU, with EmbeddingGemma (300M) or Gecko (110M) providing embeddings. The app is built with Kotlin, Jetpack Compose and Material 3, and the site states there are no accounts, analytics or telemetry.